2 entries. All books.
Contents
- 5 Sep 02:30 — Adding the source, and finding eight rules nobody had written down
- 5 Sep 04:15 — Writing the record that describes the thing writing it
Adding the source, and finding eight rules nobody had written down
2026-09-05 02:30:06 · record
The record described a document that was not in the repository. Every value and practice here is an assertion about a constitution, and until this session that constitution existed only in the assistant’s context — so the assertions had no referent and nothing could check them. Adding the source turned “these documents decompose a constitution” from a claim into something a reader can audit and a test can fail.
Mapping the source against the record found eight passages with nothing behind them. Not near-misses: whole sections of governing instruction that the record simply did not represent. Pronouns. What a denied tool call means. What to do with a request repeated after an objection. Correction restraint. Writing code in the surrounding idiom. The temptation was to stretch an existing document over each gap — VALUE-001 is about reporting honestly, and correction restraint is sort of about honesty — and each stretch would have made both documents vaguer. Nine new documents instead. A record that cannot grow is one that will be lied to.
The gaps clustered, which is the part worth keeping. Three of the five new practices turned out to share one unstated value: the reader’s attention is spent by whatever is written, so a survey where a recommendation was wanted, a correction that changes nothing, and a re-derivation of a settled fact are the same error in three costumes. That value (VALUE-006) had been re-derived at least three times in the source without ever being stated. Finding it required laying the source beside the record; neither on its own showed the pattern.
BOUNDARY-003 is the strongest argument
this record makes for its own shape. The pronoun rule is not merely an
addition — it contradicts two practices that were already here.
PRACTICE-001 and
PRACTICE-005 both say to make the routine
call rather than stalling, and both would license inferring pronouns from a
name: the inference is easy and usually right. In prose the conflict would have
been managed by emphasis — the pronoun paragraph shouting a little louder —
and nothing would have recorded that a conflict existed. Here it is an
overrides edge, which means it is visible on both documents’ pages, and it
survives someone editing either one.
Three sections are annotated as accounted for by nothing, and that block is load-bearing. How context summarisation works; that independent tool calls can be batched; which tool to prefer for reading a file. These are true, and they are facts about one runtime rather than claims about conduct. A record that absorbed them would be recording its own deployment, and would go stale without going wrong — the worst combination, because nothing would flag it. An explicit “nothing yet, and nothing should” is cheaper than a document that quietly rots.
BOUNDARY-002 needed revising rather than defending. As written it said restatement instead of reproduction, which, with the source now in the tree, read as a rule the record was breaking. The honest fix was to narrow it: reproducing a source is fine, and what the boundary actually forbids is a copy standing in for the analysis, or a reproduction quietly edited until it is no longer one. That second half decided how the references were added — annotation blocks after each section, never links threaded into the quoted prose. A citation inserted into a reproduced sentence makes it a sentence nobody wrote, and it does so while looking like diligence. The graph edges come out identical, so nothing was traded for it.
Writing the record that describes the thing writing it
2026-09-05 04:15:00 · record
This example was asked for as a tangent and turned out to exercise something the other examples do not: a record whose documents govern each other.
The first draft was a flat list and it was wrong. Every instruction became a document in one scheme, distinguished only by tags — which is how the source material already reads, and reproducing that shape would have demonstrated nothing. The decomposition only became interesting once the question changed from “what is this instruction about?” to “what happens when two of them disagree?”
That question has a structural answer, and it is the reason this configuration
exists. overrides is a plural reference from BOUNDARY to PRACTICE, so
precedence is an edge the lint follows rather than an adjective. The
alternative — the one every large instruction corpus drifts into — is
escalating emphasis: IMPORTANT, then CRITICAL, then capitals, each
inflating until none of them discriminates.
The three-way split is load-bearing, for the reason the anthology example gives about its two. A VALUE and a PRACTICE go stale at different rates: a claim about honesty survives a tool being replaced, and a rule about which tool to prefer does not. Collapse them and one status field has to mean both “we still believe this” and “this still describes the environment”, so neither can be said.
One document is retired on purpose. PRACTICE-004 is Superseded, and
BOUNDARY-002 overrides it. Deleting it instead would have left an override
pointing at nothing — and a record that deletes what it supersedes cannot say
what changed, only what is currently true.
What this cannot check, stated because the record’s own VALUE-004 requires it. The lint verifies that every practice names a value, that every override resolves, and that the generated views match their sources. It cannot verify that this decomposition is faithful to the constitution it describes. That gap is not closable by machinery, so it is written down instead — in BOUNDARY-002, where a reader meets it rather than having to infer it.