06-reference/concepts

citation version discipline

2026-10-18·reference
research-disciplineverificationdata-qualityvault-maintenance

Citation-version discipline (cite with a version/date stamp or don't cite it)

A single-week cluster of six independent 06-reference/research/ deep-research briefs (2026-10-12 through 2026-10-15) each closed the same way: a number the vault had been carrying as settled turned out to be stale, unversioned, or quietly re-derived from a different denominator than the one originally cited.

The common failure shape: a number gets cited once, correctly, against a primary source at a point in time — but the citation is stored as a bare fact ("8-12 months," "741%," "18-month durability") rather than as a fact-plus-provenance ("8-12 months, per USPTO FY2025 pendency report, as of [date]"). Once the provenance is dropped, nothing in the vault can tell the difference between a number that is still true and one that has quietly decayed, been revised upstream, or only ever held at a denominator nobody wrote down. The research brief that eventually re-pulls the primary source does the correction, but the correction is only possible because someone bothered to re-check — there is no structural signal that a given number is due for a recheck.

Why this is in the vault

Six deep-research briefs hitting the identical failure mode independently, in the same five-day window, across completely unrelated domains (trademark law, AI-coding productivity, semiconductor supply chains, metabolic-health RCTs) is a pattern, not a coincidence — it is what happens when a vault accumulates numeric claims faster than it re-verifies them. /compile-vault and /deep-research are exactly the mechanisms that catch this, but only when a brief happens to touch the same topic again; there is no proactive "this number is N months old, re-check it" signal. Naming the failure mode is the first step toward building one (a verified_as_of frontmatter field, or a standing /deep-research sweep over citations older than some threshold, are the natural next builds — not undertaken here, flagged for whoever picks this up).

Mapping against Ray Data Co