Working notes

Append-only, newest last. Things a session learned that the queue and decision files do not say.

2026-10-19 — opening the queue (batch-000, batch-001)

Patterns that predicted the outcome.

• Demonstrated warrant, non-derivation defeater. If a claim filed as demonstrated has a defeat condition about predictive power, modelling adequacy, a population’s behaviour or an impossibility, the claim is compound or misclassified. Seen in CLAIM-077, CLAIM-098, CLAIM-285 and CLAIM-128. Read the defeated_if before the body: it is the quickest test.
• A promote_when that bundles two jobs. A proof plus “and it explains X”, or a theorem plus “and humans behave so”. These became splits or narrowings: THEORY-155, -008, -001, -156.
• “The record’s own” in a THEORY. A result the record proves belongs in a CLAIM (ADR-032), as THEORY-005’s became.
• Check what the source actually proves. THEORY-087’s “regular half is derived in the sources” held only for linear-Gaussian models. Grep the NOTE’s claims table for “cited” or “approximation” before trusting a promote_when’s own summary.
• Check the statement’s quantifiers. CLAIM-264 (strong generator vs density) and CLAIM-123 (sufficient stated as necessary) both had a correct theorem stated with the wrong scope. That is a correction, not a promotion blocker.

Traps.

• luria lint takes about eight minutes. The owner, 2026-10-19: open the PR and rely on CI to run luria, rather than waiting on it locally. check.py is the local check that matters for the ledger.
• CI’s docs-check pushes a “docs: repair sources” commit onto a PR branch when sources need repair, for example unlinked codes in a changelog fragment written after luria link --fix. The run that bot commit triggers stops at action_required, and a maintainer has to approve it. Run luria link --fix last, after the changelog, and pull before pushing again.
• luria new adr takes the next free number at once (the filing mode). Concurrent branches can collide on it.
• Codes in these YAML files are not linked or upgraded by luria repair. Use concretized codes. PR #165’s tmp codes (CLAIM-371, CLAIM-370, THEORY-719, THEORY-717, THEORY-718, QUESTION-137, QUESTION-136) get numbers when it merges.
• Run the triage against a freshly fetched origin/main, never a local main. Queue version 1 was scored against a stale clone-time main (b84ad2e3), because git checkout main || git fetch never fetched once the checkout succeeded. Always git fetch origin main first, then cut the branch from origin/main. The triage script used for version 2 is triage.py here.

Open threads.

• Dependency unlocks (wave 4). CLAIM-056 rests on CLAIM-267 and CLAIM-081 on CLAIM-270, but both also rest on Proposed claims (CLAIM-106 and CLAIM-056), so promoting 267 and 270 does not unlock them alone. CLAIM-189 and CLAIM-191 apply CLAIM-270; check them once it is Active.
• The next triage should add a column for the warrant/defeater mismatch.
• Wave 2 has more seeds than the six already done. Grep THEORY for promote_when texts containing “Done”, “met” or “confirmed”, which mark a half-met condition.

2026-10-19 — batch-002

Patterns.

• Counters are different. Three of five were role: counter. They carry no defeated_if, so the warrant/defeater test does not apply; read the objects_to target instead. Promoting a counter means the record is persuaded, so pair it with a correction to the opposed entry (CLAIM-313 → CLAIM-310). A counter whose remedy is renaming (CLAIM-251) is the owner’s call, not a proof.
• Dropped qualifiers from an Active source. Diff the claim’s statement against the THEORY it grounds on. CLAIM-308 lost THEORY-348’s “deterministic”, and randomized rent-or-buy does better (e/(e−1) in expectation).
• Exclusive “only” clauses. Test them against the standard example of the field. CLAIM-251’s “only the compatibility structure engages gluing” fails on CHSH, which carries both a PR box and product laws.
• A demonstrated core no Active entry holds. CLAIM-063’s approximate identity: ARG-006 (Active) holds only the exact case. Check the ARG that concludes a claim before deciding a split is redundant.
• A NOTE’s “my inference”. NOTE-600’s “consistently connected, by my inference” was a correct one-step derivation. Such remarks are cheap split cores.

Traps.

• Worktree-isolated sessions refuse compound shell commands with git. Run check, add, commit and push as separate calls, and do not commit after a failing check. It happened once here: queue version 4’s wave counts were stale, and version 5 fixed them. Recompute the wave counts on every queue save.

Open threads.

• Wave 4 now has three concrete items: CLAIM-310 after CLAIM-313; ARG-006 citing the CLAIM-063 split for the approximate case; CLAIM-081 and CLAIM-056 once CLAIM-270 and CLAIM-267 are Active.
• The next five pending in queue order: CLAIM-268, CLAIM-081, CLAIM-107, CLAIM-188, CLAIM-193. CLAIM-081 is better taken in the wave-4 pass after CLAIM-270.

2026-10-19 — batch-003

Patterns.

• ADR-040 parents. A claim whose note says “the general content is filed separately” (CLAIM-081 → CLAIM-270, CLAIM-193 → CLAIM-269) keeps only the carry-over to a domain. Its demonstrated warrant is then usually wrong. Review the general entry instead, and correct the parent’s warrant.
• An Active granted objection. If an Active claim with role: granted objects to the candidate (CLAIM-094 → CLAIM-107), the granted point is the narrowing. Write it into the candidate’s statement and promote the rest.
• Empirical disjuncts that cannot defeat. “Shown to agree on every pair tested” (CLAIM-268) does not refute a statement about what a method guarantees. Restate it as a derivation-type defeater rather than calling the claim compound.
• Locate the singularity. In model-reduction claims, check which model is singular at which parameter. For pruning, it is the model with the candidate, at the point where the candidate is switched off (THEORY-082), not the reduced model (CLAIM-188).
• Look for an Active substitute ground. CLAIM-269 grounds on THEORY-002 (Proposed) for “kernels fix representations up to symmetry”, but THEORY-004 (Active) proves the orthogonal-equivalence form. Before recording a dependency blocker, grep the Active THEORYs for the same statement.

Traps.

• Shell quoting: a note passed on the command line breaks on an apostrophe. Quote the whole note.

Open threads.

• Wave 1’s yield fell from 3/5 to 2/5, and the remaining wave-1 candidates are mostly ones with Proposed premises. Run a wave-4 pass next: CLAIM-269 (THEORY-004 and THEORY-008 re-grounding), CLAIM-310, ARG-006, and CLAIM-056 then CLAIM-081. Each is gated on an open PR merging (#165, #167, or batch 002’s follow-ups).
• Wave 2 seeding (THEORY promote_when texts with “Done”, “met” or “confirmed”) is still untried.

2026-10-19 — batch 004 (wave 4)

Patterns.

• Count the dependent’s other Proposed premises. Promoting a formal premise unlocked an entry only when nothing else under it was Proposed (CLAIM-322) or when the remaining premise could be swapped for an Active one (CLAIM-269, onto THEORY-004). Every dependent that also rests on an interpretive premise was retained.
• Find dependents mechanically. List the Proposed entries whose rests_on or grounds include a just-promoted code, with their other non-Active premises beside them. Twelve matched. The hand-off’s guesses included two Active entries (CLAIM-310, ARG-006), which can be corrected but not promoted.
• Pattern 1 again. CLAIM-056, like CLAIM-193, is a “demonstrated” application whose formal core was split off under ADR-040, and its defeat condition is empirical. Every ADR-040 application claim with a demonstrated warrant should be suspected of this. CLAIM-081 is the next one.

Open threads.

• Batch 004’s two promotion paths both wait on #165. Once it merges, one PR can split CLAIM-269 and promote CLAIM-322.
• Wave 2 seeding is still untried, and it is the better next batch. Wave 4 has little left with zero other Proposed premises.

2026-10-19 — batch 005 (wave 2 begins)

Patterns.

• The promote_when says the proved half is done. THEORY-174 (“already proved in refereed venues”), THEORY-168 (“already proved”) and THEORY-017 (“condition is met”) each named a finished half and held the whole entry for the rest. Grep for those phrases first. Triage-002’s keyword heuristic found them, and wave 2 went 5 of 5.
• An elementary proof can be checked in the review itself. THEORY-184’s promote_when asked for “the proof checked independently of this reading”. The proof is a triangle inequality on the pair (X, PX), so the review is that check. Write the computation in the decision’s evidence, so the promotion PR can cite it.
• The record’s own derivation, found in a THEORY, goes to a CLAIM. THEORY-165’s Lipschitz bound is “the reader’s derivation”. It is correct, and its LP duality needs nothing from the paper once the primal is shown bounded. This is the same split as batch-000’s THEORY-005.
• Read the claims table in the source NOTE. “asserted, no example” (C3) and “asserted, not computed” (C7) marked exactly what THEORY-168 must keep Proposed. “followed step by step” marked what could go Active.

Traps.

• In a worktree-isolated session, shell loops over $n and multi-command git pipelines are refused. Write git show origin/main:<path> > file one per call, then grep the files.
• The ledger branch’s working tree lags main. Read entries with git show origin/main:<path>, not from disk.

Open threads.

• Batch 005’s five follow-up PRs are not opened (batch-005.yaml).
• Wave 4 after them: CLAIM-237 could ground on THEORY-174’s monotone half. CLAIM-105, -235, -319 and -320 may need only THEORY-165’s Lipschitz claim. THEORY-018, THEORY-019 and CLAIM-148 may need only THEORY-017’s core, which bears on CLAIM-269’s THEORY-019 clause.

2026-10-19 — batch 006, in progress

• Wave-4 re-grounding candidates after #176. CLAIM-105, CLAIM-201, CLAIM-219, CLAIM-235, CLAIM-319, CLAIM-320 and CASE-037 cite THEORY-165 for the continuity of the contextual fraction. That result is now held by CLAIM-372, the Lipschitz bound split from THEORY-165 in #176. Once #176 merges, each one that needs only continuity can re-ground on the new claim.

2026-10-19 — batch 006 closed (wave 2 at 15 of 16)

Is the wave-2 yield real, or is triage-002’s heuristic picking easy cases? Mostly real, with a selection effect worth naming. Triage-002 scored promote_when texts for words like “met”, “checked”, “proved” and “first half”. Those words mark entries whose filer already found part of the condition satisfied. So the heuristic does select for half-met conditions, which is exactly the shape that splits. That is a selection effect, but it is the intended one, because a split is a real promotion of a real core. The evidence that the yield is not easy cases: in batch 006, three of four promotion paths needed the record to supply a missing step (a −δI shift, a contraction and occupancy-LP argument, a column-space generalisation), not just to read a “Done”. And the first entry whose promote_when bundles nothing, THEORY-277, did not yield. Expect the yield to fall as triage-002’s keyword hits run out (THEORY-171 onward mostly score on “checked” alone), and re-triage before treating 15 of 16 as the wave’s rate.

Patterns.

• A neutral example plus “persists with damping, not shown”. A uniform shift A − δI lowers every Lyapunov exponent by exactly δ, because Φ_δ(t) = e^(−δt)Φ(t). So any neutral-at-every-instant instability gives a stable-at-every-instant one. Worth remembering for other “stable at each instant” claims.
• “Equal because both are the unique fixed point of one contraction.” Check that the two operators really are identical term by term, then the identification is immediate (THEORY-247).
• A theorem stated at full rank often holds for a fixed column space. THEORY-236’s Theorem 1 is the case col = ℝⁿ of “a GL(p)-invariant index depends only on the column space”. That generalisation is what a “write it for the centred case” request usually needs.
• Re-derive the rank argument; don’t trust the summary. NOTE-693’s sketch (“a dependence among rows of maximal |S|”) does not close as stated. The working proof uses the zero lower-degree Taylor parts.
• A promote_when that asks only for an independent proof does not split. Register the independent source as a LIT, or put the THEORY-004 standard (a step-by-step reading suffices) to the owner.

Traps.

• In a worktree-isolated fork, a long python heredoc that writes files, chained with git commands, is refused as “too complex to verify”. Put the script in the scratchpad, run it alone, then run each git command separately.

2026-10-19 — batch 007: a control sample for triage-002

What the control showed. Five mathematics THEORYs whose promote_when has none of triage-002’s keywords, drawn by a fixed rule (sorted codes, every 8th of 43). Four of five found a promotion path, against 15 of 16 for the keyword group. That is not a distinguishable difference at these sizes, so triage-002’s ranking is weakly informative at best. The class predicts yield, not the keywords: a mathematical THEORY whose promote_when asks for more than its title claims.

What drives the control’s yield. Three of its four paths (THEORY-233, THEORY-297 and THEORY-323’s growth half) rest on short record-own proofs standing in for a requested reading or second source. The keyword group’s paths more often rested on readings already done. So a high wave-2 rate partly measures how often the record is willing to substitute its own proof. That needs the owner’s ruling (batch-007.yaml’s owner_question, and THEORY-277). Until then, every decision that substitutes says so in its notes.

Implications for triaging waves 3 and 5.

• Don’t rank by promote_when wording alone. Score instead (a) whether the promote_when names something the body’s “What this does not say” disclaims, which is the bundling that splits, and (b) whether the title is mathematics with short textbook proofs. Both can be read off the entry in one pass.
• Wave 3 (source-backed, non-mathematical THEORYs) can’t use the record-own-proof route. Its yield will depend on whether the NOTE says the source was “read in full” or “followed in outline”. Score that from the NOTE’s history line, and expect THEORY-106’s shape (theorems read only in outline) to be common.
• Wave 5 (interpretive synthesis) should be sampled, not ranked, the way this batch did. No textual heuristic has yet been shown to predict yield, and a control is cheap.
• Always draw a small control alongside a ranked batch when a wave starts. Five entries was enough to show the ranking was not doing the work.

Patterns.

• A title broader than its own body. THEORY-009’s title claims “spectral SSL” while its body lists three routes (symmetrized loss, kernel precondition, composition with an adjoint) that contradict it. Read the body’s exceptions against the title.
• A promote_when about something the THEORY disclaims. THEORY-233 (trained-network magnitude), THEORY-297 (truncation stability) and THEORY-009 (asymmetric objectives) each ask for evidence on a point the body says it does not claim. Move that clause to a QUESTION and judge the title on its own.
• “Cited to” inside a THEORY’s own list of results. THEORY-323’s white-noise inequality is cited in the source to an unread paper. Grep the body for “cited” before promoting a list of results whole.

