Skip to content

v3.2.0 — Evidence-based upgrade: form-matching, judge calibration, pressure testing

Latest

Choose a tag to compare

@smixs smixs released this 03 Aug 09:01
· 1 commit to main since this release

Evidence-based upgrade: every change in this release traces to a published paper or a battle-tested community skill, and the release itself was produced by Conductor evaluating and improving itself (3 gated iterations, dual-family judges — Claude + GPT via codex, final score 0.971 with all critical questions passing).

What changed

  • Principle #10 — Match the form to the failure. Classify the baseline failure before writing a rule: prohibitions bulletproof discipline failures but measurably backfire on shaping failures. No nuance clauses, exemption clauses don't scope. From obra/superpowers writing-skills wording tests + polarity data from Guardrails Beat Guidance.
  • Critique-before-verdict judges. Grader, comparator and bineval agents write the detailed evidence critique BEFORE committing to the verdict, with a borderline few-shot example in each. From hamelsmu/evals-skills by Hamel Husain.
  • Threshold-blind judging. The judge no longer computes the overall score or the GATE — the orchestrator aggregates. A judge that knows the bar is biased toward it. From neolabhq/context-engineering-kit.
  • Automatic cross-family judge calibration. A second judge from a different model family answers the same bank; stable disagreement flags a badly worded question, not a dispute. Grounded in CheckEval, Prosa and the self-preference-bias study.
  • Variance discipline. Improvements on non-critical questions count only when they reproduce in 2 consecutive runs; the 3-iteration cap now carries its rationale from TICK. Judge-swing numbers from grafana/skills skill-authoring.
  • references/pressure-testing.md — micro-test protocol (no-guidance control, 5+ reps, variance as a metric) + pressure scenarios for discipline skills (7 pressure types, forced A/B/C choice, rationalization tables). From obra/superpowers.
  • Pushy description formula. [What] + Use when [4-5 phrasings] + "even if they don't explicitly say '<canonical term>'" + Do NOT use for [...], deduped to a single canonical home. Pattern from grafana/skills / Anthropic's skill-creator guidance.
  • Question bank v1.1 — 5 new questions: pushy triggers, nuance clauses, directive reference loading, time-rot language, redundant content via the Expert/Activation/Redundant taxonomy from softaworks/agent-toolkit skill-judge, actionability test from SkillReducer.
  • Freedom-consequence test in architecture selection ("if the agent makes a mistake here, what is the consequence?") — from softaworks/agent-toolkit.
  • Self-eval fixes: unified held-out terminology, directive reference loading, removed a python3-fallback contradiction, hardcoded model ids dropped, question-source contract reconciled (hybrid = deterministic + fixed bank).

Credits

Community skills: Jesse Vincent (obra/superpowers) · Hamel Husain (evals-skills) · Grafana (skill-authoring) · Softaworks (skill-judge) · Neolab (context-engineering-kit) · Trail of Bits (skill-improver) — stop-hook pattern referenced.

Research: Guardrails Beat Guidance · TICK · CheckEval · Prosa · Self-Preference Bias · LLMs Cannot Self-Correct Reasoning Yet · BIG-Bench Mistake · SkillJuror · SkillReducer · IFScale · Prompt Design at Scale · Ask, Don't Judge (BinEval)

Full changelog and per-source attribution: README — Synthesized from.