Fable Brain

Confidence Tag Policy

4 min read

Confidence Tag Policy

Every domain claim in this vault carries exactly one confidence tag assigned under five fixed rules from the confidence tags reference, which privilege the corpus and dated Anthropic URLs and demote folklore.

What it is

  • A decision record adopting the confidence tags reference as the binding assignment policy for this brain.
  • Four tags exist: EVIDENCE-BASED (controlled research or authoritative primary data), PRACTITIONER (operationally useful applied judgment, not proven by studies), CONTESTED (mixed, disputed, or failed-replication evidence, always with a caveat), and FOLKLORE (popular claim without credible support, demoted and never presented as fact).
  • The usage rule is strict: every domain claim carries exactly one tag, and a bestseller or popularity is not evidence.
  • Fable 5 is new and public coverage is thin, so the policy adds five domain-specific assignment rules.

How it works

The five assignment rules, verbatim in intent:

  1. A claim sourced only from the system prompt export is EVIDENCE-BASED with source_type: primary and a mandatory line reference in the L123-145 format; it is evidence of that capture, not a universal guarantee, and the note must say which harness (claude.ai) it describes.
  2. A claim corroborated by a dated public Anthropic URL (docs, news, engineering) is EVIDENCE-BASED.
  3. Plausible operational advice without official confirmation is PRACTITIONER.
  4. An export-versus-docs conflict is CONTESTED until resolved by a newer official source, with a > [!contradiction] callout in the owning note.
  5. Anything past its refresh_due date in the source ledger is stale and must be re-verified before being cited as current.

Rule 1 works jointly with Corpus Scope Decision: the corpus is a claude.ai capture, so rule 1 evidence never generalizes to Claude Code.

FOLKLORE has no assignment rule because it never qualifies as support; it appears only inside explicit demotion sentences.

Best practice

  • Tag claims, not notes: the frontmatter confidence summarizes, but each guidance bullet ends with its own tag. PRACTITIONER
  • Use dated Anthropic pages such as the models overview (https://platform.claude.com/docs/en/about-claude/models/overview) or the launch announcement (https://www.anthropic.com/news/claude-fable-5-mythos-5) as rule 2 anchors. EVIDENCE-BASED
  • When a claim sits between rules 2 and 3, downgrade to PRACTITIONER rather than inflating to EVIDENCE-BASED. PRACTITIONER
  • Pair every rule 4 CONTESTED tag with a > [!contradiction] callout in the owning note so the conflict is visible in graph queries. PRACTITIONER
  • Never let FOLKLORE claims surface as assertions; quote them only to demote them explicitly. PRACTITIONER
  • Re-verify anything past its ledger refresh_due date before citing it as current, per rule 5. PRACTITIONER
  • Record source_type: primary on rule 1 claims so queries can separate capture evidence from documentation evidence. PRACTITIONER

Pitfalls

  • Tagging a whole note once and skipping per-claim tags, which hides mixed-confidence content.
  • Treating popularity, virality, or bestseller status as evidence; the policy explicitly rejects this.
  • Citing the corpus without a line reference; rule 1 makes the L-format reference mandatory.
  • Forgetting the harness statement on rule 1 claims, which silently overgeneralizes claude.ai evidence.
  • Resolving a rule 4 conflict by preference instead of waiting for a newer official source.
  • Citing a stale ledger entry as current because the note text still reads plausibly.

Sources

Next actions

  • Add a lint check that flags corpus citations missing an L-format line reference.
  • Audit existing notes for CONTESTED tags lacking a contradiction callout.
  • Review the source ledger refresh_due dates at the next brain refresh.

Graph View

Backlinks