2026-10-19 — the owner’s rule on re-derivations (after batch 007)

Asked whether a record-own re-derivation meets a promote_when that asks for a proof the record does not hold, the owner chose to promote the THEORY (ADR-047 v2). The alternatives were filing the proof as a CLAIM only, or holding the entry. In practice, write the derivation out step by step in the decision file, as batches 006 and 007 did, and cite it in the history note. THEORY-277 is requeued on this basis.

2026-10-19 — batch 008, the first batch under ADR-047 v2

• Re-derivation works for short, self-contained proofs. It closed THEORY-160’s bracket equivalence, THEORY-158’s finite-state cycle and Basu’s theorem in a Markov category, each in a few lines. Write it in the decision file under scope.accepted, so the promotion PR can copy it into a dated note.
• It stops at constructions. THEORY-277’s Lemma C.4 reduced cleanly, but the covering construction it rests on did not close in two rounds. The trap is an enumerable set chopped into blocks in enumeration order, whose last block never provably fills. When a proof uses “chop the enumerated objects into blocks”, check how it handles the final block before trusting the bound. Record the blocked step exactly, so the next attempt needs only one retrieval round.
• Lost diagrams are a reason to re-derive. NOTE-588 could not check Fritz’s string-diagram proofs, which did not survive text extraction. An algebraic re-derivation from the prose definitions recovers the check. State the definitions relied on, because they too were reconstructed from prose.
• An existence title asked for a characterisation (THEORY-171). One checked example settles an existence statement. When the promote_when says an example “does not count”, it is asking about a stronger claim than the title makes. That calls for a narrowing.
• Keep a superseded decision traceable. When a candidate is re-reviewed, overwrite its decision file and record the earlier disposition in supersedes_review. check.py allows one file per code.

2026-10-19 — batch 009

• A third scope error in a correct theorem. THEORY-686’s title says the observations with no generative effect are exactly the left adjoints. The absence of generative effects is about binary joins, while Thm 1.115 needs all joins. A constant map to ⊤ has no generative effect and is not a left adjoint. That makes three after CLAIM-264 and CLAIM-123. When re-deriving under ADR-047 v2, test the title’s quantifiers against the smallest degenerate case: the empty join, a two-element order, a constant map.
• Frames from Stonean spectra. THEORY-037’s Heyting structure follows in a few lines once each stage is extremally disconnected. Closures of open unions are then clopen, and S ∩ cl U = cl(S ∩ U) for clopen S. A promote_when that rules out “Sub(X) is Heyting in any topos” is met by an argument internal to the clopen sub-objects.
• Full readings can already hold the core. THEORY-002 needed no re-derivation. NOTE-738 read every proof in App. B, C and E. Its title simply bundled them with a reading of the Platonic hypothesis. Check a NOTE’s version-1 history line, which says what was read, before re-deriving anything.
• The Dietrich–List result is beyond a two-round pass. Proposition 3 for the court agenda needs path-connectedness and a decisive-coalition argument. It is requeued as a dedicated pass, not deferred.
• Wave 2’s triage-002 list is exhausted except for THEORY-565. The yield stays high (27 of 30), so re-triage the remaining mathematics THEORYs by bundling (batch-009.yaml recommendation) before moving to wave 3.
• A stale directive was left on purpose. CLAIM-107 still has a THEORY-158 inactive-ok directive. #183 left it in place because #173 rewrites that file. Trim it once both #173 and #183 have merged.

2026-10-19 — triage-003 and batch 010

• Bundling scores saturate. triage_bundling.py gives the full 3 to 27 of 34 eligible mathematics THEORYs, so its ranking is mostly the tie-break. Batch 010 went 4 of 4 on its top four, which is no evidence the ranking helps: the keyword list went 15 of 16, and the keyword-free control 4 of 5. Within this class, review in code order and skip re-triaging.
• A promote_when can be mis-targeted, not bundled. THEORY-014’s condition asks for a bridge between formalisms that its body explicitly disclaims. The title was fully supported by two Read sources. The remedy is promote, with the condition moved to a QUESTION. No split is needed, because nothing in the title is unsupported.
• Recurring cores. Four of five cores this batch were short linear-algebra or spectral facts that sources state and call standard: the diffusion-map isometry (needs the 1/π weighting), the Fisher–NTK spectral identity (needs the model Fisher, not the empirical one), weight decay as a nuclear norm (needs the rank bound from the inner dimension), and tight frames and POVMs (needs equal norms within each factor). Each holds only with the hypothesis in brackets, so the narrowed statement must carry it.
• Worktree isolation. A worktree-isolated fork can’t run compound shell commands, such as git piped into tar or a heredoc after cd. Write helper scripts to a scratch directory inside the worktree (here .tmp-main/, not committed) and run them one at a time. git archive -o file.tar origin/main record followed by tar -xf gives a clean copy of main’s record for triage.

2026-10-19 — merging the audit’s PRs (#165–#188)

• Merge one PR at a time. The Docs workflow has no concurrency: group. Merges seconds apart start several docs-generate runs at once. Each concretizes the same tmp codes with different numbers, and the runs that lose the push race fail on rebase: the main runs for #182 and #184 did. Nothing broken landed, because a later run concretized everything. Still, if the last run to start is one that loses, its tmp codes stay on main until the next push. Wait for each merge’s generate step (its “docs: regenerate views” commit) before merging the next. A concurrency group on the workflow would remove the race; that is the owner’s decision.
• GitHub’s “has merge conflicts” can be stale. After main’s concretization commit, #181, #183 and #186 were refused as conflicted while git merge-tree --write-tree exited 0, and the refusal did not clear on retry. Merging origin/main into the branch (a plain merge commit, no conflicts) and pushing reset it. Check the merged tree for references to tmp codes that no longer exist before pushing.
• Conflicts are almost all inactive-ok lines. Several PRs trim the same directive. Resolve by position, and keep a code only if every side keeps it, so every trim survives. Where line counts differ, resolve by hand.
• Bot repair commits on a branch (“docs: repair sources”) leave a run at action_required. CI on the commit beneath is what counts. Read the repair’s diff before merging. So far every one has only added links.

2026-10-19 — batch 011 (wave 2)

• The tag misleads triage. triage-003 drew wave 2 from THEORYs tagged “mathematics”, but THEORY-022, -191 and -192 are empirical. Interpretability and Betti-curve studies carry the tag for their tools. Classify by what the title asserts: a theorem, or a measurement.
• Bounded Gromov δ from a tree decomposition with clique bags. If a graph has a tree decomposition whose bags have diameter ≤ 1, its geodesic triangles are 2-thin. Take the node of the tripod spanned by the three corners’ bags that is nearest a bag of p; its bag separates p from the two corners. This settles hyperbolicity for any complex grown by gluing on single faces (stacked polytopes, k-trees, Apollonian networks), and is worth reusing when a promote_when asks for “δ bounded independently of N”.
• “A tree of X tiles H^d” fails for d ≥ 3. In a face-to-face tiling, tiles around a codimension-2 face form a finite cycle, and a tree of simplices has none. It holds in d = 2 only because the codimension-2 faces of ideal triangles are at infinity. Check any “tiling by a tree of cells” claim this way.
• “The dual is a tree” depends on the flavor. With preferential or uniform gluing (s = 0, 1), several simplices share a face, so the face-adjacency graph has cliques. The gluing or parent structure is always a tree; say which one is meant.

2026-10-19 — batch 015 (wave 4, dependency pass)

• Wave 4 counts: 5 reviewed, 1 promotion path (CLAIM-269 split).
• CLAIM-107 was decided in batch 003 (narrow-and-promote, carried out in #173); its decision file is overwritten with supersedes_review, disposition needs-correction. CLAIM-193 (batch 003, needs-correction, carried out) and CLAIM-269 (batch 004, split) likewise. The queue entries for these three need their batch and disposition updated.
• CLAIM-148 and CLAIM-164 are new to the queue: add them as wave-4 candidates, decided.
• THEORY-127’s dependents, for a later batch: CLAIM-192 (Proposed, interpretive) and CLAIM-271 (Proposed, interpretive) use only the court- agenda possibility, that premise-wise and conclusion-wise majority can disagree, so each can re-ground on THEORY-725. CLAIM-030 (Active, interpretive) cites ‘a rational group cannot judge proposition by proposition from its members’ judgments’, which is the general impossibility (Propositions 2–3), still THEORY-127’s; THEORY-725’s Proposition 1 is one instance of it, enough as a formal anchor if the text says ‘can fail to’, not ‘cannot’.
• Pattern: an Active claim with a Proposed ground and an inactive-ok directive is a cheap wave-4 find after a split. Grep for directives naming a split parent.

2026-10-19 — batch 017 (wave 3, control)

• Wave 3’s control: 2 of 5, both formal. Non-formal entries 0 of 3. Triage-004’s components did not discriminate.
• triage_wave3.py’s depth takes the weakest source, which hides a full reading of the source that holds the core (THEORY-078: depth 0, though NOTE-479 read the derivations in full). Consider taking depth from the NOTE of the source that the accepted clause cites, or the maximum over sources.
• triage_wave3.py’s firsthand misfires on entries whose ‘What was actually shown’ opens ‘Nothing is measured’ (THEORY-470). Add negative phrases such as ‘nothing is measured’ and ‘no theorems’.
• The mathematics tag under-marks formal results. Triage-004’s own report of untagged formal entries (THEORY-248, -369, -198, -315) and this batch’s THEORY-078 and THEORY-309 point the same way.
• Annealed or mean-field derivations (THEORY-199) are a recurring non- yielder. The source proves the approximation exactly, not the title’s claim about the exact model. Read the claims table for ‘strong for the map’ or ‘numerical fidelity’.

2026-10-19 — batch 014

• Check which direction of a Proposed ground is used. Three claims listed THEORY-156 (general Blackwell converse) as a dependency. Two used only the definition of a garbling, and one used the converse only outside finite outcome spaces, where THEORY-717 (Active) covers the rest. Grep the body for the sentence that cites the ground before recording a dependency blocker.
• A sign question in a THEORY is a cheap re-derivation. THEORY-312’s promote_when asked only to settle a remainder’s sign. Lemma B.0.1 closes on three intervals using (1 + x/2)² ≤ e^x on [−2, 0], and Theorem B.1 needs only log(1 + x) ≥ x/(1 + x) and E Var − HSIC(Z,Z) = Var μ₁.
• Pattern 1 recurs in the fidelity line. CLAIM-212 and CLAIM-228, both ‘the record’s reply of 2026-10-10’, are demonstrated over a reception- study defeater. Expect the other 2026-10-10 replies in pragmatic- transport to have the same misclassification, and check them in wave 5.

2026-10-19 — batch 012 (wave 2)

• Exact hyperbolic trigonometry beats small-angle substitutions. cosh d = cosh(r₁ + r₂) sin²(θ/2) + cosh(r₁ − r₂) cos²(θ/2) turns an angle-packing lower bound into two lines with constants (THEORY-196).
• Minimax uniqueness is fragile. A worst case is attained by one adversarial target, and any encoder that is perfect on that target’s plane ties. Test ‘the unique minimizer’ with a small tilt and the S-lemma dual, min over λ of λ_max(R + λ(S − cI)), which is a few lines of numpy (THEORY-239).
• Counting arguments: the negative half (more independent equations than unknowns, so the solutions are null) becomes a proof via the incidence- bundle dimension. The positive half (balanced counts, so solutions exist generically) can fail over the reals on open sets. A 2 × 2 rotation is the standard witness (THEORY-257).
• A fourth and fifth scope error in correct theorems (THEORY-239’s uniqueness, THEORY-234’s ‘only when’). Read ‘unique’, ‘only’, ‘exactly’ and ‘generic’ in a title against the body’s disclaimers before re- deriving.
• Triage-003’s eligibility admits empirical THEORYs about mathematical objects (THEORY-254, tangles). Screen the title’s verb.
• Worktree isolation: python one-liners with file writes through the Write tool to a path outside the worktree are refused. Write scratch scripts with a heredoc into a scratch directory inside the worktree (here .tmp-b012/, not committed). scipy is not installed, so a grid over λ replaces minimize_scalar.

2026-10-19 — batch 013 (wave 2)

• Fixed-point subspaces settle symmetry-cluster claims. For an equivariant vector field, Fix(Σ) is invariant for every subgroup Σ. Isolated desynchronization is Fix(K), with K the product of the other factors, and it needs the moving factor’s support to be the cluster (THEORY-276). Use this instead of equal-input computations.
• An empirical THEORY can hide a record-own invariance-in-law argument (THEORY-288). With equivariant initialisation and optimiser, the law of the learned representation inherits the architecture’s symmetry group. The argument needs an orthogonally equivariant optimiser. Adam breaks it, leaving only signed-permutation invariance. Check the optimiser before filing any such claim.
• Numerical scaling THEORYs (THEORY-304) yield their lower bounds to a rescaling argument (lift-up at streamwise wavenumber 0). Their constants and sharpness need the computation. File the lower bounds as a CLAIM and keep the parent.
• A fourth correct-theorem, wrong-scope case (THEORY-299), after CLAIM-264, CLAIM-123 and THEORY-686. An identification exact for the source’s normalised variant (squared norms) is read as covering the original models. Check which variant a source actually works with.
• Re-deriving a promote_when’s lemma chain can remove a Proposed dependency. THEORY-301’s Cor. 1 is one line, so its promotion does not rest on THEORY-008.

