v0.5.0-kaizen — Hephaestus Kaizen Engine

v0.5.0-kaizen — Hephaestus Kaizen Engine

Explore freely. Prove ruthlessly. The loop now improves itself.

This is the largest release since the project rebrand. The Kaizen Engine turns
Hephaestus from a hypothesis-driven improvement loop into a self-improving system
by adding a third face (Reflector), a belief workspace (Global Workspace analog),
containment for new deadlock modes, four idea channels that supply diversity at
volume, and crash-safe runtime infrastructure.

All 21 v0.5 issues (#52–#72) closed via PR #73.
191 tests pass. Praxis evidence bundle: .praxis/runs/v05-full-evidence.jsonl.

Highlights

• Belief workspace (beliefs.yaml) — narrow, capacity-limited, single-writer
  (Reflector). Hard cap 12 active beliefs. Every belief carries a kill_criterion
  (falsifiability contract). Capacity competition for entry.
• Reflector agent — the third face. Shared-assumption extraction, relaxation
  probe, perspective tour. Decorrelated model family from orchestrator. Runs
  offline (idle / plateau), like human sleep consolidation. Writes only
  beliefs.yaml and narrative.md — never code, never hypotheses.
• relies_on field + blame propagation — the load-bearing change. When a
  hypothesis FAILs, blame propagates upward to every belief it relied on. Three
  failures on B-X → B-X becomes a suspect. Schema-breaking: required on all
  new hypotheses from this release forward.
• Containment engine — authority matrix (28 role pairs, terminal referee
  for every pair), suspect TTL (no permanent lockout), curiosity exemption
  (Red Team blocks MERGE, never RUN), belief min-residency + eviction evidence,
  frame-shift cooldown + ratchet hysteresis.
• Four idea channels — analogy (cross-project lessons + casting), random-leap
  (bisociation replay + dream mode + external entropy), affect (Brier calibration
  • mood modulation + boredom rotation), context (whisper inbox + morning
    briefing + regime detector). All independently feature-flagged, default
    disabled, separate daily budgets. Disabled = zero spend, zero artifacts.
• Provenance tagging — every hypothesis + merge carries a provenance tag.
  Per-channel hit-rate reports. A merge without provenance fails schema validation.
• Tick transaction journal — durable, atomic write-and-rename, phase tracking,
  operation-key idempotency, stale-lock recovery. Proves idempotent across
  worker dispatch, blame propagation, merges, lessons, provenance, spend
  accounting, channel output, and reflector consolidation.
• Canary suite — 11 end-to-end scenarios: belief capacity 12/13, TTL expiry,
  direct blame, authority coverage failure, ratchet restart, prompt-injection
  isolation, disabled-channel zero side effects, deadlock recovery, eviction,
  reflector shadow isolation, duplicate-tick idempotency, crash recovery.

Phase 0 — debt cleared (#52, #53, #54, #55)

• T2 Challenger / T3 Arbiter personas — templates/SOUL.challenger.md (read-only,
  blind, decorrelated, one rebuttal round) and templates/SOUL.arbiter.md (raw-evidence-
  only, binding merge/reject, falsifiability). Bootstrap wiring.
• Self-grade diff gate — templates/scripts/self-grade-diff.py. Orchestrator
  verdict cannot be more optimistic than mechanical evidence verdict. Wired into tick.
• Pre-registration lock — templates/scripts/prereg-lock.py. Locks metric,
  threshold, direction, relies_on at dispatch. SHA-256 tamper detection.
• Repo hygiene — schema/tests_tmp/ removed; tests promoted to schema/tests/;
  .gitignore tripwires for Windows path artifacts.

Phase 1 — Belief workspace + Reflector (#56–#60)

• schema/beliefs.schema.json — capacity cap 12, kill_criterion, blamed_by,
  historical_beliefs[].
• relies_on on hypotheses → blame propagation → suspect detection.
• narrative.md — one-page project story, rewritten each consolidation.
• SOUL.reflector.md + reflector-dispatch.sh — three questions per run.
• Stagnation / momentum signals in state.json; suspect-belief hard rule in
  orchestrator SOUL and prompts/tick.md.

Phase 2 — Containment (#61–#64)

• Authority matrix — 28 role pairs, every blocking state has a terminal referee
  with timeout + safe-default HOLD.
• Suspect TTL (default 5 ticks) — no permanent frame lockout.
• Curiosity exemption — belief-test hypotheses run from protected budget.
• Belief min-residency (M=3) + eviction audit log.
• Frame-shift cooldown (K=8) + ratchet hysteresis (one notch per R-window,
  direction reversal requires full window of evidence).

