DR-104: Ratified docs outrank the codebase as evidence of correctness

Context

The reference set, runbooks, platform-direction.md, the OKF doctrine and corpus-reframe-review.html were deliberately corrected (S504 rulings R1–R14) so the platform has a north star, precisely because large parts of the codebase are known-inaccurate: half-repointed work that later sessions are fixing. Initiative 12 exists to close that gap.

Nothing stated that precedence anywhere. In S515, four separate verdicts — three by sub-agents, all accepted by the Coordinator — were derived from the codebase and each was overturned by the owner. The four inferences all looked like diligence:

“record-run.ts has 10+ production importers, so it is live, not residue.” It has 18 callers, of which one passes the current cocoindex fields. The S507 audit had already diagnosed the module as “lib/pipeline/ cron-helper vintage, pre-cocoindex” — many callers is the symptom that phrase describes, read as its refutation.
“There is a row on staging, so the write path is correct.” Platform staging and prod are internal dev environments pre-launch. Population proves nothing, and emptiness may mean unexercised rather than dead — which also re-opened a retire verdict on pipeline_runs.items_updated/items_skipped.
“grep found nothing in scripts/census/, so the manifest coupling was never built.” It searched the column census, not the corpus fixture gate — which already carried a hardcoded DRIVER_MANIFEST_DEST_PATHS array.
“Every assertion binds a live table, so the spec matches the target model.” The specs bound live tables while framing themselves around retired concepts, and used binding-admission gate — a compound fusing the two gates DR-025 separates, defined in zero decision records and coined in one session working note, yet present at 10 code sites.

Decision

In rebase-class work, ratified documents are the authority on what is correct; code is evidence only of what exists. The precedence order is: corpus-reframe-review.html (R1–R8) and the Decision Register, then platform-direction.md and reference/entity-glossary.md and the OKF doctrine, then the task’s own spec chain — and only then the tree.

An existing reference in code is the symptom under investigation, not proof the thing referenced is right. Specifically, none of the following is evidence of correctness: caller count, row population on any Platform DB, absence from a grep, or assertions binding live tables.

This binds sub-agent dispatch too: a brief citing only task files, specs and code reproduces the codebase’s errors.

Alternatives considered

Leave it as reviewer judgement. Rejected — it was reviewer judgement in S515 and failed four times in one session, including cases where the audit of record already said the opposite.
Ban codebase evidence outright. Rejected — the code is the only source for what exists today, which is exactly what a census or impact pass needs. The rule is about precedence when doc and code disagree, not about ignoring the tree.
Encode it only in platform-context.md. Insufficient on its own: the anchor is read at session start and its pointers were treated as optional depth. It needs a citable DR-NNN that a verdict can be checked against, plus the skill change.

Consequences

A retire/keep/rename verdict must cite a ratified document, not a code observation. Where the doc is silent, that silence is a finding to raise, not licence to infer from the tree.
platform-context.md carries the operational form of this rule as ## Evidence precedence — docs outrank code, naming the four traps above; the start-session skill makes reading it binding and carries it into dispatch briefs.
A term absent from reference/entity-glossary.md is a drift risk (S504 R6). The source-lifecycle vocabulary — binding gate, knowledge admission, retention classes — is not yet in the glossary; id-408 owns closing that.
Costs a verification step on work that would previously have shipped on a plausible code read. That is the intended trade.