2026-10-19 — batch 016 (wave 3, ranked)

• Wave 3’s yield comes from its formal members. Of five top-ranked entries, the two that yielded were formal results without a mathematics tag, promoted by re-derivation. The three empirical ones were held by measurements the record has not read. Expect that pattern to continue, and route untagged formal entries to wave 2.
• Look first at the step a NOTE took on trust. THEORY-481’s NOTE-658 said the Wishart fourth-moment step was ‘followed as statements, not re- derived’. Re-deriving it found the source commuting the rotated empirical covariance Ŝ, which is not diagonal at finite N. That makes its per-mode decoupling and s^∞ exact only as N → ∞, or for isotropic inputs. Grep NOTE history lines for ‘followed as statements’, ‘not re- derived’ and ‘taken as stated’ before trusting an ‘exact’ claim.
• A concentration argument settles high-dimensional nearest-neighbour bounds. For fixed N and uniform data, all squared distances are d/6 + o(d), and the midpoint of any two points is at d/8 from the query. So every convex cluster of at least two points meets the nearest-neighbour ball with probability tending to 1. The same trick may serve THEORY-366 (Beyer et al.’s distance concentration).
• triage_wave3.py’s firsthand markers miss ‘stated by no source’, ‘the filer’s synthesis’ and ‘nothing was tested’. Its depth score reads ‘read in full’ per NOTE, not per claim. Both fired falsely in this batch.
• A source-derivation error is a needs-correction even when the record read the source faithfully. Correct the THEORY with status unchanged, add a dated note to the NOTE, and requeue the THEORY for narrow-and- promote, so a status change never rides on a correction.

2026-10-19 — the parallel round (batches 011–017), summary

• Yields. Wave 2’s last 16, in code order: 11 of 16 (011 at 2 of 5, 012 at 4 of 5, 013 at 5 of 5, plus THEORY-312 in 014). Wave 4 (015): 1 of 5, with three corrections to Active entries. Wave 3: the ranked top five (016) and the control five (017) each went 2 of 5. So triage-004’s ranking did not predict yield. All four paths were formal results that carried no mathematics tag (THEORY-111’s core, THEORY-369, THEORY-309, THEORY-078’s algebra).
• The tag is the wrong class boundary. Batch 011 found wave-2 entries that are measurements (THEORY-022, -191, -192). Batches 016 and 017 found wave-3 entries that are theorems. Classify by what the title asserts.
• The owner’s rule on single studies (ADR-047 v3). One study read in full is enough for an empirical THEORY whose title claims no more than that study measured. The replication request moves to a QUESTION. This should raise wave 3’s yield on the hedged empirical entries, and it leaves causal or generalising titles, such as THEORY-470, where they are.
• Trap: history order. A new history: entry goes last, not first. luria lint reads the last entry as the current version and fails with “history: ends at version N but the document says N+1”. Four PRs from this round prepended entries (#190, #193, #201, and ADR-047 v3 on the ledger) and were fixed by reordering. Also, a CLAIM’s grounds takes only THEORY, CASE or LIT codes; another claim used as a worked example is cited in the body, not listed (#196).

2026-10-19 — batch 022 (wave 5 opens, fixed-rule sample)

• Wave 5’s population (sample_wave5.py): 219 codes, 179 Proposed interpretive CLAIMs and 40 multi-source non-mathematics THEORYs. The THEORY half of the definition (at least three distinct LIT sources) is a proxy for “synthesis” and was not validated; THEORY-058 fit it well.
• An interpretive warrant over an observational defeat condition is pattern 1’s mirror (CLAIM-052). The fix is a warrant correction to empirical, not a split, when the claim is a single generalization that is simply untested. ADR-043’s “empirical” covers untested hypotheses whose support would be observational.
• An existence title with a documented case is the cheapest promotion path in every wave so far (THEORY-171’s in batch 008, CLAIM-328 here). For interpretive claims it still goes to the owner.
• Claims in one argument line should be decided root first. CLAIM-217 cannot move before CLAIM-151, which cannot move before CLAIM-004 and CLAIM-061. Sampling such claims independently spends reviews on entries that are blocked by construction.

2026-10-19 — batch 021 (wave 4, THEORY-127 dependents)

• Wave 4 counts: 3 reviewed (CLAIM-192, -271, -030), 0 promotion paths. CLAIM-206 and CLAIM-126 are wave-1 candidates: 2 reviewed, 0 promotion paths.
• All five decisions are new; none supersedes an earlier review. Remove CLAIM-192, CLAIM-271 and CLAIM-030 from the requeue list.
• Pattern 1, a demonstrated warrant left over a split-off formal core, has now appeared in five ADR-040 parents: CLAIM-081, -193, -056, -269 and -126. Check every ADR-040 parent still marked demonstrated, in one sweep.
• Pattern: dependents of a split parent usually cite its general result only for an existence statement. ‘Cannot always’ or ‘can fail to’ is settled by one instance, so the split’s core is enough.

2026-10-19 — batch 023 (THEORY-127, Proposition 3)

• The ultrafilter argument fits in a decision file. For a judgment aggregation impossibility, the work is finite. - List the agenda’s minimally inconsistent sets. - Show conditional entailment connects all literals. - Find one minimally inconsistent triple that stays consistent with two of its members negated. The rest is the standard finite- ultrafilter lemma. Reusable for any “no independent, unanimous, non- dictatorial rule” claim in the record.
• A general title can be false although its worked case is a theorem. “Interconnected propositions” fails on affine agendas, for example {p, q, p↔q}, where issue-wise parity is a non-dictatorial, independent, unanimous, robustly rational rule. When a THEORY states a social-choice impossibility for a vaguely named class, test it on an affine agenda before promoting the general wording.
• The record holds no Dietrich–List paper, and a grep for “Dietrich” finds an unrelated LIT-912 (flow). Under ADR-047 v2 the re-derivation stands in for it.
• Trap: codes inside written paths. luria repair turns a code inside a path such as record/promotion-audit/decisions/THEORY-127.yaml into a link, which breaks the path. In record entries, cite a decision file as “the promotion audit’s decision file for THEORY-127”, not by path. QUESTION-140, -141, -147 and -148 on main already carry broken paths (found by batch 023, not yet fixed).

2026-10-19 — batch 020 (wave 3)

• Triage-004’s top scores carry no information: five 7-of-7 entries gave one promotion. Sort wave 3 by what the title asserts (theorem, measured result, argument) instead.
• ADR-047 v3’s first use. A model-run result, in hedged existence form and read in full, promotes once its scope is narrowed to the model actually run (THEORY-295: a memoryless single-test consensus). Expect more of these among the social-epistemology simulations.
• Exclusive “only” again (THEORY-546). Test a title’s “only when” or “requires” against the dilemmas its own body reports. The body named multiple occupancy as the alternative to non-branching.

2026-10-19 — batch 018 (wave 3, ADR-047 v3)

• First use of ADR-047 v3. “Read in full” was taken to allow figures and supplements unseen, provided the title rests only on results the text reports. Both NOTEs (672, 668) qualify. A title resting on a supplementary table known only by summary would not.
• A promote_when that asks for a confirming instance of a logical consequence (THEORY-255: “a second model that also places the deviation outside the meet”) is met by the proof of the consequence. What remains, which published models drop which assumption, is a QUESTION.
• Triage-004’s maximum score (7) went 2 of 3 here and 2 of 5 in batch 016. The three 7s reviewed here split along formal content, which the rubric doesn’t score. Batch 017’s formal-content flag is the fix.
• THEORY-235 should move to wave 5 in the queue: a single-source philosophical argument with open objections.

2026-10-19 — batch 019 (wave 3, formal)

• ADR-047 v3 needs the measurements, not only the reading. NOTE-816 (THEORY-292) and NOTE-758 (THEORY-256) both say “read in full” and “figures read through captions only”. The difference was that LIT-986 states its numbers in the text and LIT-937 does not. Grep the NOTE’s history for “captions” and check that the title’s numbers appear in the extracted text.
• A short derivation of a preparation-noncontextuality bound without outcome determinism. Put u = μ1 − μ3 and v = μ4 − μ2. The equivalence gives |u|, |v| ≤ ν pointwise, and |σa + τb| + |σa − τb| = 2max(|a|, |b|) ≤ 2. Together these give σ⟨X⟩ + τ⟨Z⟩ ≤ 1. Dropping the equivalence gives 2, the gbit. The same template may serve other symmetric-orbit inequalities.
• Screening lemmas from full capture. If I(H:E) = I(V:E) with H drawn from V alone, then I(V:E|H) = 0, and any separation statement that holds given V passes to H. This is the core of THEORY-198, and it is a general tool for information-bottleneck claims in the record.
• “Stays finite” in a bound’s title should name its finiteness condition. THEORY-248’s Lipschitz bound is finite exactly when the output law’s rate-distortion function is finite at ε/(2L).

2026-10-19 — round 018–023, summary

• Yields. Wave 3: 018 at 4 of 5, 019 at 4 of 5, 020 at 1 of 5. Wave 4: 021 at 0 of 5, with four corrections. Wave 5: 022 at 1 of 5, the owner’s call. THEORY-127 (023): narrow-and-promote, after Proposition 3 was re-derived. Every wave-3 path was a formal result or a measurement read in full (ADR-047 v3). Arguments and conceptual analyses went 0 of 7.
• Sort wave 3 by what the title asserts (batch 020). A theorem goes to re-derivation, a measurement to v3, and an argument to wave 5. Triage-005’s route flag does this. Its score does not.
• v3 needs the numbers in the text. “Read in full” excludes results that the NOTE saw only in figure captions (THEORY-292).
• Warrant correction is the next cheap work. Wave 5’s first sample promoted almost nothing. It found misclassified warrants, and a mechanical count suggests 36 more. ADR-040 parents marked demonstrated are a second sweep (pattern 1, five cases now).

2026-10-19 — batch 028 (warrant-correction sweep)

• Two readings of ADR-043 are in use. The pilot’s: an empirical claim with no observation under it is interpretive (CLAIM-117, CLAIM-063, CLAIM-056). Batch 022’s: it is empirical (CLAIM-052, #204). Until the owner rules, don’t correct warrants in either direction on that ground alone.
• Term-of-art false positives. “Empirical model” (contextuality) and “experiment” (Blackwell’s statistical experiments) fire the cue list without naming anything observed. warrant_sweep.py excludes the first; the second needs a reading.
• Check the defeater’s quantifier against the title’s. An existence title (“can”) with a frequency defeater (“at a rate no greater than”) is a mismatch, but read the history first: CLAIM-238 (Active, granted) is why the 2026-10-10 restatements turned possibility defeaters into prevalence tests, so the fix may be the title, and that is a design choice.
• A title already corrected by its own body. CLAIM-074’s body says the transient/persistent dichotomy in its title “does not hold”; the title is unchanged. Grep bodies for “does not hold” or “needs stating” in a later pass.
• Split parents and earlier corrections reappear in any mechanical sweep (CLAIM-056, CLAIM-063). Skip codes whose last decision was carried out unless their entry changed since.

2026-10-19 — batch 024 (wave 3, re-derivation route)

• The re-derivation route went 5 of 5. Every title asserted a theorem, and each disposition separated a proved core from an extension to trained models, random starts or real data.
• A dual certificate is a short route to max-margin geometry claims. For min ‖X‖_* under margin constraints, the polar factor of the candidate is a certificate whenever its columns are nonnegative combinations of the constraint normals, which reduces to a sign check on (ẐẐᵀ)^{+1/2} (THEORY-286). Reusable for other unconstrained-features results.
• “Linear attention” in a title should say whether it is normalized. Sun et al.’s identity holds for the unnormalized form only (THEORY-169).
• Logistic per-mode dynamics recur: THEORY-182 (tied symmetric factorization) and THEORY-186 (two-layer linear) reduce to u̇ = 2u(λ − u) once the start is aligned. A split core of this shape takes a few lines to re-derive.

2026-10-19 — batch 025 (wave 3, re-derivation route)

• A reconstruction can show that a toy is not recoverable from the record. For THEORY-320, rebuilding the toy graph from NOTE-833’s description reproduced Theorem 5.2 exactly, but not Theorem 5.3’s threshold or Lemma 5.4’s harm case. That shows the set-up is underdetermined, not that the paper is wrong. Record the reconstruction and its mismatches in the decision file, so the next reading of appendix B can settle it in one round.
• The Donsker–Varadhan approximation gap closes by two-sided truncation. The negative part of log dP/dQ is P-integrable whenever MI is finite: ∫ p log(q/p) over {p < q} is at most 1, by log x ≤ x − 1. The estimation gap closes by weak convergence of the empirical measures, uniform over an equicontinuous family. Both are reusable for other variational-bound estimators (NWJ, InfoNCE) the record may hold.
• Representer theorems over the completed RKHS need only projection onto a finite span and strict monotonicity. Existence needs lower semicontinuity and coercivity, and a counterexample without lower semicontinuity takes one line. Use this template when other kernel results are stated over expansion classes rather than H_k.

2026-10-19 — batch 026 (wave 3)

• A conditional theorem should be promoted on the hypotheses its proof uses, not the ones its source names. THEORY-478’s source assumed ‘a tree heat kernel’. The proof needs only a Hölder bound and a separation bound on Hellinger distances, and the admissible range is α < θ/β. The source’s range (θ = 1) and the reader’s repaired range (θ = 1/2) are then two instances of one theorem, not a dispute.
• Multiscale sums of the form Σ_k 2^(−kα) h_k with a Hölder regime below a crossing scale and a separation regime above it are equivalent to d^(αβ) exactly when α < θ/β. The coarse-scale geometric sum is the only place the exponent condition enters. A reusable template for multiscale diffusion distances.
• ‘No Bell-like no-go in a prepare-measure scenario’ has a one-line proof: send λ = (x, a). Any distribution with Y independent of (X, A) is then realized. Worth reusing wherever a source states a realizability claim without derivation.
• Triage-005’s formal flag, read against content: one of three doubted flags was right. The doubt followed the subject tags (philosophy, quantum foundations), not whether the title relates defined notions.
• THEORY-614 is the first re-derivation-route entry to stop at one unread appendix step. Record the step exactly (likelihood summed over delimiter states) so that a single retrieval can close it.

2026-10-19 — batch 027 (wave 3, missed formal)

• Triage-005’s prose-title misses were real formal results (3 of 4). Its formal flag reads titles; these four have formal content in the summary and body (“exact”, “conserved”, an SDE equation, a composition identity). Score the summary too.
• An existence clause is proved by a constructed example, even inside a paper’s own restricted protocol. THEORY-176’s “can stitch differently” took a two-position, one-channel example in which a 1×1 stitcher cannot undo a position swap. When a promote_when asks for a measured instance of a “can”, build the smallest case first.
• Narrowing to the derivation’s hypothesis is the commonest fix for formal THEORYs read off a paper’s theory section: THEORY-187 (decoupled, balanced start) and THEORY-211 (the SDE model, both groups noisy, product at stationarity). The title’s law is kept and its scope is stated.
• Promoting THEORY-176 removes one Proposed premise from CLAIM-126, CLAIM-225 and CLAIM-199. A wave-4 recount should follow once the PR merges.
• ADR-047 v3 with captions-only figures: THEORY-195’s results are stated in the paper’s text, so batch 019’s rule is met, but the curve collapses were not seen. Medium confidence is the honest record of that.