Phase 3 — Idea channels (#65–#68)

• Analogy — ~/.hermes/lessons.jsonl cross-project corpus with dual writing
  (concrete + denominalized abstract) + casting sessions.
• Random-leap — bisociation replay (mid-distance ledger pairs), dream mode
  (seeded entropy, generation/filtering separated), external entropy (arXiv/HN).
• Affect — Brier calibration (confidently-wrong is costly, agreement NOT
  rewarded), mood as parameter modulation, boredom → family rotation.
• Context — whispers/ inbox (30s, weakest evidence), morning briefing,
  regime detector. Security: external streams earn workspace entry through
  capacity + competition, never write to beliefs.yaml directly.

Phase 4 — Provenance + measurement (#69)

• provenance field on hypothesis + merge records (elimination | analogy | replay | whisper | external | relaxation).
• Per-channel section in reports/<date>-tick.md template.
• Rolling hit-rate aggregator script.
• A merge without provenance fails schema validation.

Phase 5 — Runtime reliability (#70, #71, #72)

• feature-flags.py — gradual rollout with safe defaults (shadow mode). Active
  blocked until readiness-check.py passes.
• tick-journal.py — durable transaction journal. Atomic write-and-rename,
  phase tracking, operation-key idempotency, stale-lock recovery.
• channel-budget.py — atomic daily idempotent accounting with file locking,
  stable operation keys, UTC reset.
• tick-runtime.py deterministic wrapper with journal phase transitions;
  tick-gate.sh calls tick-runtime.py init as mandatory pre-tick hook.
• channel-dispatch.py — deterministic channel dispatcher.
• .github/workflows/v05-ci.yml — full test matrix on every PR.
• Canary suite — 11 scenarios, 19 assertions.

Bootstrap integration

• bootstrap.sh — DF_REFLECTOR_* defaults, adapter extraction, template vars,
  reflector profile install, v0.5 script copy, extra ledger dirs, registry update.
• bootstrap.ts — AdapterConfig.reflector_model/chain, decorrelation check
  against orchestrator, profile install, v0.5 script copy, registry update.
• adapters/v7-alphaforge/project.yaml — reflector_model/reflector_chain
  fields (deepseek ≠ claude orchestrator, satisfies decorrelation invariant).
• templates/prompts/tick.md — Phase 4 channel dispatch + reflector dispatch.

Invariants (enforced by tests)

• 191 tests pass across schema/tests/ (41 correction + 19 canary + 131 legacy).
• A merge without provenance fails schema validation.
• bootstrap.sh bash syntax verified (bash -n clean).
• Disabled channel = 0 USD spend, 0 artifacts, 0 proposals.
• Reflector writes only beliefs.yaml and narrative.md.
• Authority matrix covers all role pairs; new profiles fail bootstrap without entry.
• Provider-chain decorrelation: orchestrator ≠ worker ≠ challenger ≠ arbiter ≠
  reflector (bootstrap validateChainDecorrelation() fail-closed).

Migration from v0.4

bash migrate-ledger.sh .alphaforge/orchestrator/state.json --backup

migrate-ledger.sh converts v1 hypothesis files (empty relies_on) into v0.5
schema. The relies_on field is required (may be empty []) on all new
hypotheses from this release forward.

Install / upgrade

# Fresh install
curl -fsSL https://raw.githubusercontent.com/ddawnlll/hephaestus/main/install.sh | bash

# Upgrade existing
cd ~/.hephaestus && git pull && bash bootstrap.sh /path/to/project --adapter v7-alphaforge

What's next

• v0.6: channel dispatch hardening (deterministic pre-commit + cron-side enforcement).
• v0.6: reflector active-mode rollout (default shadow until readiness passes for
  30 consecutive ticks across all deployed projects).
• v0.6: cross-project lessons.jsonl consolidation endpoint.

Known limitations

• Reflector active mode requires readiness check + containment verified; default
  is shadow mode (reflector_proposals.yaml, never touches beliefs.yaml).
• Channel dispatch is wired in the tick prompt; deterministic channel-dispatch
  scripting (pre-commit hooks, cron-side enforcement) is a hardening opportunity.
• bootstrap.sh registry update uses a Python heredoc — bash bash -n passes,
  but prefer bootstrap.ts (Bun) for new adapter work.

---

Full Changelog: https://github.com/ddawnlll/hephaestus/blob/v0.5.0-kaizen/CHANGELOG.md
PR: #73
Praxis Evidence: .praxis/runs/v05-full-evidence.jsonl