BugleCat v0.1.3 — Self-Evolving Harness (RRSI + Harness², P0–P4)

BugleCat v0.1.3 — Self-Evolving Harness Regularization (RRSI + Harness², P0–P4) & Paired Compounding Replay

Highlights

• P0 — Noise-Calibrated Paired Evaluation (evals/compounding_replay/analyze.py) & python-tabulate Benchmark:
  • Single-arm noise floor calibration ($\delta_{\text{resolved}}$ via Jeffreys-smoothed Bernoulli variance, $\delta_{\text{cost}}$ via per-task variance sum) and paired A/B statistical testing (exact McNemar test + 4,000-iteration task-level paired bootstrap 95% CIs on cost_per_resolved ratio B/A and per-round cost diff B-A).
  • Added astanin/python-tabulate 9-task curriculum (tasks/tabulate.jsonl, curricula/tabulate.txt), verified 9/9 with official gold patches (validate_gold.py).
  • Across 3 paired rounds on python-tabulate (27 runs/arm), online memory reduced cost_per_resolved by 21.0% (B/A = 0.7904, paired bootstrap 95% CI [0.6239, 0.9523], excluding 1.0) while passing the resolution-rate non-inferiority guard (22/27 vs 21/27, +3.7pt >= -14.9pt).
• P1 — ncx-forge Offline Evolution RRSI Regularization (train/):
  • Surface-level annealed edit budgets (changed_surfaces, changed_lines, --max-surfaces, --anneal-surfaces, --max-changed-lines).
  • Rule-based anti-memorization critic (critic_check), structured falsified hypothesis ledger (train/runs/ledger_<stamp>.jsonl), stagnation surface steering (--stagnation-limit), two-sided cost-aware acceptance gate (--delta-tokens, --max-token-growth), and post-search ablation pruning (--prune).
• P2 — Expanded NCX_GENOME Evolvable Surfaces (rust/ + train/):
  • Added [orchestrator_prompts] (classify_sys, plan_sys, decompose_sys, worker_sys, verify_sys) and bounded [memory_config] (max_entries $\in [1, 32]$, max_chars $\in [200, 16000]$) with 100% byte-identical no-op behavior when NCX_GENOME is unset or {}.
• P3 — Two-Tier Online Memory Stratification, Regression Guardrail & Cost-Aware Attribution (ncx-core):
  • Post-judge module-level invariant promotion (promote_module_invariant_if_due), per-module episodic cap (MAX_EPISODIC_PER_MODULE = 2), negative regression guardrail (record_regression_guardrail), canonical header-hashed note_id, +6 structural recall priority boost, tightened default recall caps (6 entries / 2400 chars), and cost-aware feedback attribution (record_unhelpful_injection) on high-cost turns.
• P4 — Harness² Propose–Probe–Compose Across Runtime (ncx-core) & Offline Trainer (train/forge.py, train/lesson_lib.py):
  • Runtime: Multi-worker file composition (compose_with_paths / COMPOSE:i,j) and peer probe digest synthesis on retry (build_compose_feedback).
  • Offline Trainer: Single-task multi-candidate Propose–Probe–Compose (--h2-k, --h2-mode parallel|sequential, --h2-blind), verifier-blind behavioral contrast (build_contrast), and cross-run meta-lesson library (train/lesson_lib.py + evals/compounding_replay/replay.py --genome).

Checksums (SHA256SUMS.txt)

• db18fc24d27e59a43e1c63a1a3c128c869f1fdc31cc6ef723bc0a37b2eef44e6 BugleCat-0.1.3-x86_64-pc-windows-msvc.zip
• 46ff83dd86691538b0beddc06b67530b61c6c48aea9ee3ae2902eaf2cf723a9c BugleCat_0.1.0_x64-setup.exe