2026-10-19 — batch 029 (ADR-040 parent sweep)

• The ADR-040 parent population is closed: all eight parents are now interpretive once CLAIM-081’s correction merges. Pattern 1 needs no further sweep over ADR-040 parents. Watch new splits instead: when a split leaves the parent only a reading of a domain, correct the warrant in the same PR.
• A regex for “empirical” defeat conditions misfires on terms of art (“empirical model” in contextuality). Read the defeat condition before counting it. Batch 022’s count of 36 interpretive claims with empirical- sounding defeat conditions may hold similar false positives.
• When a decision defers a warrant review until a premise is decided, as batch 003 did for CLAIM-081, record the trigger in the requeue list. Otherwise it waits until a sweep finds it.

2026-10-19 — batch 030 (wave 5, roots)

• The roots option in sample_wave5.py (—roots_only) drops population members whose rests_on premises include a Proposed interpretive claim, queued or not. Of 212 in the population, 159 are roots.
• Batch 022’s mechanical count of interpretive claims with empirical defeaters over-reports. An interpretive claim can rightly have an observational defeater when its support is argument (CLAIM-327). A warrant is misclassified only when the support, not the defeater, is observational (CLAIM-052). Before a correction sweep, read each one’s support, not only its defeated_if.
• Decide argument clusters root first in one owner question. CLAIM-046 sits under CLAIM-206 and CLAIM-217, and CLAIM-241 objects to CLAIM-217. Ruling on CLAIM-206 against CLAIM-217 without the other two would leave both of them stale.
• A theorem read in full can still hide inside a multi-LIT synthesis (THEORY-109’s Theorem 1). Check every wave-5 THEORY’s NOTE claims table for “strong (proof)“.

2026-10-19 — round 024–030, summary

• Yields. Wave 3, re-derivation route: 024 at 5 of 5, 025 at 5 of 5, 027 at 5 of 5 (the formal entries triage-005 missed), and 026 at 2 of 5. The three entries whose formal flag triage-005 doubted split one promotion, THEORY-231, and two retentions. So the route predicts yield well, and its errors run toward missing formal entries rather than over-flagging them.
• Sweeps. The warrant sweep (028) found no miswarranted claims in 10. The ADR-040 parent sweep (029) found one, CLAIM-081. Both are mostly done. Under ADR-043 v2, the 37 interpretive claims with empirical defeaters are correctly labelled.
• The owner’s warrant rule (ADR-043 v2, #221). warrant records how a claim is supported now. An untested empirical claim is interpretive, and its defeat condition can still be an observation. CLAIM-052 is reverted. Do not relabel a warrant only because the defeater is empirical.
• Wave 5 roots (030). Drawing roots of argument lines worked: 3 of 5 had a path. Two of them are interpretive promotions, so they go to the owner.

2026-10-19 — batch 031 (wave 3, source route)

• Sort class predicted all five outcomes in batch 031: two arguments retained, one single-study measurement promoted under v3 (THEORY-216), one cross-study generalisation retained (THEORY-115), one theorem-class entry retained (THEORY-116). Triage-005 scored all five at 6, which carried no information.
• A theorem-class title can state the direction its source does not prove (THEORY-116: the title is LIT-669’s Conjecture 1; App. F.3 proves the converse). Check direction before re-deriving, and grep the Active THEORYs for the converse. Here THEORY-732 (iii) already holds it.
• When a narrowing excludes a universal clause, grep the dependents’ bodies for that clause, not only their grounds. CLAIM-273 cites THEORY-216 for “the lasting artworks are the ones whose coalitions keep placing under overwriting”, which the narrowing drops, so its bullet is restated in the same PR.
• Queue: THEORY-073 and THEORY-223 move to wave 5 (argument class). THEORY-115 and THEORY-116 stay retained in wave 3. THEORY-216 is decided, narrow-and-promote.

2026-10-19 — batch 032 (wave 3, source route)

• Wave 3’s source route, first five: 1 of 5, and that one (THEORY-324, Ashby) was a theorem in prose. Triage-005’s formal flag missed it. Flag prose deductions (“only if”, “necessary and sufficient”, “every equilibrium is”) as formal.
• A trigger condition in a sufficiency theorem. “Changes the parameters when the variables leave their limits” is ambiguous between a jump on crossing and a change for as long as they stay outside. Only the second makes every equilibrium adapted. A jump on crossing can land in a field with an unadapted equilibrium and stay there. Check “when … leaves” wording in any feedback or homeostasis title.
• The depth score reads a full reading of a secondary source as full depth (THEORY-560, a narrative review; THEORY-411, an encyclopedia entry). ADR-047 v3 needs the study itself. Treat results a review reports as cited, not read.
• Necessity relative to a rule. Ashby’s “only if” holds for systems that follow the trial-and-error rule, and becomes nearly definitional once the rule is assumed. A fixed regulator stable across the whole environment class needs no channel. When a promote_when widens an “only if” to “any system that achieves X”, look for a regulator that achieves X without searching.

2026-10-19 — batch 033 (wave 3, triage-005 control)

• Triage-005’s control: the route flag predicted 2 of 4. It was right on both arguments (THEORY-071, -485) and wrong on both measurements (THEORY-178, -306), each routed to source and promoted under ADR-047 v3. Sort by what the title asserts (batch 020). The flag catches theorems only.
• empirical_hedged scored 0 on both yielders. THEORY-178’s measured signature is a model ensemble that the title does not name, and THEORY-306’s contrast (MAE and data2vec) sits in the title’s last clause. Read ‘What was actually shown’ for a model run or a table, not the title.
• Thomas’s rule is not enough for synchronous Boolean necessity claims. A free negative cycle (a’ = ¬b, b’ = a, period 4) can be phase-shifted by a transient clamp. Beside any other oscillation, that gives a new joint attractor with no positive circuit in reach. Check for it whenever a Boolean attractor claim says “only through a positive circuit”. THEORY-178’s condition holds in its network only because the network has no such oscillator outside the reach of positive circuits.
• An “explicit in X” clause in an empirical title can be a one-line identity (THEORY-306: −H(p̄) = D_KL(p̄ ‖ U) − log K). Check it before treating the title as measurement only.
• Move THEORY-593, THEORY-071 and THEORY-485 to wave 5: each is an argument or conceptual analysis read in full, with nothing to measure or re-derive.

2026-10-19 — batch 034 (possibility theses)

• Batch 028’s CLAIM-033 and CLAIM-115 decision files were on its unmerged branch at this batch’s revision. This batch’s files replace them, with supersedes_review naming batch 028. When both branches are integrated, keep batch 034’s files (check.py allows one file per code).
• A defeater restated in place leaves its title behind. Whenever a history note says a defeated_if was restated, check the title in the same review. All three of the 2026-10-10 in-place restatements CLAIM-238 prompted left a possibility title over a prevalence test.
• A split parent can keep the split-off core’s quantifier (CLAIM-105: its “can” is CLAIM-320’s, since 2026-10-15). After an ADR-043 or ADR-040 split, the parent’s title should say only what the parent keeps.
• CLAIM-238’s comparative half was answered with new entries (CLAIM-220, -202, -204), not by restating the old ones. So CLAIM-115 and CLAIM-139 still carry marginless defeaters by design. Don’t correct them; whether CLAIM-220 supersedes CLAIM-115 is the owner’s open decision.
• CLAIM-139 and CLAIM-143, also named by CLAIM-238, are unreviewed for defeater fit. CLAIM-143’s second disjunct is an impossibility, like CLAIM-077, and probably fits; CLAIM-139 is comparative, like CLAIM-115.

2026-10-19 — batch 035 (possibility theses, part 2)

• Read CLAIM-238’s own lists before calling an objected-to claim a possibility thesis. It sorts its targets into three shapes: nulls with no margin, “can” theses, and impossibility conditions. The remedy differs by shape: a margin, a prevalence claim, or an argument. CLAIM-238’s version-2 history names which targets have been restated; a target at version 1 is still bound in full.
• Exclusive “only” again, in a CLAIM (CLAIM-143). Test a singular “after X is fixed” against quantifying over a class of X. Here a disagreement can survive every admissible correspondence, so the claim holds only with “a class of correspondences” in place of “the correspondence”.
• THEORY-628’s outcome matches triage-005’s route (source, retained). With it the control sample (THEORY-071, -178, -306, -485, -628) is fully reviewed; tally it against the route when integrating. The queue also marks THEORY-176 control: true, though triage-005’s control list does not name it; reconcile. Arguments filed as THEORYs go to wave 5.
• A record-built toy model does not meet a promote_when that asks for a worked case of a conceptual distinction (THEORY-628). It builds the distinction in by construction. ADR-047 v2 covers proofs, not constructed cases.
• A coded-corpus result is a measurement under batch 020’s rule, and ADR-047 v3 can apply (THEORY-167). Check that the coding data, not only its summary, is in the text read. Propp’s per-tale schemes were garbled in the OCR and unread, and they carry the order claim.
• THEORY-167’s Source line understates NOTE-615’s reading (it names four chapters; the NOTE read all nine). Fix it when the entry is next edited.

2026-10-19 — round 031-035 tally and triage-006

• Round yield: 4 promotions (THEORY-216, -324, -178, -306), 3 title corrections (CLAIM-033, -087, -105, in #222), 16 retained. The source route went 1 of 5 in each of batches 031 and 032, and the one in 032 (THEORY-324) was a prose deduction that the formal flag missed.
• Triage-005 control tally (THEORY-071, -178, -306, -485, -628): the route flag sent all five to source; two yielded (THEORY-178, -306, both measurements under ADR-047 v3). Batch 020’s sort by what the title asserts got all ten of 031 and 033 right. Keep sorting by the title; use the flag only to order the queue.
• THEORY-176 has no control flag in the queue; batch 035’s note about it was a misreading and needs no change.
• Triage-006 replaces the formal flag (signals: title, cites, relation, derived; minus a philosophical title; threshold 2). On the 30 labelled entries it gains recall (0.90 vs 0.76) at equal precision, but it was fitted to those 30. Every labelled false flag scores exactly 2, as do 53 of 82 flagged; treat the score-2 band as noisy. THEORY-290 is wrongly penalised by the philosophical-title signal (“argument” there means debate content); review it as formal.
• Batch 034’s decisions for CLAIM-033 and CLAIM-115 replace batch 028’s; CLAIM-077’s retain-proposed replaces batch 001’s needs-correction.

2026-10-19 — owner ruling: prevalence over possibility

• The owner accepted #222’s restatement of CLAIM-033, -087 and -105 as prevalence claims, as easier to corroborate and harder to reject. When a possibility thesis’s defeater tests a rate, restate the title as a prevalence claim (CLAIM-238’s remedy); do not weaken the defeater to “the cited instances fail on inspection”.

2026-10-19 — batch 040 (wave 5, requeue)

• CLAIM-265’s defeat condition was met by its own counterexample: a verdict reached over a whole class “without any correspondence being fixed”. When a defeater names a condition, check whether a case the record already holds meets it.
• An argument whose title, cut down to what it establishes, becomes the premise or a conditional on it (THEORY-223, THEORY-593) is a conceptual proposal waiting on its experiment. Retain it. Do not keep narrowing.
• An argument THEORY can have a real narrow-and-promote path in wave 5 (THEORY-073: restrict to the formulation and applications the critiques examined). ADR-047 has no route for arguments, so it goes to the owner like an interpretive CLAIM promotion.
• A correction that gives a CLAIM a proof (CLAIM-265) moves it to wave 1. Requeue it there as demonstrated once #227 merges.
• Requeue: CLAIM-265 to wave 1 after #227. CLAIM-143 after CLAIM-004 is decided and the owner rules on carrying the class wording into it. THEORY-223 and THEORY-593 wait on studies; no requeue.

2026-10-19 — batch 041 (wave 5, requeue)

• Judging an argument on its own terms in wave 5 did not change any of the three wave-3 verdicts. Each promote_when asks for something an argument cannot supply (a mechanical procedure for THEORY-071, a formally stated worked case for THEORY-628), or for readings the record lacks (THEORY-485). Requeuing an argument from wave 3 to wave 5 is only worth a review if its promote_when, its premises or its rival changed in between. Check git log on the entry, its NOTE and its rival before spending a review.
• The one route by which these arguments could move is a title that reports the argument (“X argue that …”). That is batch 032’s open owner question on THEORY-351. THEORY-071 and THEORY-628 both wait on it. Ask it once for the class.
• THEORY-485 is the cheapest of the three: its promote_when can be met by readings (Ney 2016 and 2021, Chalmers and McQueen 2022, Bikaraan-Behesht 2022) plus an argument for P7. Queue a reading task.
• A numerical model result read in full counts as one study under v3 (THEORY-178, now THEORY-290). Narrow to the model’s sizes and strategy space, and keep the equilibrium-selection caveat in the body when a result depends on which equilibrium is chosen.
• Triage-006’s philosophical-title penalty is a word match. THEORY-290’s “argument” was debate content. Before applying the penalty, check whether the word is the entry’s subject or its method.

2026-10-19 — batch 036 (wave 3, triage-006 formal 1-5)

• Triage-006’s top five went 4 of 5 by re-derivation; the fifth (THEORY-080) derives too but is the record’s own synthesis. Before re- deriving a formal_route entry, read its summary for “the filer’s synthesis” or “stated by no source”: those go to a CLAIM under ADR-032, not to promotion.
• A source theorem about a sequence of networks can hold as stated only under strong connectivity. When a title says “the consensus” or “the limit belief”, check whether the source’s statement allows several closed groups, and narrow to one if the derivation needs it (THEORY-321).
• “Learned not at all below X and in full above it” (THEORY-293) is a recurring overstatement of threshold theorems whose upper bound carries a margin. Check both ends of a threshold.
• For a “learns modes in order of eigenvalue” title, check that the order is stated for a common initial scale; the step time depends on the initial scale too (THEORY-229).
• The disintegrated change of measure plus the subgaussian square-moment bound reproduces Sefidgaran & Zaidi’s Theorem 1 constants exactly; a reusable check for PAC-Bayes-shaped titles.

