decision

ADR 0010: Karpathy's "lint" *is* the dreaming pass - correcting ADR-0009's reading of S8

Context#

ADR-0009 added the dream stage and cited three sources. Two of them - S6 (OpenAI) and S7 (Anthropic) - had been ingested. The third had not:

A third source sharpened the framing. Karpathy's LLM Wiki gist describes this same pattern (raw sources / maintained wiki / schema) and names three operations: ingest, query, lint. The kit has all three. What it does not have is what S6 and S7 both argue is the load-bearing one - a maintenance pass that runs on its own clock rather than inside the ingest.

That gist has now been ingested as S8 (sources/260731_llm-wiki/), with its nodes gated before topics/memory.md was reopened, specifically so the prior conclusion would be re-derived rather than inherited. It does not survive.

The first two assertions hold - the three-layer architecture (S8 n4) and the three operations (S8 n6) are both accurate readings.

The third is wrong. §Lint reads in full:

"Periodically, ask the LLM to health-check the wiki. Look for: contradictions between pages, stale claims that newer sources have superseded, orphan pages with no inbound links, important concepts mentioned but lacking their own page, missing cross-references, data gaps that could be filled with a web search. The LLM is good at suggesting new questions to investigate and new sources to look for."

Against the eight classes the dream stage works:

"Periodically" and "ask the LLM to" place it outside the ingest, on its own trigger. It is the maintenance pass, described eight weeks before either vendor published.

How the error was made, which is the part worth keeping. The prior reading matched the word "lint" to validate.py, concluded the kit already had that operation, and stopped. The two share nothing but the name: validate.py checks form (ADR-0004), and every item on Karpathy's list is a judgement. The kit's own contract says exactly this, one sentence from where ADR-0009 reasons: "a green validator means the shape is right, not that the thinking is."

Decision#

Record that S8's §Lint is an independent third statement of the decoupled maintenance pass, and correct ADR-0009's characterisation of it. ADR-0009's decision stands unchanged; only its reading of one source changes.

Nothing about the dream stage is altered - not its trigger, its classes, its output, or its prohibition on running inside an ingest. What changes is what the brain believes about the evidence behind it.

What the third source is worth, stated precisely, because it is easy to overclaim:

The closest S8 comes to the rationale is by accident, and it is recorded as a divergence (S8 d1): §Why this works claims LLMs "don't forget to update a cross-reference" while §Lint instructs you to hunt for missing ones. Both cannot be true. Believe §Lint - it is the operational section, and the six-item list is an admission that integrate-on-ingest leaves defects behind. The usable form: LLM bookkeeping is cheap enough to be worth doing repeatedly, not so reliable that doing it once is enough. That is the observation underneath claim 59, not an argument for it.

Alternatives considered#

• Silently fix the sentence in ADR-0009. Rejected. An ADR is a dated record of what was decided and why; editing the reasoning after the fact destroys the thing it exists to preserve. The correction gets its own numbered entry and ADR-0009 gets a pointer.
• Reopen the decision itself. Rejected, and it was considered seriously: if one of three cited sources was misread, was the stage justified? Yes, and more strongly than before. The misreading understated the support - S8 argues for the pass, not for its absence. The stage would stand on S6 and S7 alone, and on the concrete defect that motivated it (the stale "no measurements" claim in memory.md).
• Treat S8 as raising confidence in claim 59. Rejected under the independence rule's sibling principle: S8 corroborates the practice and is silent on the reason. Recording it as support for the rationale would launder a coincidence into evidence. memory.md now splits the claim's confidence explicitly - practice: three sources, one non-vendor; rationale: still S7 alone.
• Rename the kit's stage to "lint" for consistency with the earliest source. Rejected. The kit already has a validate.py, and "lint" in every engineer's usage means a form checker - which is precisely the confusion that produced this error. dream is a worse metaphor (S6 and S7 borrow it from sleep without citing any of that literature) but an unambiguous name in this repo. Pick the name that cannot be mistaken for something you already have.

Consequences#

• memory.md gains a third source and an explicit split: the decoupled-curation practice now has three sources including one non-vendor and earliest; the rationale still has one.
• rag.md goes seed -> emerging on S8, its first source in the brain's history, and the topic's scope widens to cover maintained knowledge layers as the alternative to query-time retrieval - rather than spawning a knowledge-bases note that would put two halves of one argument in two places.
• The rag / memory boundary drawn in ADR-0007 needed a qualifier, and got one in both notes. "Who authored the corpus" does not separate them: an LLM Wiki is system-authored about someone else's documents. The refined line is what the knowledge is about - external sources (rag) versus the system's own experience (memory). Recorded here rather than as its own ADR: it is one qualifying sentence in two notes and creates no new structure. If a second source straddles the line, that earns an ADR.
• A named failure mode for this kit, and the real lesson. The defect was a claim made about a source that had not been ingested. It was wrong the day it was written, it lived in a file nothing re-reads, and a dream pass could not have caught it - reading all of brain/ would show a perfectly coherent story, because the contradicting evidence was outside the brain. The general rule: citing a source you have only skimmed creates a claim with no gate behind it. Prefer "reviewed, not yet ingested - claims about it are ungated" over a confident one-line summary, or ingest it first.
• Revisit when: a source straddles the rag/memory boundary a second time; or S8's n10 (index-file navigation replacing embedding retrieval at ~100 sources) is externally tested, which is the highest-value deep-research target either topic now carries.