Decision: epistemic-status standard (claim-level maturity tags)
Current truth (the standard)
Maturity is a property of the claim, not the page — real pages mix maturities (the signal-set evidence table proved this). Every substantive claim in research/ and product/ carries one of five inline tags. The operating rules live in AGENTS.md (§ Epistemic status); this page holds the gates and the why.
BD/customer/people pages split in two: records of events (met X, said Y, stage Z — attribution is the claim) stay untagged; reads of the situation — interpretive claims someone would act on (who the champion is, budget/intent/timing, why a deal is stuck, whether a relayed signal is real, competitive reads) — get tagged like everything else. BD evidence that passes the gates: demonstrated behavior or an independent source linked to the raw/ transcript ([evidenced]); signed contracts, public record, or direct confirmation + a second source ([fact]). A third-hand signal (e.g. an investor rumor relayed on a call) is [idea] no matter how exciting.
The five states and their promotion gates
A tag is not an adjective — it is a gate that was passed. If the gate’s requirements are not met, the claim stays at the lower state. No exceptions, including for claims we’re excited about.
| Tag | Means | GATE — required to enter (all items) |
|---|---|---|
[idea] | Someone proposed it; untested. May not even be a coherent bet yet. | Attribution: who said it + when (date). That’s all — entry state. |
[hypothesis] | A deliberate, falsifiable bet we are (or plan to be) working. | 1) Stated so it can be false. 2) A written “what would settle it” — the observation that would refute it. If you can’t write the refuting observation, it stays [idea]. |
[evidenced] | We have real proof, but from one angle. | 1) At least ONE of: first-party measured evidence (run artifacts, experiment, corpus analysis) OR one independent external corroboration. 2) A link to the evidence (wikilink or URL — the page/dataset/repo where it lives). 3) A date in the tag: [evidenced 2026-07-04]. 4) Scope + known confounds stated in the claim (“Claude only, n=4.3k runs, confounded with content quality”). |
[fact] | Settled, within its stated scope. | Everything in [evidenced], PLUS one of: (a) reproducible on demand — a re-runnable probe/script exists and has been re-run on ≥2 separate dates with the same result, or (b) ≥2 independent evidence sources agreeing. AND: human sign-off (Vincent or Elliot) unless path (a) verified it mechanically. Platform-behavior facts MUST carry platform + as-of scope in the claim text — hosts drift; the tag date is the last verification. |
[refuted] | We tested it and it came back negative. | Same gate as [evidenced] (evidence link + date + scope). Negative knowledge is kept forever: flip the tag on the existing claim, never delete the row. |
Gate mechanics (who may move a claim)
- Promotion is evidence-gated; demotion is free. Any human or agent may demote a claim or flag it stale on sight, no justification beyond a sentence. Promotion requires the gate above, checked at write time.
- Promotions/demotions touching
[evidenced]/[fact]/[refuted]get a LOG line (promote |/demote |) naming the claim and the evidence link — the audit trail. - Mechanical lint:
tools/check-epistemic.shfails on[fact]/[evidenced]/[refuted]tags missing a date or an evidence link, and warns on[fact]tags older than 90 days (re-verify or demote). Run it before committing wiki changes. - Ingest never promotes. A transcript saying “X is true” produces
[idea](attributed) at most — “Theresa said X” never becomes “X” by ingestion. Only measurement or corroboration promotes. - Numeric
confidence:frontmatter on hypothesis pages is retired — a hand-set float goes stale silently; the tag distribution on a page is its live confidence signal.
Vocabulary migration (old → standard)
The signal-set page’s local vocabulary maps as: Evidenced → [evidenced] · Decoded → [fact] (it met the reproducible-on-demand bar) · Candidate → [hypothesis] (if testable) or [idea] · Tested negative / Refuted → [refuted] · Operational prerequisite stays as a scope note inside the claim, not a maturity state.
Why
- The failure mode this prevents: an agent reads a hedged idea, misses the hedge, and builds a premise on it. Hedges in prose don’t survive excerpting; an inline tag travels with the claim wherever it’s quoted (the ICD 203 / GRADE insight: uncertainty must live in the running text, since readers — human or LLM — may never see metadata).
- Grounded in a research pass (2026-07-06) across the Admiralty/NATO A1–F6 system, ICD 203 analytic standards, Wikipedia verifiability/quality grades, ADR/RFC lifecycles, GitLab maturity tiers, GRADE medical evidence certainty, and emerging LLM-wiki conventions. Convergent findings: one axis with ≤5 discrete states (two-axis systems and fine granularity fail even for trained analysts); dates over adjectives (self-decaying); promotion gated on evidence, not opinion (the IETF rule — no promotion without implementations); labels costing >~10s at write time don’t get applied.
- Rejected alternatives: numeric confidence (unoperationalizable, goes stale silently — and agents’ own early guesses self-reinforce when re-read); two-axis source-reliability × credibility (Admiralty’s 36 cells — raters can’t sustain it); page-level quality grades (pages mix maturities; a page-level label lies).
Open questions
- Should the 90-day
[fact]staleness window differ by domain (platform-behavior facts drift faster than market facts)? - Wire
check-epistemic.shinto the dormant dream-cycle / VPS nightly pass when that goes live?
Timeline
- [2026-07-06] (decision — Vincent) Standard accepted: five claim-level tags with hard promotion gates, default reading rule, lint gate, LOG audit trail. Retrofit order: hypotheses pages + product-thesis first, opportunistic elsewhere. Prompted by the risk of agents construing ideas as facts; designed from the signal-set page’s organically-evolved vocabulary + external research.