2026-10-19 — batch 037 (wave 3, triage-006 formal 6-10)

• Triage-006’s formal route went 4 of 5 on its entries 6–10. The miss, THEORY-110, is a false flag of a nameable kind: an empirical conjecture whose body cites theorems proved for toy models. The “cites” signal fires on those. When the cited theorems are about a different model from the title’s, the entry is not formal.
• Check every “only when A and B” in a title against the exact condition. THEORY-185’s “weak and alike” turned out to be “alike relative to their size” (min s > s_(1) s_(2)). Equal strengths keep the analogies at any strength, so “weak” was a first-order assumption, not a necessity.
• A Kronecker or tensor-product eigenstructure gives clean attribute modes only when each factor’s top eigenvector is constant (balanced attributes). Without that, even the top mode is a product. Check balance before reading “the eigenvectors are the attributes”.
• A bound stated as “Bayes risk at most X above” can be X in expectation and 2X pointwise (THEORY-161’s TV bound). The tower rule gives the expectation form. Check which one a title means.
• A conditional title (“When P, then Q”) whose promote_when asks for P to be measured: under ADR-047 the measurement sits outside the title and moves to a QUESTION (THEORY-185). This is the same move as THEORY-324’s wider form. It is flagged in the decision in case the owner reads such a promote_when as a hold.
• A reading claim about a field (“the frameworks of 2021–2025”) can be held to the papers read in full, by analogy with ADR-047 v3. The field- wide form moves to a QUESTION (THEORY-161).
• THEORY-110 could be split in the manner of THEORY-109 once the record has checked Lin et al.’s Theorems 10, 11 and 18 (NOTE-535 read them for statement only).

2026-10-19 — batch 038 (wave 3, triage-006 formal 11-15)

• Triage-006’s formal route went 4 of 5 on its entries 11–15, all at score 3. The miss (THEORY-270) was a theorem whose long proofs the NOTE read only in outline. Read a NOTE’s history line for “overviews” or “not checked line by line” before routing a theorem to a one-batch re- derivation.
• A flagged constant inconsistency can be a rescaling. THEORY-238’s proof applies Lemma 3 at λ/2, so “λ/n < 1.36” in the proof and “λ/n < 0.68” in the lemma agree. Check the argument the lemma is applied at before calling the constants wrong.
• Check whether a promote_when was met by sources added later. THEORY-197’s promote_when asked for a second derivation, and versions 2 and 3 had already added two sources (LIT-887, LIT-891) that give it. Neither version note said the condition was met.
• THEORY-266’s promotion rests on Theorem 4 of LIT-233, which is THEORY-263’s framework and was reviewed elsewhere in this round. If that review finds a defect in Theorem 4’s proof, revisit THEORY-266. #230 also trims THEORY-266’s and THEORY-238’s names from THEORY-263’s inactive-ok directives, so expect a directive-line conflict if a THEORY-263 PR touches the same lines.
• Numerical side conditions are worth one script. Lemma 3’s second inequality is stated by the paper for x < 1.3, but the lemma needs it to 1.36. It holds there, by 0.0025.
• THEORY-270 should be requeued for a dedicated re-derivation of LIT-941 App. C, the common-nonsense upper and lower bounds. Once that is done, narrow the title to the stylised models and promote, moving the promote_when to a QUESTION.

2026-10-19 — lint trap: codes in titles

• Never put a bare code in a title:. luria repair links it in the body heading but not in the front matter, so title: and the heading disagree and lint fails (#230: THEORY-278, two new QUESTIONs). Name the source by its authors instead.

2026-10-19 — batch 039 (wave 3, triage-006 control)

• Triage-006 control: the formal flag went 1 of 5. Only the score-3 entry (THEORY-205) is formal. All four score-2 entries are measurements or arguments, and so were the four labelled false flags. The ‘relation’ signal fired on definitions and estimator formulas. Raise the threshold to 3, or require ‘relation’ together with ‘derived’ or ‘cites’.
• Batch 020’s title sort got 5 of 5 again. Measurements flagged formal still promote under v3 (THEORY-275, -365), so the score-2 band is worth reviewing as v3 work, not as re-derivation work.
• IB titles saying “first transition”: the linear-stability point is β = 1/ρ_m², the squared top singular value of the normalised channel. The point where the trivial encoder stops being optimal is 1/s*, the KL contraction coefficient, and s* ≥ ρ_m², so that point can come much earlier. They coincide in the symmetric Ising check (s*/ρ² = 0.999999) and differ 27-fold on a skewed 3-state chain. Name the bifurcation.
• The philosophical-title penalty’s word list missed THEORY-162 (“is part of its meaning”). Thesis-shaped titles about what meaning, content or understanding “is” should count.
• Checking a NOTE’s quotations against the source’s extracted text (THEORY-365) is quick when the arXiv PDF is reachable, and it settles batch 018’s ‘title rests only on results the text reports’ directly.

2026-10-19 — round 036-041 tally

• Triage-006’s formal route went 12 of 15 on entries 1-15 (batches 036-038), against 1 of 5 for its control (batch 039), where the one hit was the only score-3 entry. Two control misses still promoted as measurements under ADR-047 v3. Raise the formal threshold to 3, or require relation with derived or cites; send score-2 entries to a v3 review.
• Batch 020’s sort by what the title asserts got all five control entries right again.
• Wave 5 (batches 040-041): the six requeued arguments moved none of their wave-3 verdicts except THEORY-073, which has a narrowing path that is the owner’s call. Before requeuing an argument, check git log for a change to its promote_when, premises or rival.
• THEORY-080: the exact half is the record’s own synthesis; file it as a demonstrated CLAIM (ADR-032) rather than promoting the THEORY.

2026-10-19 — batch 046 (follow-ups)

• A proof written the day before can contradict its own definition. CLAIM-265’s steps 2 and 4 applied maps to both descriptions while the definition maps one onto the other. Recheck a fresh proof against its stated definitions, not only its conclusion.
• Stochastic maps are the class that makes “every pair is compatible” true on a single context (the constant kernel); deterministic maps do not. Use them for a vacuous-class example.
• A record-own result split from a THEORY can be filed Active at once when its proof is self-contained (CLAIM-370, CLAIM-384); its decision file is its review.
• THEORY-162 joins THEORY-485 and THEORY-167 as entries whose next step is a reading task, not a review. Track reading tasks separately from the requeue list.

2026-10-19 — batch 042 (wave 3, triage-006 formal 16-20)

• Triage-006’s formal_route 16–20 went 3 of 5 to promotion, with every title in the band re-deriving; the two misses are a report title (THEORY-344, the owner’s open question) and the record’s own synthesis (THEORY-019, the THEORY-080 pattern). Entries 1–20 now stand at 15 of 20.
• A title that reports “is stated” with the proof elsewhere (THEORY-387) is a theorem in waiting: if the record can re-derive the stated result, narrow the title to assert it for the case re-derived and move the rest to a QUESTION, rather than holding it as a report.
• NOTE-763 misreads the dual radii of Calo et al. as (1−γ)/6 and (1−γ)/2; they are 6/(1−γ) and 2/(1−γ). Fractions extracted by pdftotext -layout lose their bar; check a radius or step size against the lemma that uses it before transcribing it.
• THEORY-019 joins THEORY-080: two derived results filed as THEORYs that the record assembled itself. Both corrections can go in one PR filing two demonstrated CLAIMs.

2026-10-19 — batch 045 (wave 3, triage-006 score-2 band)

• The score-2 band went 5 of 5 under the batch 020 sort (8 of 10 with batch 039). Queue the remaining 44 score-2 entries for wave 3 in batches of five; do not re-derive by default, sort by title.
• “Least total X over models” versus “sum of least X per part”: the two agree only when one model reaches every part’s minimum at once. In Contextuality-by-Default that is noncontextuality (THEORY-177). Check any title that sums per-part minima.
• When a title says a quantity “equals” another over a protocol, check whether the source’s identity is per step and the protocol result a bound (THEORY-026).
• When a kernel result holds because two kernels are proportional, the proportionality is the theorem; state its setting (THEORY-294: one- dimensional, zero bias, classes split by the origin).
• A cheap simulation can stand beside an ADR-047 v3 promotion: THEORY-253’s two-task and rotated-quadratic results reproduced in seconds.

2026-10-19 — batch 044 (wave 3, score-3 and re-derivations)

• Triage-006’s score-3 band went 2 of 3 (318, 585), and the miss (083) is the false-flag kind batch 037 named: end-state theorems quoted in the body of an empirical claim. Its formal route is now exhausted; the score-2 band should go to v3 review, as batch 039 advised.
• A re-derivation can find the title wrong as well as unproved. In THEORY-585 the source asserts the failure and the record had hedged it; proving it showed the failure set is decided by a context’s projections (support on the first factor), not by whether the context factorises, so the “only on V₁⊗ℂ1” clause was false (Bell-basis contexts agree). Correct the title to what was proved, then promote.
• Source proofs carry small errors worth recording in the decision: Lin et al.’s Theorem 6 allows α ≠ 0 where its proof gives α > 0, and their Theorem 10’s probability is not derived; Goldwasser et al.’s App. C.1 computes τ(A_i) as if A_i ⊆ T. None changed a bound the record keeps, but each changes what may be quoted.
• The split pattern (ADR-040, THEORY-109) recurs for empirical conjectures that cite toy-model theorems: THEORY-110 now, and the false flags batch 037 named. Check the cited theorems’ proofs once, then split, rather than re-reviewing the conjecture.
• Queue: THEORY-083 requeued after a reading task (Figures 2–7 of arXiv:2008.08186 into NOTE-476). THEORY-270 and THEORY-110 leave the requeue list.

2026-10-19 — batch 043 (wave 3, triage-006 score-3)

• Triage-006’s formal flag went 5 of 5 on the score-3 entries after the top 20. Together with batches 036–038 that is 17 of 20 at score 3 or more, against 1 of 5 for the score-2 control (batch 039).
• A source that asserts a sufficiency theorem “by a similar proof” can still be settled by re-derivation in one batch when the necessity proof already computes the boundary condition: at equality the condition becomes tangency (THEORY-189).
• “Tight only for X” titles: test the claim at p = q and at other trivial pairs; the intended reading is usually “for every pair” (THEORY-313).
• “Tight” applied to a bound family that includes InfoNCE: InfoNCE’s supremum is at most log K and generally below I, so say “attains its supremum” and name the members that then equal I (THEORY-237).
• When a negative clause (“does not make the minimizer X”) has a promote_when that accepts a finite counterexample, a numerical counterexample with many restarts settles it at medium confidence (THEORY-305). Record the generating recipe in the decision file, since scratch scripts do not survive the session.
• Check every constant a body quotes against the source text: THEORY-189’s K ≤ 2ε/(1 − ε²) misread Eq. 25, which gives K ≤ ε/(1 − ε²).

2026-10-19 — round 042-046 tally

• Triage-006’s formal route is exhausted: 17 of 20 at score 3 or more promoted (batches 036-038, 042-044), against 1 of 5 for the score-2 control. But the score-2 band yields under ADR-047 v3 review: batch 045 went 5 of 5, so 8 of 10 across both samples. Only 3 of the 9 checked score-2 entries were formal. Work through the remaining 44 by title.
• Misses on the formal route fall into four recognisable kinds: a report title (THEORY-344), the record’s own synthesis (THEORY-019, -080, which go to a demonstrated CLAIM under ADR-032), a theorem read only in outline (THEORY-270, now re-derived), and an empirical conjecture citing toy-model theorems (THEORY-083, -110; split the theorems out once checked).
• A re-derivation can show a title false, not just unproved (THEORY-585, -313, -237). Correct the title to what was proved, then promote.
• Recheck a fresh proof against its own definitions: CLAIM-265’s day-old proof contradicted its definition and was repaired in #237.
• Put the recipe for a numerical counterexample in the decision file; scratch scripts do not survive the session.

2026-10-19 — batch 051 (wave 3, score-2 band)

• The score-2 band went 4 of 5 here. With batches 039 and 045 that is 12 of 15. Three of the five are formal in fact, so 6 of 14 checked score-2 entries are formal. Sorting by title kept predicting every outcome.
• A possibility title (“an intervention can reproduce X without Y”) is settled by an existence construction. The promote_when’s request to show the measured cases are instances is wider than the title and moves to a QUESTION (THEORY-282).
• A promote_when asking for “a second reader’s line-by-line check” is met by an ADR-047 v2 re-derivation in the decision file. When the re- derivation also shows the named counterexample cannot exist, no QUESTION is needed (THEORY-284, -281).
• For an “only when” over deterministic maps, check which hypotheses the necessity direction actually uses. THEORY-281’s needs neither disjoint groups nor strict concavity; only sufficiency needs disjointness.
• In a single-study measurement, drop any clause the source states in words with no number (THEORY-272’s “unless old ones are forgotten”) and keep the measured one (no stored pattern stable at N = 100, n = 500).
• THEORY-302: requeue to wave 5.

2026-10-19 — batch 052 (follow-ups)

• A reading task can be done inside a batch when the source is on arXiv: render the figure pages (pdftoppm) and read the plots, not the extracted text, which scrambles multi-panel figures.
• Reading the plots changed the verdict’s shape, not just its confidence: the captions said every measure “decreases” and the NOTE said “monotone across all 21 cells”; the plots show non-monotone cells and a partial collapse on ImageNet. A captions-only reading of a grid of curves tends to overstate both completeness and monotonicity.
• Check whether large drops in a training curve coincide with learning- rate cuts before crediting them to the phase the paper names.
• The THEORY-080 pattern applied cleanly a second time (THEORY-019): the split CLAIM carries the proof and the THEORY keeps the reading and its promote_when. Neither THEORY promotes on the split, since what remains asks for a held source.
• THEORY-019 supports CLAIM-193 and CLAIM-269. Whether they should rest on the new claim instead is for their next review.

2026-10-20 — tmp codes in the ledger

• A decision file for a CLAIM filed in a batch is named by its tmp code (CLAIM-384). docs-generate concretizes the code in record/ but not in record/promotion-audit, so after the filing PR merges, rename the decision file and every ledger reference to the permanent code (CLAIM-384 became CLAIM-384).

