Topics — What a work is about. Nucleation holds whatever reading the owner files here, whether or not an Anthology of the SOTA topic could also hold it; instruction for ML practice belongs to the anthology (ADR-048). The words are seeded from the clusters the reading-time triage of 2026-09-25 found, and grow by decision.
Learning theory — why learning generalizes and how it proceeds — compression and description length, Kolmogorov complexity and sufficient statistics, rate–distortion and PAC-style bounds, model selection; and training dynamics: solvable models of learning trajectories, stages, plateaus and phase transitions in training, and the implicit bias of optimisation. A work whose subject is how to train a model better is ML practice and goes to the anthology (ADR-036).
7 of 191 CASE documents. Back to the full index.
| # | Title | Summary | Status |
|---|---|---|---|
| CASE-113 v2 | A deep network under conventional training: competing losses, an external optimizer and a scheduler, with three architectures that move governance inward and crossed interventions that locate it | The owner’s application case, U30–U31 of the exchange of 2026-10-17: a deep network during training, taken apart with the record’s individuation and volition frameworks. A network is trained on tasks with conflicting gradients by an optimizer with its own state and a scheduler. A31’s Systems A, B and C move the regulation of conflict from the optimizer into the learner by stages, and its crossed interventions say where a regulatory condition sits. Stipulated: the outcomes of the interventions are not known, except the optimizer arm in THEORY-253’s solvable model. | Active |
| CASE-117 v3 | Emergent specialization and recovery in a population of adaptive operators: identically parameterized operators whose contributions, influence and training opportunities co-evolve over a changing task distribution, compared across fixed and learned features, uniform, endorsement-based and contribution-based influence, coupled and decoupled training exposure, and imposed and endogenous diversity | The study proposed in the synthesis the owner relayed on 2026-10-18 for QUESTION-083 and the questions filed with it. Operators differ only by initialization and experience. It measures whether stable niches emerge, whether specialists recover after long irrelevance, and whether the organization adapts after losing a dominant specialist. The decisive comparison holds the final diversity of operator usage roughly fixed while changing the mechanism that produced it. Ordinary mixture-of-experts routing is the baseline. Even positive results would not show self-constitution (CLAIM-288). | Active |
| CASE-118 v2 | Two competing components with interaction matrix A = [[1, ρ], [ρ, 1]]: with equal growth rates their invasion growth is 1 − ρ, the smallest eigenvalue of the Gram matrix, and with growth rates 1 ± δ the same matrix supports coexistence only when |δ| < (1 − ρ)/(1 + ρ), so feature geometry alone does not fix coexistence | From the synthesis the owner relayed on 2026-10-18, checked. The two-component Lotka–Volterra system with symmetric overlap ρ: invasion growth 1 − ρ equals λ_min(A), and κ(A) = (1 + ρ)/(1 − ρ), so overlap weakens recovery and ill-conditions the features together. Holding A fixed and setting r = (1 + δ, 1 − δ) gives the coexistence equilibrium n* = (1/(1 + ρ) ± δ/(1 − ρ)), feasible only when |δ| < (1 − ρ)/(1 + ρ). The same representation geometry supports coexistence under one supply and exclusion under another. | Active |
| CASE-144 v2 | The evaluator that stops tracking: two populations of specialists with identical capabilities and task streams, one scored against independent outcomes and one by an evaluator that weights evidence agreeing with its previously favoured specialists, after a shift that makes a marginalized specialist uniquely competent | Proposed in the exchange of 2026-10-23, “Framework Overview Request”, at A5 (section 5) with A4’s “corrigibility trap”, after the owner asked which experiments could “kill multiple birds with one stone” (U4) and which questions simulation could answer (U5). One design bears on CASE-092, CASE-108 and CASE-117 at once, and separates four causes of a failure to adapt. A5’s central question: “Under what conditions does an organization become unable to recognize or act upon evidence that its own organization of interpretation and authority has rendered inconsequential?” Not run. | Active |
| CASE-155 v3 | Authority as a replicator population: for a distribution π of control over processes with mean qualities q, the expected Maynard-Cross update is π_a q_a / v, Bayes’ rule when q is each process’s likelihood; the Maynard dynamic is the Taylor dynamic slowed by 1/v, so the two share their orbits; mean quality climbs at the rate Var_π(q); and when a process’s score rises with its own share, ℓ′ = ℓ + h + c(2π − 1) has an unstable threshold π* = (1 − h/c)/2, above which a process the evidence disfavours captures the organization | Filed at the owner’s direction, 2026-10-24, after asking whether the tools of Soma et al. (THEORY-899) or related work could formalize the record’s abstract governance dynamics. Four results, each checked when filed, two by script. It reads authority as the population vector of a replicator dynamic. Evidence-tracking authority is then Bayesian, and authority-dependent scoring gives lock-in with an explicit correction threshold. | Active |
| CASE-157 v2 | Boosting the vote: in the owner’s 2014 model, voters of fixed accuracy drawn from Beta(2, 2) vote in a weight-averaged election, and every voter’s weight is multiplied by 1 + f when their own vote was right and 1 − f when not, with a fraction replaced each round at weight 1; with full-information feedback the weighted result rises from 0.50 toward 0.98 as f grows, and churn only costs in a world that never changes; the model escapes the hive mind’s absorbing states because every voter is scored, heeded or not, and it assumes away the capture its author predicted | Filed at the owner’s direction, 2026-10-24, after the owner suggested taking inspiration from the experiments in dmarx/BoostTheVote. A thought-experiment written in R in 2014, read at commit b5bca8f and re-run here through a Python port (scratch/formal/boost_the_vote.py). It is the full-information counterpart of the replicator construction. What it shows, what it assumes, and the three-axis extension it suggests are recorded, and the extension was run on 2026-10-25. | Active |
| CASE-158 v2 | Audited governance against a contextual router after a reversal of expertise: specialist denoisers with frozen, separately trained operators propose for one shared latent at each step, and spatial grants are revised from independent held-out probes of every proposal, available only from the next episode, against a fixed allocation, a conformity-scored allocation and a contextual mixture-of-experts router; the question is whether governance does anything a router does not, and under full-information audits the record predicts it cannot | Proposed at A46 and reported at A47 of the exchange of 2026-10-24, continued, after the owner’s “close the loop”. It is CASE-144’s design, run without citing it, with competence frozen so that only the allocation learns, and with a conventional router added as the rival that CLAIM-278 requires. A47 reported that the router won on every seed. Its results come from a sandbox the share does not carry, and the record has not run it. What would make it informative is one arm per failure of QUESTION-271. | Active |