Skip to content

0.2.0: refutation contagion + analyst/critic skills + release hygiene - #28

Merged
nyashkn merged 3 commits into
mainfrom
claude/session-03914w
Jul 23, 2026
Merged

nyashkn merged 3 commits into
mainfrom
claude/session-03914w

Conversation

@nyashkn

@nyashkn nyashkn commented Jul 23, 2026

Copy link
Copy Markdown
Owner

What

Three related changes that package insight-kit for agent use and close the refutation story, cut as 0.2.0.

1. Refutation contagion — guard_refuted_inputs

The republish guard answered "was THIS claim refuted?". This extends it one edge outward: "does this claim's input_claims chain reach anything refuted?".

  • A claim is refuted-at-the-source when its claim_id carries a standing refutation from a sealed run, or a non-guard critic refutes its record in-run.
  • Taint propagates transitively along input_claims (a DAG): if U is refuted and D derives from U, and E from D, both D and E are flagged. A claim that is itself refuted-at-the-source is a taint source, not a contagion finding (the republish guard / the refuting critic already surfaces it).
  • Per contaminated draft/published claim: emit a marked contagion critic (carrying the ancestor provenance + the input_claims path) and apply a high-severity critique (V16 record-then-enforce).
  • Idempotent: deterministic per-target critic id, and the guard's own critics are excluded from the in-run seed, so a contagion critic can never re-seed its own target and shrink the findings list. Surface, never block.

Exports guard_refuted_inputs + ContagionFinding. 7 new tests: standing + in-run sources, transitive propagation, published downgrade, supported-ancestor clears, clean-chain no-op, idempotency, and the self-reseed regression.

2. Analyst + critic-council skills, over a harness-agnostic method

Packages how an agent uses insight-kit, keeping the method separate from the harness binding (gate logic stays L1 Python; archetype prompts stay data; the binding stays thin — so a pi binding can reuse the same spec).

  • docs/method/ — the canonical method. analyst-loop.md (discover → compose → run → seal → verify) and critic-council.md (the mechanical-vs-semantic audit split, a perspective-diverse lens panel, the majority-refute kill rule, and recording verdicts as gated critic claims + contagion propagation).
  • .claude/skills/insight-kit-analyst/ — producer binding + a verify_run.py self-check helper (provenance + lineage + standing-refutation report), exercised against a sealed growth_demo run.
  • .claude/skills/insight-kit-critic/ — critic-council binding: dispatches the archetypes, applies the kill rule, records verdicts via ik_claim_emit, and propagates them with guard_refuted_inputs.
  • .claude/agents/ik-critic-{correctness,methodology,staleness,support}.md — the four lenses as first-class subagents, each returning a stance.

Producer and critic are deliberately different agents — a producer must not grade its own homework.

3. Release hygiene

Version had drifted to 0.1.7 with no tags past v0.1.0a3 and no changelog.

  • CHANGELOG.md (Keep a Changelog + SemVer), starting at 0.2.0, capturing this branch's work plus the already-merged cross-run substrate, Hamilton lineage, Layer-2 metrics, and measure catalog.
  • Bump pyproject 0.1.7 → 0.2.0 (propagated into uv.lock).
  • release.yml: on a v* tag, verify the tag matches pyproject, build sdist + wheel, and open a GitHub Release whose notes are the matching CHANGELOG section. Release flow documented in CONTRIBUTING.

Testing

Full suite 931 passed, 1 deselected; ruff clean. The verify_run.py helper was run against a live sealed growth_demo run (reports registered_input for base measures, payload + input_claims for the Layer-2 metric).

Notes

🤖 Generated with Claude Code

https://claude.ai/code/session_01Joez4MiYjd6owPcRQNcKrv


Generated by Claude Code

claude added 3 commits July 23, 2026 16:57
Version had drifted to 0.1.7 with no tags past v0.1.0a3 and no changelog.
Establish release hygiene:

- CHANGELOG.md (Keep a Changelog + SemVer), starting at 0.2.0 which captures
  the cross-run substrate, Hamilton lineage, Layer-2 metrics, and the measure
  catalog shipped on this branch.
- Bump pyproject 0.1.7 -> 0.2.0.
- release.yml: on a v* tag, verify the tag matches pyproject, build sdist+wheel,
  and open a GitHub Release whose notes are the matching CHANGELOG section.
- Document the release flow in CONTRIBUTING.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Joez4MiYjd6owPcRQNcKrv
The republish guard answers "was THIS claim refuted?". guard_refuted_inputs
extends that one edge outward: "does this claim's input_claims chain reach
anything refuted?".

- A claim is refuted-at-the-source when its claim_id carries a standing
  refutation from a sealed run (source "standing") or a non-guard critic
  refutes its record in the current run (source "in_run").
- Taint propagates transitively along input_claims (a DAG). A claim that is
  itself refuted-at-the-source is a taint SOURCE, not a contagion finding —
  the republish guard / the refuting critic already surfaces it.
- Per contaminated draft/published claim: emit a marked contagion critic
  refuting it (carrying the ancestor provenance + the input_claims path) and
  apply a high-severity critique (V16 record-then-enforce).
- Idempotent: deterministic per-target critic id, and the guard's own critics
  are excluded from the in_run seed so a contagion critic can never re-seed its
  own target and shrink the findings list. Surface-never-block.

Exports guard_refuted_inputs + ContagionFinding from the gate. 7 tests cover
standing + in_run sources, transitive propagation, published downgrade,
supported-ancestor clears, clean-chain no-op, idempotency, and the self-reseed
regression. Full suite 931 passed.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Joez4MiYjd6owPcRQNcKrv
… method

Package insight-kit's usage as directly-invocable Claude Code skills, keeping
the method separate from the binding so a pi binding can reuse it.

- docs/method/: the harness-agnostic method. analyst-loop.md (discover →
  compose → run → seal → verify) and critic-council.md (mechanical vs semantic
  audit, perspective-diverse lenses, majority-refute kill rule, verdicts as
  gated critic claims + contagion propagation).
- .claude/skills/insight-kit-analyst/: producer binding + a verify_run.py
  self-check helper (provenance + lineage + standing-refutation report).
- .claude/skills/insight-kit-critic/: critic-council binding — dispatches the
  archetypes, applies the kill rule, records verdicts via ik_claim_emit and
  propagates them with guard_refuted_inputs.
- .claude/agents/ik-critic-{correctness,methodology,staleness,support}.md: the
  four archetype lenses as first-class subagents, each returning a stance.

Producer and critic are deliberately different agents — a producer must not
grade its own homework. Gate logic stays L1 Python; archetype prompts stay
data; the binding stays thin. verify_run.py exercised against a sealed
growth_demo run. Full suite 931 passed.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Joez4MiYjd6owPcRQNcKrv
@nyashkn
nyashkn merged commit 48ecd92 into main Jul 23, 2026
6 checks passed
@nyashkn
nyashkn deleted the claude/session-03914w branch July 23, 2026 17:34
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

2 participants