2026-10-19 — batch 048 (wave 3, score-2 band)

• The score-2 band went 4 of 5 (12 of 15 with batches 039 and 045). Keep working through the remaining 34 by title.
• “Approximately closed” or “to first order” in a theorem title: bound the error, then check whether it vanishes uniformly or only in one regime, and narrow the title to the regime (THEORY-190).
• A promote_when that asks for a limit proof of a model’s signatures, where the record’s own derivation is the approximation being set aside, is a reading task, not a re-derivation (THEORY-188: the rigorous random- hyperbolic-graph literature).
• “Does not by itself favour X over Y” is settled by constructing a Y model that meets the regularity (THEORY-166’s classical model with order events).
• A mechanism verb in a measurement title (“interpret by inverting”) narrows to “matched” when the evidence is a fit of condition means with no rival fitted (THEORY-172).

2026-10-19 — batch 049 (wave 3, score-2 band)

• The score-2 band went 4 of 5 here, 12 of 15 across batches 039, 045 and 049. Keep working through it by title.
• “A symmetry implies a mirror” depends on the group containing an involution with a nontrivial fixed subspace. Positive rescaling (ReLU’s λ > 0) contains none, so check the sign before carrying a sparsity corollary (THEORY-213).
• A weight-decay threshold for “every minimum on the symmetric subspace” needs only a lower bound on the loss’s Hessian, not smoothness: the even function’s gradient vanishes at zero, so f(z) − f(0) ≥ −(c/2)‖z‖² (THEORY-213).
• A “can” thesis about an information measure promotes by construction once the title names the measure and the intervention distribution; its interpretive reading moves to a QUESTION (THEORY-222).
• Recomputing a quoted exponent from the paper’s own equations is a cheap check that can close the theorem route (THEORY-203’s ν ≈ 1.89 against a quoted 1.78).
• Requeue THEORY-203 if THEORY-199 is promoted and the ν gap is explained or a second parameter set is measured.

2026-10-19 — batch 047 (wave 3, score-2 band)

• Score-2 band: 4 of 5 here, 12 of 15 across batches 039, 045 and 047. Two of the four (THEORY-030, -133) were formal in fact and both titles were false as written; the other two were measurements narrowed under ADR-047 v3. Keep sorting by title.
• Landauer-type titles: with reversible joint dynamics of memory and environment, the floor is on the logical entropy lost, H(X | f(X)), so many-to-one maps on known data have no floor either. “Learning has no minimum cost” is false in Goldt and Seifert’s own framework; what has no floor is the heat (THEORY-030).
• Counterfactual “exactly when” titles: check that the precondition (the higher-level property is a difference-maker) is in the title, and whether the condition is “would” (all nearest worlds) or “might” (some) (THEORY-133).
• “Suitably arranged” in a determination result usually means a connectivity condition; state it (THEORY-117: the seen-pair graph must connect every value). A span condition needed for a classifier is not needed for determining the embeddings.
• THEORY-085’s CFAC - KFAC identity is the record’s own one-line algebra (class-weighted); file it as a demonstrated CLAIM with THEORY-019 in the next follow-up batch.

2026-10-19 — batch 050 (wave 3, score-2 band)

• The score-2 band went 5 of 5 (13 of 15 across batches 039, 045 and 050). Keep working through it by title.
• When a title says two quantities agree “only to within” some term, check whether the source shows the term is reached or only bounds the gap. THEORY-244’s entropy gap is reached (a point mass); THEORY-259’s alpha is only an upper bound, and the conditional form of domination replaces it by the likelihood family’s complexity, constant over H.
• A promote_when that names a simulation as sufficient can often be met in the review itself (THEORY-261: Kanerva’s basic SDM at up to 200,000 locations, about 20 s per run). Record the design, seeds and results in the decision file; the scripts do not survive the session.
• For a theorem over “some labelling”, look for a convexity argument that covers every labelling at once before following a case analysis (THEORY-271: a nonnegative affine function vanishing on half the cube vanishes on a facet).
• Jaeckel’s SDM fidelity formula is slightly conservative at small p, because it treats the target’s own activation count as additive noise, and slightly optimistic at large p, because Hamming-ball overlaps vary more than Poisson. Its optimum (2MT)^(-1/3) still sits between the measured and predicted optima.

2026-10-20 — round 047-052 tally

• The score-2 band went 21 of 25 (batches 047-051), so 26 of 30 with batches 039 and 045. Sorting by what the title asserts predicted every outcome; the misses were an argument (THEORY-302), a reading task (THEORY-188), a scaling argument (THEORY-203) and the record’s own algebra (THEORY-085, which goes to a demonstrated CLAIM).
• Several titles were false as written and were corrected before promotion: THEORY-030 (only the heat has no floor), THEORY-133 (missing difference-making condition, “would” vs “might”), THEORY-190 (closure only at large separation), THEORY-259 (“only to within” was a bound).
• A reading task can be done inside a batch when the source is on arXiv (THEORY-083): render the figure pages and read the plots.
• A promote_when that names a simulation as sufficient can be met inside the review; record the design, seeds and results (THEORY-261).
• With ten-odd PRs touching shared inactive-ok lines, conflicts recur after nearly every merge. resolve_directives handles most; a hunk where one side dropped a whole directive line needs a hand merge (keep the intersection).

2026-10-19 — batch 056 (wave 3, score-2 band tail)

• The score-2 band’s tail went 1 of 4. The three misses are kinds the band has shown before: an unsupported principle, an argument with a derivation attached, and a measurement reported only in plots.
• Re-derive a source’s own proof of an axiom, not just its conclusion. Ince’s subset-equality proof is right about the co-information and silent about the sign rule, which is what fails. Pointwise measures with a sign condition need the condition checked in every property proof.
• THEORY-673’s first clause is the Price equation, which re-derives in four lines. When the wave-5 review takes up the argument, decide whether to split the identity off as its own Active THEORY (ADR-040).
• THEORY-675 carries the anthology-candidate flag and is a loss-landscape measurement; it is a candidate for the anthology rather than further review here.
• What remains unqueued in wave 3 (triage_wave3.py scoring against the queue at 6759e51b, excluding this batch): 331 eligible entries. 15 are still flagged formal at score 2 or more (THEORY-332, -336, -362, -367, -368, -377, -380, -419, -459, -486, -527, -535, -548, -580, -660), which are the remaining score-2 band. After them, 316 remain: 84 at formal score 1, 224 at 0 and 8 at -1. Ten carry the philosophical-title penalty, which misfired once (THEORY-290). The formal flag has nothing left to rank, so a further pass is worth it only if it sorts by what the title asserts. A triage-007 that tags each remaining title as theorem, measurement or argument (with a fixed control sample), and queues the measurement and theorem titles for ADR-047 v3 and v2 review, is the next useful step; the score-2 band went 26 of 30 on exactly that sort.

2026-10-19 — batch 054 (wave 3, score-2 band)

• The score-2 band went 3 of 5 here, so 32 of 40 across batches 039, 045, 047–051 and 054 (039 and 045 went 8 of 10, 047–051 went 21 of 25). Sorting by title predicted every outcome. The two misses are an argument whose proved core is trivial (THEORY-459) and a theorem outside the reach of a review (THEORY-486).
• A theorem a paper states without proof can itself be false as stated. GKT’s Theorem 6.5 puts the unit-rule cycle through t, but a cycle on any tuple in one of t’s derivation trees makes a coefficient infinite. The paper’s own algorithm propagates that correctly. Check a stated criterion against the paper’s algorithm, which often encodes the true condition.
• Check a quoted observational-equivalence example’s parameters, not only its conclusion. Pearl’s M and M′ agree with δ = γ only when x = u (σ₁ = 0). The general pair needs δ = γσ_u²/(σ_u² + σ₁²) (THEORY-419).
• When an argument’s formal step is a cardinality count, test it on an infinite structure first. The backward shift on ℓ² is a surjective, non- injective linear map of a space onto an isomorphic copy, which is why De Haro and Butterfield need a physical argument for infinite state spaces (THEORY-459).
• Before retaining a “needs a refereed version” entry, check the arXiv API for a later version or journal reference. THEORY-486’s source is still v1 with none, so requeue only when that changes.
• Requeue THEORY-459 to wave 5. THEORY-486 waits on a refereed or revised source.

2026-10-19 — batch 053 (wave 3, score-2 band)

• The score-2 band went 4 of 5 here, 30 of 35 with the earlier samples. Sorting by title again predicted every outcome.
• A threshold rule’s “competitive within c” can fail by its overshoot. When the trigger fires on the statement that crosses a threshold, a step of size f times the threshold adds about f on each side of a cycle. Online-SI’s 3 becomes 3 + 2f. Check lower-bound examples for an overshoot before trusting that the bound is tight (THEORY-332).
• “Stated without proof” titles on algebraic maintenance rules (counting, delta rules) can often be re-derived from a telescoping identity. A small brute-force check against full recomputation is cheap and catches stratification mistakes (THEORY-362).
• For provenance invariance, monotonicity alone makes minimal witnesses semantic: every witness contains a single-derivation image and every such image is a witness. Rewrite-system results whose rules are not given cannot be re-derived, so move them to a QUESTION (THEORY-367).
• In a multi-system measurement title, check which systems were measured at which settings. Indyk and Xu measured seven of their ten indexes only at L = 0.1n, and ran three from a survey’s reimplementations, not their authors’ code (THEORY-368).
• THEORY-336 waits on a reading task, a stage-comparison primary study such as Geisler (1989) or Banks, Geisler and Bennett (1987), and on the report-title ruling.

2026-10-19 — batch 055 (wave 3, score-2 band)

• The score-2 band went 3 of 5 here; with batches 039, 045 and 047-051, 29 of 35. Four of the five are formal in fact. Sorting by title again predicted every outcome.
• A promote_when asking for a null comparison can be met inside the review, and the answer can be “it is the null” (THEORY-548: community variability falls with S exactly as summing independent oscillators does). State that in the narrowed title rather than dropping the clause.
• THEORY-548: τ_c is not set by the largest growth rate alone, as NOTE-061 inferred; from S = 40 up its correlation with π/(2 max r) is about 0.2, and τ_c falls below that bound. Worth a dated note on NOTE-061 when it is next touched.
• A short printed proof can rest on definitions in a companion paper (THEORY-580 on “The Topology of Causality”); fetch the companion and check the definitions, and note that its numbering differs by arXiv version.
• THEORY-527 is the third conditional theorem promoted with its promote_when moved to a QUESTION (after THEORY-185 and -133). Verdicts argued in prose from a definition are not consequences a review can check; drop them from the title.
• PhilSci-Archive fails TLS verification from this environment (THEORY-535); a reading task there needs another route to the text.

2026-10-19 — batch 057 (follow-ups)

• A reading task can find that the rigorous literature narrows a title rather than confirming it: Candellero and Fountoulakis show the global clustering of the hyperbolic random graph vanishes at every temperature when the degree exponent is at most 3, a transition the source’s simulations did not report.
• Name which clustering coefficient a title means. The average of local coefficients and the global (transitivity) coefficient behave differently under heavy tails.
• A paper-length rigorous proof cannot be re-derived in a batch. Whether a held, peer-reviewed proof that was skimmed meets “a proof” is an owner question (THEORY-188; compare THEORY-270 in batch 038).

2026-10-20 — round 053-057 tally; triage-006 finished

• The score-2 band finished at 40 of 54 promoted (batches 039, 045, 047-051, 053-056). The earlier “26 of 30” in these notes undercounted; the counts per batch are in the batch files. The tail (batch 056) went 1 of 4.
• Again several titles were false as written and were corrected before promotion: THEORY-332 (the 3x bound becomes 3 + 2f under overshoot), THEORY-380 (a theorem stated without proof was false; the cycle need not pass through t), THEORY-714 (subset equality fails; the proof never checks the sign rule). Re-derive a source’s unproved theorems and every axiom proof, not just conclusions.
• A reading task can narrow rather than confirm (THEORY-188: global clustering vanishes for degree exponent <= 3). Name which coefficient.
• Before retaining an entry waiting on a refereed version, check arXiv for a newer version (THEORY-486). PhilSci-Archive is TLS-blocked here.
• What remains of wave 3: 316 unqueued entries at formal score <= 1. The formal flag has nothing left to rank; a triage-007 that sorts by what the title asserts, with a control sample, is the next step.

2026-10-19 — batch 058 (wave 5, arguments)

• Batch 041’s lesson held again: none of the four had changed since its wave-3 review, and wave 5 moved none of the three pure arguments. Check git log before requeuing an argument; requeue only when a promote_when, premise or rival has changed, or when the entry carries a separable formal core.
• A separable formal core is the one thing wave 5 reliably yields (THEORY-673, like THEORY-109 and THEORY-110). Whether it becomes a THEORY or a demonstrated CLAIM turns on whether a source states it (THEORY) or the record assembled it (CLAIM, ADR-032).
• THEORY-302 and THEORY-660 now wait on reading or data tasks, not reviews; THEORY-459 waits on a structural argument from its source or a rival. Track them off the review queue.
• Main still fails lint on NOTE-981, THEORY-753, THEORY-757 and THEORY-202; #272’s CI will stay red until #271 merges.

2026-10-19 — batch 059 (reading tasks)

• What could be fetched, and from where. arXiv: Chalmers and McQueen (2105.02314), full text. gumer.info: Propp’s Morphology, Russian HTML, apparently the 1969 edition, encoded cp1251 (decode it explicitly). It carries the appendices but not the fold-out schemes.
• What was closed. Unpaywall, OpenAlex and Semantic Scholar list no open- access copy of: Geisler 1989; Banks, Geisler and Bennett 1987; Geisler 1984; Banks, Sekuler and Anderson 1991; Mandelkern, Bounded Meaning (2024) and ‘Bounded Modality’ (2019); Ney 2016 and 2021; Bikaraan- Behesht 2021/22. No copy was found for Heim 1990 or Elbourne 2005 either. The 1968 English Propp at archive.org is lending-restricted.
• Sources that were blocked. philpapers.org returns 403 and PhilSci- Archive fails TLS through the proxy, so philosophy preprints are mostly unreachable from a session. cps.utexas.edu redirects to a page with no PDFs, and jov.arvojournals.org returns 403. Philosophy and vision- science reading tasks for closed works need owner-supplied copies; requeueing them unchanged will not help.
• Reading the source can narrow a promotable scope even when the clean copy is missing. Propp’s appendix states that five tabulated tales were dropped, partly because their order was not strictly canonical. So an order claim checked on the remaining 45 is partly a selection effect, and a v3 title must say so.
• A source cited as a rival can turn out to argue against the citing THEORY’s conclusion while still being represented accurately. Chalmers and McQueen are LIT-164’s “explicit, testable rival”, and they also argue that collapse is a causal-closure gap. Record both points.

