Gate-fold episode record (third episode)

Captured 2026-08-18 by the session that performed the fold, with human inputs quoted verbatim from the live conversation. Nothing here is reconstructed.

This is the episode the workshop was waiting for: the retained epistemology entering operative review machinery. Two semantic review gates were authored, installed with explicit rests-on lineage to the theory notes, and validated by live review runs.

Exact human inputs

The fold directive, with a discuss-first constraint:

OK - now lets work on folding the epistemology into the writing machinery - discuss it first.

The commission, deciding scale (the agent had recommended starting with one gate) and the validation design:

add the two gates - then run the validation with Opus sub-agent for the validation run all semantic gates not just the two new - so that it is more similar to the most common way of running it

The discriminative experiment:

can we run the gates on the previous version of the note?

Acceptance and the calibration follow-up:

OK - record it and revise the gate

Agent contributions

  • Coverage map. Surveyed the existing semantic gate suite and mapped it onto the theory: semantic/explanatory-reach already implements the structural threshold; grounding-alignment audits cited-support routes; no gate implemented the epistemic threshold. The two-threshold distinction from the warrant note thereby became the division of labor between the old gate and the new one.
  • Gap identification. Named the two uncovered failure modes: central-claim scope exceeding presented support with no transfer relation, and epistemic-status collapse (including support spreading to neighbouring claims), the latter operationalizing the previously unenforced mixed-epistemic-status note.
  • Gate design. Authored semantic/unwarranted-scope and semantic/epistemic-status-blur: failure modes, tests, the pass conditions (stated transfer relation or conjecture marking carries surplus scope; informal status markers suffice), neighbour boundaries, and examples.
  • Sequencing argument. Recommended gates before COLLECTION.md or skill changes — measurable through the review pipeline, and clean of the separate skill-revision claim.
  • Execution. Pipeline run and off-pipeline historical run, both with Opus workers; finalization; the post-run boundary clarification prompted by worker feedback.

Human contributions

  • The fold directive itself and its timing.
  • The scale decision: both gates at once, against the agent's one-gate-first recommendation.
  • The validation design: run the full semantic bundle, not just the new gates, "so that it is more similar to the most common way of running it".
  • The discriminative experiment: testing the gates against the previous version of the calibration note was the human's idea, and it produced the episode's strongest result.

Installation evidence

  • Gates committed as 4e935c24; boundary clarification committed with this record.
  • Both gates carry rests-on footer edges to Natural-language theories carry warrant claim by claim and scope by scope; epistemic-status-blur also rests on Mixed epistemic status must be preserved below the document level, unwarranted-scope also on the derivation-and-inheritance note. The theory-to-machinery lineage is explicit in the artifacts.

Validation results

  • Pipeline run (review job 7183, all 10 applicable semantic gates, worker model claude-opus-5[1m], partition claude-opus-5): all PASS on the current calibration note (agent-context-is-constrained-by-soft-degradation-not-hard-token-limits.md), with discriminating INFO findings from both new gates (the "whenever" universal quantifier as the furthest-reaching transfer; a mechanism-bridge sentence marked a section away from its status marker).
  • Historical A/B (off-pipeline, same model, note at dc67ccca, 2026-07-29): epistemic-status-blur returned WARN on exactly the flat "shared mechanism" sentence that commit 36adf3cf (2026-08-03) later fixed by adding the quarantined working-hypothesis section. The gate independently detected a defect the note's real revision history fixed, and registers the fix as the INFO-grade residue on the current version. unwarranted-scope passed both versions: the old version already carried its transfer relations and interaction-surface boundary, so the historical defect was status blur, not scope overreach.

Evidential status

This episode establishes installation: the retained epistemology now operates as review machinery, with explicit lineage. The A/B result is retrodictive validation — the gate discriminates a real past defect from its fix — not yet the prospective improvement the compounding pathway's third step requires. That step needs a later writing episode whose outcome is better because these gates ran.

One attribution correction recorded for accuracy: the fix commit 36adf3cf predates the theory episode, so the current calibration note's clean pass is not evidence that the theory improved that note; it is evidence that the gates recognize the difference between the defect and the fix.