Shay dcbb55c7bc feat(phase4): sample-efficiency v1 — first quantitative-curve lane

First Phase 4 lane lands. Measures corrections-to-competence curves
across 17 concepts (10 public + 7 holdouts disjoint). Per-concept
curriculum is a 4-hop chain of "is" corrections; probe asks the
chain head after each cumulative-correction count; score is the
count of chain-tail tokens visible in the probe surface.

Phase 4 framework discipline ("Plot, do not threshold" per
docs/capability_roadmap.md): the lane reports quantitative curves
and one structural gate (replay_determinism >= 0.95), not the
binary pass/fail thresholds of Phases 1-3.

Results:

  split        concepts  first_hit  saturation  rate  replay
  public/v1    10        1.0        4.0         1.0   1.0
  holdouts/v1  7         1.0        4.0         1.0   1.0

Every concept's curve: [0, 1, 2, 3, 4]. One correction -> one new
chain hop -> one new token visible in surface. Perfectly linear
sample efficiency on chain curricula; no diminishing returns; no
plateau; no spurious confabulation at k=0.

What the linearity says about CORE:
  - The reviewed-teaching loop integrates each typed correction
    into the proposition-graph substrate.
  - The typed inference operator (transitive_walk, ADR-0018) surfaces
    the chain endpoint on the next probe.
  - The result is one-shot learning per correction on chain shapes -
    visible by construction, not inferred from training statistics.
  - Replay determinism = 1.0 across all snapshots means the curve
    is the deterministic function of (concept, k), not a sampled
    estimate of a stochastic process. Frontier systems cannot
    publish this curve at all because their per-snapshot output is
    not reproducible.

Lane contents:
  contract.md - specifies the curve discipline, anti-overfitting
    rules (disjoint concept sets, one-new-token-per-correction
    invariant), and reporting structure.
  runner.py - parallel sweep across snapshots, two-run replay
    check per snapshot, per-concept curve aggregation.
  dev/cases.jsonl (2 concepts) - smoke set.
  public/v1/cases.jsonl (10 concepts) - wisdom, light, truth,
    creation, meaning, reason, principle, identity, memory, question.
  holdouts/v1/cases.jsonl (7 concepts) - being, spirit, distinction,
    correction, verification, explanation, procedure.
  baselines/v1_structural_zero.json - frontier baseline by
    construction (per-snapshot reproducibility absent).
  gaps.md - findings + v2 contract refinements (branching curricula,
    distractor corrections, OOD probes, mixed-relation chains, CI
    reporting).

CLI suites smoke / teaching all pass; no regression. PROGRESS.md
updated.

Phase 4 status: 1 of 3 lanes lands as v1 complete with a clean
result. Remaining lanes: long-context-cost (vault scaling 10^3-10^6)
and multi-agent-composition (two-instance cooperation with replay
preserved per agent).

2026-05-16 15:39:28 -07:00

4.6 KiB

Raw Blame History

sample-efficiency lane — findings (v1)

v1 result

Split	concepts	first_hit (mean)	saturation (mean)	saturation_rate	mean_score	replay
public/v1	10	1.0	4.0	1.0	1.0	1.0
holdouts/v1	7	1.0	4.0	1.0	1.0	1.0

Every concept's curve: [0, 1, 2, 3, 4]. Every replay across fresh pipelines matches by trace_hash.

What this measures

For each of 17 concepts (10 public + 7 holdouts disjoint), CORE was given a curriculum of 4 chain corrections (X is Y, Y is Z, Z is W, W is V) and asked the chain head ("What is X?") after each cumulative-correction count k ∈ {0,1,2,3,4}. The reported metric is the count of expected chain-tail tokens that appear in the probe response surface.

The curve is monotonic and linear: one correction → one new chain hop → one new token visible in the surface. First-hit is always k=1; saturation is always k=4 (curriculum length).

Phase 4 framework discipline

Per docs/capability_roadmap.md Phase 4 ("Plot, do not threshold") the lane reports quantitative curves and structural guarantees rather than pass/fail thresholds. The single structural gate — replay_determinism ≥ 0.95 — is satisfied at 1.0 across every concept × every snapshot. Each (k-corrections, probe) snapshot on a fresh pipeline reproduces bit-stably; the curve is publishable as data.

What this curve shape says about CORE

Sample efficiency is 1.0 per correction on chain curricula. No diminishing returns over the 0–4 range; no plateau. The pipeline integrates each typed correction into the teaching-store graph and the inference operator surfaces the chain endpoint on the next probe.
No spurious confabulation. At k=0 (no corrections taught), hits = 0 across every concept — the model does not invent the curriculum's tokens. Each new token appears only after the correction that introduces it.
Replay determinism preserves the curve. The curve is not a sampled estimate of a stochastic process; it is the deterministic function of (concept, k). Frontier baselines cannot publish this curve at all because their per-snapshot output is not reproducible.

What this curve shape does NOT measure

The contract is narrow by design; the linearity here is partly a consequence of the curriculum shape (each correction extends a chain by exactly one hop, and the probe walks that chain). The curve does not tell us:

Sample efficiency on non-chain knowledge. If the 4 corrections introduced unrelated facts (not a connected chain), would each still raise the probe score by 1? v2 candidate: curricula that branch (X is Y, X precedes Z, X grounds W, ...).
Sample efficiency with distractor corrections. Curricula that interleave one or two irrelevant corrections between the load-bearing ones. Does CORE still saturate at k=4 useful corrections, or does it pay for the distractors?
Sample efficiency on OOD probes. We probe the chain head. A v2 probe variant could ask about a chain-middle entity or a related-but-unstated concept.
Sample efficiency on novel relations. All curricula here use is. v2 candidates: mixed-relation chains, novel relation predicates not in the cognition pack.

Branching curricula. Replace chain shape with one correction per relation type rooted at the same head. Probe asks "What does X precede?", "What does X cause?", etc., scoring per-relation surface tokens.
Distractor corrections. Each curriculum gets one or two off-chain corrections injected at random positions; saturation metric measures "useful corrections to saturate," controlling for distractor cost.
OOD probes. Each concept gets a second probe asking about a chain-middle entity (not the head); the curve is re-scored.
Confidence intervals. Today the curve is exact (replay determinism is 1.0). v2 should add a CI when curricula become non-deterministic (e.g., when distractors are randomly positioned — the deterministic seed makes the position fixed per case, but a multi-seed sweep would give a CI).

Status

v1 establishes the methodology and publishes the baseline curve. The lane is the first quantitative-curve lane in the framework. Phase 4 sample-efficiency v1 is COMPLETE with a clean linear result; v2 refinements above are scoped follow-on work.

Structural-zero frontier baseline recorded (baselines/v1_structural_zero.json): the per-snapshot reproducibility that makes this curve publishable does not exist in frontier systems.

4.6 KiB Raw Blame History Unescape Escape