2026-10-20 — triage-007, batches 058-059, and a broken main

• Triage-007 sorts the 327 unqueued wave-3 entries by what the title asserts: 76 theorem, 115 measurement, 130 argument, 6 report. On the 97 labelled entries (fitted to them) theorem precision 0.82 / recall 0.95, measurement 0.89 / 0.67, argument weak (0.60 / 0.40): conditional-shaped arguments are called theorems. Sort each route entry by hand before re-deriving.
• Wave 5 (batch 058) moved none of the pure arguments again; its only yield is a separable formal core (THEORY-673’s Price equation, split as a THEORY because a source states it).
• Reading tasks for closed works (batch 059) need owner-supplied copies; philpapers 403s and PhilSci-Archive fails TLS from here.
• Main broke when another session’s PRs (#246, #247) merged half-written entries; the completed versions were on its branch. Porting that commit (#271) fixed main. When main fails lint on files no audit PR touches, look for the fix on the originating branch first.
• Decision files outlive the PRs they cite, so tmp codes in them go stale when docs-generate concretizes main. After each round, map every X-tmp… in decisions/ to the file whose formerly: lists it, and replace it. Batch files may keep them, though replacing them there too is clearer.

2026-10-20 — batch 060 (triage-007 theorem route 1-5)

• Hand vs classifier on triage-007’s top 5 theorem calls: 3 of 5 agree. Both disagreements are philosophy titles shaped like results. THEORY-453 is a colon-joined position statement and THEORY-476 an “If … then” conditional. Both carry the signals theorem_title, cites, shown_thm and derived. So even at margin 3 the route mixes in arguments. Sort by hand before assigning re-derivations.
• The route ranks by margin, then rubric score. The three real theorems were margins 5, 5 and 4, and both false positives were margin 3. A margin cut at 4 would have been exact on this sample. The next batches can check whether margin >= 4 stays clean.
• Two of the three real theorems promoted because the missing steps were textbook-sized: a Laplace-method kernel expansion, and superposition coding plus a KKT check. The third, Isett’s Onsager theorem, is paper- length and joins THEORY-188 and THEORY-270 on the skimmed-proof question.
• The Hamming degrading channel has a closed form worth reusing. a(x) = min(p(x), theta), T = (kappa1/kappa2) I + 1 (a1 - a2)^T / kappa2. It was derived here, not copied from the paper’s (49). Numerical recipe in the THEORY-751 decision file.

2026-10-20 — batch 061 (triage-007 theorem route 6-10)

• Triage-007’s theorem label went 0 of 5 on entries 6-10. Entries 6-10 sit at margin 3, below the margin-5 and margin-4 heads. Four are conditional or modal-shaped arguments (“Granted that … so”, “reaches only”, “only if”, “as long as … cannot”), the failure the labelled check already showed. The fifth is a numerical study, flagged from its supplement’s analytics. The theorem route below margin 4 should be treated as unsorted. Sort it by hand, and expect mostly arguments.
• A measurement whose figures a NOTE read only from captions can be promoted under v3 once the figures are rendered from the arXiv PDF and checked (THEORY-194, like THEORY-083). Record which figures, and what each shows, in the dated note.
• Interpretive glosses appended to a measurement title (“which variables a compression keeps is set by what it must stay informative about”) leave the title at promotion and go to a dated note. When other entries cite the account for the gloss (THEORY-202 here), the note should say that the gloss is the record’s reading.
• Check the source’s own “small” derivations even when they are not in the title. Supplement Eq. 32 of LIT-873 is not normalised as printed; the ratio that follows is still right. That correction is in the decision file only, since the THEORY’s body does not use Eq. 32.
• THEORY-034 joins the reading tasks for closed works: Ladyman 1998 and Every Thing Must Go are Unavailable. THEORY-153 waits on field data. Keep both off the review queue.

2026-10-20 — batch 062 (triage-007 theorem route 11-15)

• Hand against classifier: 2 of 5 agree. That is below triage-007’s fitted theorem precision of 0.82, as expected at margins 2-3. - The two arguments were called theorems on a conditional clause (“when the appraisals rest on”) and on a textbook-theorem clause (“cannot increase along a processing chain”). - The scaling law was called a theorem on “because … provided”. - Even so, the misclassified scaling law promoted by the measurement route. Route membership is a worse predictor of yield than the hand sort.
• Scaling-law and training-dynamics titles split two ways. - THEORY-204’s derivation rests on assumed scaling forms, so it is a measurement with a conditional derivation. ADR-047 v3 narrows it to the measured match and moves the causal “because” to a QUESTION. - THEORY-209’s model result is exact, and its subject is deep networks, so it is split rather than narrowed (ADR-040). Rule of thumb: when the THEORY’s subject is the real system, split off the model result; when the subject is the measured system, narrow.
• A paper’s “broken detailed balance” can mean only a nonzero current. Kunin et al. define it as J ≠ 0. Under their own Assumption 3 the linear-regression state is time-reversible once velocities are flipped: each Hessian mode is an equilibrium oscillator at its own temperature. Check the definition before carrying “nonequilibrium” into a title. A one-line numerical check suffices: C(t) = εC(t)ᵀε.
• Check a paper’s text claim against its own table. LIT-888’s text says the fits overlap “for up to 10 of the first 12 points”; its table shows all eight TinyStories windows overlapping. Its TinyStories intervals also drift downward with fit length, which neither the text nor the THEORY mentioned.
• Both arXiv PDFs (2107.09133v4, 2602.07488v3) fetched cleanly with pdftotext -layout. Equation fractions come out split across lines, so reconstruct constants by re-deriving them, not by reading the text.
• THEORY-529 and THEORY-126 wait on a measure or a study, not on a wave-5 review. Track them off the review queue.

2026-10-20 — batch 063 (triage-007 measurement route 1-5)

• Hand against classifier: 3 of 5 agree outright, 1 partly (THEORY-201, a derivation with measurements attached), 1 wrong (THEORY-039, a cross- work synthesis whose title lists measured-sounding outcomes). That is in line with the 0.89 precision triage-007 reported. Its top measurement entries are good v3 material, but expect ML titles (THEORY-246, -201, -039) to carry a general thesis or a synthesis alongside the measurement.
• The commonest v3 move in this batch was dropping a generalising or causal lead clause and keeping the measured relation (THEORY-246, -334). Check the lead clause first: in both it was the part the study did not measure.
• “Drop clauses the source states only in words” bit twice: THEORY-246’s relearned-head result (Figure 6) and THEORY-334’s “keeps only the property needed” went to the body. Numbers in the running text (Ballard’s 50 and 129 ms, Kornblith’s “block 13 of 16”) were enough to keep a clause.
• A promote_when asking for “a test by others than the authors” can be met by the review’s own re-run when the experiment is small enough to write (THEORY-385: HNSW in about 170 lines of Python, 30,000 points, roughly an hour of CPU in all, with the machine shared). Record the design and every number in the decision file.
• The re-run contradicted one of the paper’s subsidiary claims: candidate extension, which Malkov and Yashunin recommend for clustered data, capped recall at about 0.97 against 0.999 without it. Worth a dated note on NOTE-916 when it is next touched.
• A leading-order derivation in the source (THEORY-201, App. D–F) can be re-derived in a few steps and checked by exact computation on sampled instances. The check showed where “to leading order” bites: at v = 16, m = 4 the exact covariance at depth 4 is a quarter below the formula. Put that limit in the split THEORY’s “does not say”.
• THEORY-039 is a THEORY whose summary says it is “the record’s synthesis across the five, not a claim any of them makes”. Under ADR-032 that belongs in a CLAIM (interpretive warrant). This is another case for the open owner question on record-own results filed as THEORY or Active (CLAIM-384, -385, -391).
• The scripts (hnsw_rules.py, rhm_corr.py, rhm_noise.py) live in the session scratchpad and are not committed; the decision files hold the recipes.

2026-10-20 — batch 064 (triage-007 control)

• Control verdict: the classifier’s zero score was right on class for 5 of 5 (none is a theorem or measurement title) but missed what moves an entry. One carried a derivation in its body (THEORY-072, split); two carried an empirical premise or measurement clause inside an argument’s title (THEORY-410 “each intention matches one value of it”; THEORY-528 “memory research finds that external stores lack …”), which make them reading tasks, not wave-5 arguments. The missing signals are body-level: a section marked as the record’s own inference from a definition, and “finds”/“matches” clauses inside an argument. A title-only classifier cannot see the first.
• Definitions are a source of splittable formal cores, not only theorems: THEORY-072’s core is a short proof from Montévil and Mossio’s Definition 4, which the paper states as a property without proof. Look for “the record’s reading” next to a quoted definition.
• Reading tasks with named sources: THEORY-410, Sadtler et al. 2014 (Nature) and Golub et al. 2018 (Nature Neuroscience) on decoder re- mapping (bear on the refutation clause; may yield a correction, not a promotion), Gallego et al. 2020 (Nature Neuroscience) on stable latent dynamics; THEORY-528, experiments on external stores (cognitive- offloading literature), whose measurement clause is already worded wider than the source (internal side measured, external side argued) and wants a correction when the reading is done.
• THEORY-152’s promote_when is empirical and its title has six clauses; neither wave 5 nor one v3 study can move it. Leave unqueued unless the owner asks for an exit-voice-loyalty-neglect reading, and split the title first.
• THEORY-646 is the one wave-5 case: its promote_when (an exclusion defence of whole-person functional states) may ask more than its title (realism does not require RTM). THEORY-133, which supplies the difference-making criterion, is Active, but no entry applies it to a whole-person belief.
• Montévil and Mossio’s HAL deposit (hal-01192916, /document) fetches as a PDF from here; the record’s LIT-582 url points at it.
• Process note: a pkill -f "luria lint" in this batch matched any running lint, so a parallel fork’s local lint may have been killed around 14:20 UTC. Never pkill by command pattern while forks run.

2026-10-20 — round 060-064 tally (triage-007 routes)

• 11 of 25 have promotion paths: theorem route 6 of 15 (batches 060-062), measurement route 4 of 5 (063), control 1 of 5 (064).
• Hand sort against triage-007: theorem label right on 5 of 15. At margin >= 4 it was right 3 of 3 (THEORY-763, -744, -751); at margin 2-3 it was right 2 of 12, the rest conditional- or modal-shaped arguments. Measurement label right on 3 of 5, partly on 1. The control’s zero score was right on class for all five.
• The hand label predicts yield better than the route: two misrouted “theorems” (THEORY-194, -204) promoted as measurements.
• Measurement titles in ML often lead with a general thesis next to the measurement; in both batch-063 narrowings the lead clause was the part not measured.
• Definitions are a second source of splittable formal cores (THEORY-072). When a THEORY’s subject is the real system, split off the model result; when its subject is the measured system, narrow it (batch 062).
• Never pkill -f "luria lint": parallel forks share the process table, and batch 064’s kill took out batch 060’s lint.
• ADR-048 (2026-10-19): do not check the anthology before filing.

2026-10-20 — narrowing loss and triage-008 (ADR-047 v4)

• The owner noticed that narrowing had dropped the wider claims with nowhere to go: of 93 narrowed THEORYs, 2 had a Proposed carrier of the wide form, 38 only a QUESTION, 53 nothing outside their own body. ADR-047 v4 routes each dropped clause (Proposed THEORY, Proposed interpretive CLAIM, note, or QUESTION).
• Triage-008 sorted 120 entries (93 narrowed, 21 promoted as-is, 6 from this round): kind 1 54 clauses (2 already carried, by THEORY-618), kind 2 2, kind 3 39, kind 4 30. 52 codes need filing; 26 dropped nothing.
• Several pre-narrowing titles already named their own limits (“is stated”, “not shown”); those limits are kind 3.
• #279 merged on the owner’s “merge the PRs”: a review’s own small re-run counts as a test by others than the authors (THEORY-385).

