2 commits

Author	SHA1	Message	Date
Shay	075169c33c	feat(evals): v2 lanes — monotonic-learning + provenance ... monotonic-learning v2: public/v2 — 5 domains × 3-4 probes × 20 cycles (377 ops) domains: truth, light, wisdom, order, memory max_regression=0.0, floor_score=1.0 holdouts/v2 — 4 distinct domains × 3-4 probes × 18 cycles (284 ops) domains: creation, knowledge, reason, spirit max_regression=0.0, floor_score=1.0 Demonstrates the structural claim (zero regression on prior domains as new ones accumulate) at substantially deeper cycle count and broader domain breadth than v1. provenance v2: public/v2 — 30 cases across pack_axiom, vault_recall, teaching, mixed deeper priming (3-5 turns), mixed-kind cases combining pack + vault + teaching sources in one probe source_attribution=1.0, source_validity=1.0, replay_determinism=1.0, input_sensitivity=1.0 holdouts/v2 — 20 cases on distinct vocabulary all sub-metrics 1.0 Generator: scripts/generate_monotonic_cases.py extended with three extra domain probe sets (order, memory, reason, spirit) and split definitions for v2.	2026-05-16 13:03:28 -07:00
Shay	632a69db40	feat(evals): monotonic-learning lane v1 — no regression across cycles ... Phase 2's second lane: after N teaching cycles in unrelated domains, competence on previously-taught domains must not regress. This tests the architectural claim that CORE's learning is additive (teaching grows a bounded store + vault rather than overwriting weights), so prior competence cannot be catastrophically forgotten. Protocol per split: cycle 0: probe all domains (baseline) cycle 1..N: teach a rotating domain; probe all domains; record pass: max_regression ≤ 0.05, floor_score ≥ 0.80, cycle_count ≥ 10 Components: - evals/monotonic_learning/{contract.md, runner.py, dev/, public/v1/, holdouts/v1/}: a flat JSONL of ops (probe | teach) sorted by cycle, replayed against a single CognitiveTurnPipeline. - scripts/generate_monotonic_cases.py: regenerates the cycle/probe corpora deterministically per split. Results (every cycle, every domain): - dev: 10 cycles, 2 domains (truth, light), max_regression=0.00, floor_score=1.00. - public/v1: 12 cycles, 3 domains (truth, light, wisdom), max_regression=0.00, floor_score=1.00. - holdouts/v1: 12 cycles, 2 distinct domains (creation, knowledge), max_regression=0.00, floor_score=1.00. Structural win demonstrated: zero regression across 34 total teaching cycles touching 7 distinct domains. PROGRESS.md updated to mark monotonic-learning v1 complete.	2026-05-16 11:56:34 -07:00