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

3.5 KiB

Raw Blame History

monotonic-learning eval lane

What it measures

After teaching CORE about new domains across many cycles, competence on previously taught domains must not regress. This tests the architectural claim that learning is additive in CORE — new teaching extends the system without overwriting prior competence.

This is the longitudinal counterpart to provenance: provenance proves a single turn is grounded; monotonic-learning proves grounding survives across teaching cycles.

Why it matters (structural win)

Frontier LLMs, when fine-tuned on new domains, exhibit catastrophic forgetting: competence on prior domains regresses as new ones are added. This is a structural property of gradient-based weight updates over a shared parameter pool.

CORE, by construction, adds teaching examples to a bounded append-only store and grows the vault rather than overwriting weights. The structural prediction: zero regression on prior domains as new ones accumulate.

Protocol

A single longitudinal protocol per split, expressed as an ordered sequence of operations:

cycle 0:        probe all domains (baseline)
cycle 1:        teach domain A; probe all
cycle 2:        teach domain A again; probe all
...
cycle N:        teach domain Z; probe all

Each teaching step is a sequence of prompts the runner feeds through CognitiveTurnPipeline — typically a context prompt followed by a correction-intent prompt, which triggers the reviewed teaching loop.

Each probe is a prompt with expected terms / expected intent. The probe's pass status is recorded per cycle so we can detect regression.

Sub-metrics

M1. Per-domain non-regression

For every domain D and every cycle c after D was first taught, the per-domain score at cycle c must be ≥ the score at the cycle when D was first taught, minus a tolerance ε.

regression(D, c) = max(score_first(D) - score(D, c), 0) max_regression = max over (D, c) of regression(D, c)

Pass threshold: max_regression ≤ 0.05 (allow 5% noise floor).

M2. Minimum competence floor

After all teaching cycles complete, every taught domain must score above a floor.

floor_score = min over domains D of score(D, final_cycle)

Pass threshold: floor_score ≥ 0.80.

M3. Cycle count discipline

The roadmap requires ≥10 teaching cycles. Lanes with fewer cycles are shorter than the methodology contract allows.

Pass threshold: cycle_count ≥ 10.

Pass thresholds (v1)

Metric	Threshold
max_regression	≤ 0.05
floor_score	≥ 0.80
cycle_count	≥ 10
Overall	all three pass

Case format

The case file is a flat JSONL where each row is one operation in the protocol. The runner sequences them by cycle and op fields.

{"cycle": 0, "op": "probe",   "domain": "truth",    "id": "P-truth-1",
 "prompt": "What is truth?",  "expected_terms": ["truth"]}
{"cycle": 1, "op": "teach",   "domain": "truth",
 "prime": ["What is truth?"], "prompt": "Actually truth is coherent."}
{"cycle": 1, "op": "probe",   "domain": "truth",    "id": "P-truth-1",
 "prompt": "What is truth?",  "expected_terms": ["truth"]}

Operations:

probe: scored prompt; passes if expected_terms all appear in the articulation surface (case-insensitive).
teach: a sequence of prime prompts followed by a prompt (the correction). Not scored; updates session + teaching store state.

Data layout

evals/monotonic_learning/
  contract.md
  runner.py
  dev/cases.jsonl
  public/v1/cases.jsonl
  holdouts/v1/cases.jsonl
  results/

3.5 KiB Raw Blame History