2026-10-20 — narrowing sweep, batches 065-069

• 25 codes, 26 new Proposed entries: 25 THEORYs (THEORY-780..804) and one interpretive CLAIM (CLAIM-396, THEORY-275’s generalization, which was the record’s own). THEORY-190 split into two theses; THEORY-194’s gloss was already carried by the Active CLAIM-152.
• (batch 065) Each new Proposed THEORY answers the QUESTION the original narrowing created, so the question now has a proposed answer and the wide claim inherits the old promote_when unchanged. Use that link in later batches.
• (batch 065) Triage-008’s existing-carrier check looked only at extends, refines and complements on Proposed entries. A kind-2 gloss can already be held by an Active CLAIM that grounds on the narrowed entry (CLAIM-152 for THEORY-194). Later batches should grep claims.d for grounds naming the code before filing a kind-2 CLAIM.
• (batch 065) luria new —draft takes JSON, not markdown; flags or a JSON object with the frontmatter keys plus body.
• (batch 066) Each narrowing in this batch had already moved its promote_when to a QUESTION that asks exactly the dropped claim, so the new Proposed THEORY answers that QUESTION as well as extends the narrowed entry. Expect the same for most kind-1 entries whose QUESTION was created by the narrowing.
• (batch 066) THEORY-187’s new entry is titled narrower than triage-008 suggested only in one respect, the ε law from small random weights, which Saxe, McClelland and Ganguli argue is slow but do not derive. The inherited promote_when still asks for the ε law too, so it asks a little more than the title; left as inherited, per the sweep’s rule.
• (batch 066) THEORY-163’s authors also read the fit as languages being shaped by a pressure for efficient compression. The paper gives no process (NOTE-631), so that causal reading was not folded into the new entry and is named under “What this does not say”; it would be a separate kind-1 thesis with a promote_when of its own.
• (batch 066) textwrap’s default break_on_hyphens splits codes (“THEORY-” / “tmp…”) and hyphenated words across folded YAML lines, which a folded scalar then rejoins with a space. Wrap with break_on_hyphens=False.
• (batch 066) After the PR merges, replace the five tmp codes above with the codes their files carry under formerly:.
• (batch 067) Every narrowed entry in this batch already had a QUESTION holding its old promote_when; the new Proposed entry inherits that promote_when and names the QUESTION as the test. The QUESTION asks, the new entry asserts.
• (batch 067) Check whether the source asserts the general form before filing kind 1. The pre-narrowing title can be the record’s own generalization (THEORY-275), and then the carrier is an interpretive CLAIM, not a THEORY.
• (batch 067) textwrap breaks codes and hyphenated words at hyphens; use break_on_hyphens=False when wrapping YAML notes that hold codes.
• (batch 068) Each of the five narrowings had already moved its old promote_when to a QUESTION, and that promote_when tests exactly the dropped thesis. So the new Proposed THEORY inherits the promote_when verbatim and answers that QUESTION; QUESTION and THEORY now ask and propose the same thing.
• (batch 068) Triage-008’s suggested titles sometimes over-reach the source in the same way the pre-narrowing titles did (the optimizer as “the” cause; “arithmetic” for addition). Check each suggested title against the NOTE’s claims table, and against results in the same paper that cut against the generalization, before filing.
• (batch 068) A triage-008 clause can bundle two theses (THEORY-204’s horizon mechanism and the universality-class conjecture). Split them, and file only what the source asserts with some evidence.
• (batch 068) The narrowed Active entries get an inactive-ok-file directive naming the new Proposed entry, a version bump with a history note, and a dated body note; no status or title change.
• (batch 068) THEORY-253’s summary still names its QUESTION by a stale tmp code (“QUESTION- tmpihczu”, broken across a line, so luria repair never concretized it); it is QUESTION-193. Fixed in this batch, since THEORY-253 is touched anyway.
• (batch 069) Triage-008’s suggested titles end “as X proposes”; that is the report-title form still open with the owner (THEORY-344, -351, -071, -628, -336). The filed titles state the thesis as a finding and name the proposer in the summary and body instead.
• (batch 069) When the old promote_when has already moved to a QUESTION, the new Proposed THEORY carries it as its own promote_when and answers that QUESTION, so the question shows the proposal that would answer it.
• (batch 069) One triage-008 clause can hold two theses (THEORY-190’s model degree flow and its real-network inference); file one entry per thesis.
• (batch 069) triage-008’s who_asserts can name the wrong paper of a pair (THEORY-243); check against the NOTE.
• (batch 069) THEORY-800 keeps Hopfield’s 0.15N; the analytic value usually cited is about 0.138N, unread here. If the record reads Amit, Gutfreund and Sompolinsky, the figure may need correcting.

2026-10-20 — narrowing sweep, batches 070-072; sweep complete

• 15 codes, 13 new Proposed THEORYs (THEORY-805..817). THEORY-362 and THEORY-377 were not narrowing losses. Across 065-072: 40 codes, 39 new Proposed entries. 13 codes wait on owner rulings.
• Recurring triage-008 errors, now checked by every batch: suggested titles over-reach or drop the source’s hedge; who_asserts names the wrong authors; clauses listed that were never in a title.
• batch-071.yaml reached main unparseable (a note opening with a quoted phrase and continuing past the quote). check.py now parses every batch file; quote YAML notes that start with a quotation mark.
• (batch 070) All five narrowings had moved their old promote_when to a QUESTION asking the dropped thesis (QUESTION-157, -171, -156, -195, -176); each new entry answers it.
• (batch 070) THEORY-255’s old promote_when tested the record’s own generalization of Theorem 9 (now in the narrowed Active title), not the emergence thesis, so the new entry has its own promote_when; inheriting unchanged is not always right.
• (batch 070) Counter-results in the same paper trimmed two suggested titles’ scope (THEORY-256’s strong-noise persistence; THEORY-293’s PopQA extension left out).
• (batch 070) The narrowing batches named in triage-008 contexts are batch-019, -033, -018, -044 and -036 (from the decision files); triage-008 does not record them, so check the decision file.
• (batch 071) All five narrowed entries already had a QUESTION holding the old promote_when (QUESTION-192, -151, -168, -212, -198); each new THEORY inherits it verbatim and answers that QUESTION.
• (batch 071) {“Triage-008’s suggested titles over-reached in two ways here, as batches 068 and 069 found”: ‘one widened the class past what the source names (THEORY-391), one asserted a mechanism the source cannot identify (THEORY-267). One was weaker than the source (THEORY-295).’}
• (batch 071) No existing CLAIM carries any of the five; no kind-2 clause in this batch.
• (batch 072) Triage-008 also listed clauses from as-is promotions (THEORY-362) and clauses still in the Active title as hedges (THEORY-377). Neither is a narrowing loss; check the pre-narrowing and current titles before filing.
• (batch 072) Triage-008’s who_asserts can name the wrong authors outright (THEORY-026: ‘Garner and coauthors’ for Fiderer et al.); check against the LIT’s first_author and the NOTE.
• (batch 072) Triage-008’s suggested titles can drop the source’s own hedge (THEORY-365’s low-dimension qualification). The filed title keeps it, because the paper’s own high-dimensional measurement departs from the claim.
• (batch 072) A new Proposed entry for one source can sit beside an earlier sweep entry with the same mechanism from different sources (THEORY-816 beside THEORY-785); check for overlap before filing and say how they differ in the body.

2026-10-20 — batch 073 (triage-007 measurement route 6-10)

• Hand against classifier: 3 of 5 agree (THEORY-333, -376, -421, all measurement). The two misses, THEORY-680 and THEORY-684 at margin 3, are a psychology framework whose one empirical clause is a review’s report of other studies, and a philosophy argument on testimony. Both titles state findings in the indicative (“has different consequences”, “holds”), which reads as measurement to a title classifier.
• Measurement titles already written to the study (THEORY-333, -376) promote with little or no change; the yield came from earlier careful filing, not from this review.
• Figure inset statistics count as numbers in the text for ADR-047 v3 when the PDF text layer carries them: THEORY-421’s Figure 2 ROOT, MIN, MEAN and MAX values are in the extracted text of arXiv:2306.09118v1, and the paper’s prose repeats two of them.
• A formal remark inside a measurement title can be kept if it is re- derived and stated conditionally (THEORY-421’s “a loss built only on pairwise distances leaves the origin arbitrary”); here the condition matters, because HGCN’s node-classification loss is not pairwise-only.
• Reading task: THEORY-680’s reappraisal-suppression contrast could be split off as a measured THEORY once Gross (1998, JPSP 74:224-237) is read; the framework itself goes to wave 5.
• After the PR merges, three tmp codes in this file and in decisions/THEORY-421.yaml need their permanent codes.

2026-10-20 — batch 074 (triage-007 measurement route 11-15)

• Hand against classifier: 2 of 5 agree. All five sit at margin 3, and the three misses are arguments whose titles use empirical words (“empirical hypothesis”, “open empirical question”, “may exploit a cognitive process”). With batch 063 (3 of 5 at margins 3-4), the measurement label below margin 4 looks no better than a coin.
• ADR-047 v4 changes what a narrowing PR carries. In both narrowings here the old promote_when tested the dropped claim rather than the measurement, so it moved with that claim to the new Proposed carrier instead of to a QUESTION. Expect this whenever a title’s measured part is fully supported and its promote_when asks for something else.
• A figure with no printed numbers keeps a clause out of an Active title under v3, but under v4 it need not be lost: THEORY-360’s over-rotation result is carried, with the reading it supports, by the Proposed entry.
• The UCSD copy of Kirsh and Maglio (adrenaline.ucsd.edu) now redirects to HTTPS with a certificate that does not verify from here; NOTE-878’s reading, with its script-checked quotations, was relied on. arXiv PDFs fetch fine.

2026-10-20 — batch 075 (triage-007 measurement route 16-20)

• Hand against classifier at margin 2: 3 of 5 agree. The two misses (THEORY-495, -575) are theses whose titles list measured-sounding examples or behaviours; both were the ones the directive flagged. Over batches 063 and 075 the measurement label is right outright on 6 of 10 and partly on 1, better than the theorem label at margin 2-3.
• Under v4, when a narrowed entry’s old promote_when tests exactly the claim the narrowing sets aside, it goes with that claim to the new Proposed THEORY and no QUESTION is needed: the Proposed entry is the record’s test. All three promotions here took that route.
• THEORY-766 needed no clause removed: its title claimed no causal attribution, but “raises reports” invited one. Appending the expectation confound makes the v3 condition plain, and the causal reading the promote_when tests is a source’s thesis (Ehrsson et al.), so it gets a Proposed carrier rather than a QUESTION.
• THEORY-105’s edge clause failed v3’s figures rule (extrapolation curves read only from captions), so the measured part kept is narrower than the old title’s headline. Rendering App. F/G from the arXiv PDF could later support an edge clause on the narrowed entry.
• THEORY-495 is a reading task before it is a wave-5 argument: one primary study read in full (Hamann et al. 2011, Nature 476) would let the sharing contrast split off as a measured THEORY.

2026-10-20 — round 073-075 tally (triage-007 measurement route 6-20)

• 8 of 15 promoted, all under ADR-047 v4: every narrowing filed its dropped generalization as a Proposed THEORY in the same PR (THEORY-823 to THEORY-828), or noted it when it was a mis-statement (THEORY-376).
• Hand sort against triage-007’s measurement label: 8 of 15 agree. All 7 misses are arguments whose titles use empirical words or list measured- sounding examples (THEORY-547, -436, -627, -680, -684, -495, -575). Across batches 063 and 073-075 the measurement label is right on 11 of 20; below margin 4 the hand sort decides.
• When the narrowed entry’s old promote_when tested exactly the dropped claim, it moved to the new Proposed carrier and no QUESTION was needed.
• Owner, 2026-10-20: a pre-acceptance manuscript read in full meets “one study read in full” (THEORY-123, Barsade 2001).
• adrenaline.ucsd.edu fails TLS verification from here; arXiv fetches.

2026-10-20 — batch 076 (triage-007 theorem route 16-20)

• Hand against classifier: 1 of 5 agree at margin 2 (THEORY-374). Across batches 060-062 and 076 the theorem label is right on 6 of 20, and on 3 of 17 below margin 4. The misses here fired on “only if” (THEORY-141), “on that reading … can lack an answer when” (THEORY-563) and “is the form … and its content shapes” (THEORY-060). THEORY-102 is a measurement whose title states a dissociation.
• A survey-backed complexity result can be promoted on re-derivation when the primary papers are unread: the inverse-rules certain-answer theorem is a canonical-database plus homomorphism argument, and the closed-world hardness a short 3-colourability reduction, both a page. The survey’s “many cases” hedge gives way to the exact condition re-derived.
• When the old promote_when tests the wider claim and not the measurement (THEORY-102’s replication and percept-input dissociation), it moves to the v4 carrier unchanged; no QUESTION is needed.
• THEORY-060’s well-being half is a reading task: Adler (2012) read in full could let that half be split off as a measurement.

2026-10-20 — batch 077 (triage-007 theorem route 21-25)

• Hand against classifier: 1 of 5 agree, all at margin 2. Across batches 060-062 and 077 the theorem label is right on 3 of 3 at margin >= 4 and 3 of 17 below it. The misses here are an argument with a definitional opening (THEORY-429), an analogy (THEORY-173) and a report of a court holding (THEORY-480).
• A reported theorem can be narrowed to the half the review re-derives, with the other half filed as a Proposed THEORY under ADR-047 v4. For THEORY-379 the upper bound re-derives by a Chebyshev argument and a simulation; the lower bound (all correct algorithms) needs an adversary argument and was not attempted.
• Re-deriving a reported theorem can expose an under-stated condition. “Independent orderings” is not enough for Fagin’s bound: two deterministic lists in opposite orders are independent and defeat it. The narrowed title states the uniformity the derivation uses.
• Interpretive theses with an exact identity attached split cleanly (THEORY-412 like THEORY-673): the identity becomes Active, the thesis stays Proposed and extends it.
• The recipe for THEORY-379’s simulation is in its decision file; the script (fa_sim.py) is in the session scratchpad and not committed.

2026-10-20 — batch 078 (triage-007 theorem route 26-30)

• Hand against classifier: 1 of 5 agree at margins 1-2, and the one agreement is partial (a conditional theorem bundled with a thesis). Across batches 060-062 and 078 the theorem label is right on 3 of 3 at margin >= 4 and on 3 of 17 below it. Below margin 4 it is not a route.
• The re-derivation of the stratified conjunction principle uses Hrbacek’s Closure principle (r/n is V-standard when r and n are), not induction on n, because induction fails across levels. Recipe in the decision file.
• Reading task for THEORY-457: André and Baumard (2011), “The evolution of fairness in a biological market”, Evolution 65:1447-1456, doi:10.1111/j.1558-5646.2011.01232.x, the pairwise proportionality result the target article cites. If read and re-derived it could be split off.
• THEORY-525 needs a systematic study of the fan-celebrity identity asymmetry before a wave-5 review could move it.

2026-10-21 — round 076-078 tally (triage-007 theorem route 16-30); triage-007 exhausted

• 5 of 15 have promotion paths: THEORY-374 (view-containment complexity, re-derived), THEORY-102 (narrowed to one intracranial study) and THEORY-379 (Fagin’s threshold algorithm, narrowed to random orderings) promoted; THEORY-647 and THEORY-412 split off exact formal cores. New entries THEORY-829 to THEORY-832; the dropped wide claims of THEORY-102 and THEORY-379 went to Proposed THEORYs under ADR-047 v4.
• triage-007’s theorem label below margin 4 was right on 6 of 25 across batches 060-062 and 076-078; at margin >= 4, 3 of 3. Its routes are now exhausted.
• Re-deriving a reported theorem can expose an under-stated condition: independence of the sorted lists is not enough for the threshold algorithm’s bound (two lists in opposite orders need depth N/2).
• A survey’s hedge (“many cases”) can be replaced by the exact condition the re-derivation finds, and promoted.