DR-106: Not every docs-site document is a north-star doc

DR-106: Not every docs-site document is a north-star doc

Context

DR-104 settled that ratified docs outrank the codebase as evidence of correctness in rebase-class work. It did not say which docs are ratified, and the omission cost two sessions.

S515 ruled RETIRE on pipeline_runs.items_updated from row counts. S516 re-opened it and a lane ruled WIRE, grounded on documents — apparently satisfying DR-104. The owner overturned both. The lane’s chain ran: phase-b-prerequisite-1-verification.md states “RESOLVED-RETAIN per feedback-findings-review §5.2.1”; that document actually reads DEFER-COCOINDEX | Determine after cocoindex op-ledger investigation. The cited ruling was not in the cited document. A supporting anchor, 09-diagrams.md, renders both columns in an ERD — but its own §1 boundary is “no new schema or flow content” and §8 names supabase/types/database.types.ts as its source for live-schema columns, making it code evidence one step removed.

The owner then ruled the surrounding families out of date, mid-session, which invalidated evidence already relied on.

Decision

Citing a stale document is the same error as citing the code. Three families are not ratified authority:

• initiatives/core-product/canonical-pipeline/intended-architecture/ — stale except 01-vision.md.
• canonical-pipeline/phase-0-investigation/ — historical record of how decisions were reached, not current authority.
• Any spec or invariant belonging to a task id below ~130, unless a current document re-affirms it.

07-collapse-list.md is the qualified case and is read asymmetrically: its [RATIFIED-RETIRE] rows are usually still sound and are a reasonable starting citation subject to checking for a later flip; its retain / NOT-COLLAPSING entries are the ones most likely to have flipped and carry no column-level detail; absence from it is evidence of nothing.

Two rules follow:

1. When a verdict rests on “doc A says X per doc B”, open doc B.
2. Retain-the-table is not retain-the-shape. A closing ruling can keep a table while enumerating a shape that excludes the columns in question — check what was enumerated, not the verdict word.

For pipeline and corpus work the current authority is reference/cocoindex-pipeline.md, the cocoindex skill (scripts/.claude/skills/cocoindex/), corpus-reframe-review.html, the OKF doctrine, the Decision Register, and reference/platform-context.md.

Alternatives Considered

• Leave DR-104 as-is and treat staleness case by case. Rejected: two sessions already failed the same way, and the second failure was by a lane explicitly briefed on DR-104. The rule was followed and still produced a wrong verdict, so the rule was incomplete rather than ignored.
• Archive or delete the stale families. Rejected: they hold the reasoning behind live decisions and are the only record of how several rulings were reached. The problem is citation as authority, not existence.
• Mark staleness per file with frontmatter. Not rejected on merit — it is a better long-term shape than a rule in a register, but it is a docs-wide mechanical change and nobody owns it yet. Recorded here as the standing rule until someone does.

Consequences

• A verdict grounded only in intended-architecture/ or phase-0-investigation/ is not grounded. Sub-agent briefs doing rebase-class work must carry this list.
• Some questions become unanswerable from the documents, which is a legitimate and useful outcome — it is what routed pipeline_runs’ counters to id-410 rather than to a third wrong verdict.
• The pre-130 cut-off is a heuristic, not a boundary in the ledger. It will produce false positives on specs that are still correct; the cost of checking is lower than the cost of the two verdicts it would have prevented.
• Operational detail and the read-never-written diagnostic that came out of the same session live in reference/platform-context.md under “Not every doc is a north-star doc”.