Shay 819c8b81ac feat(phase3): compositionality, multi-step-reasoning, introspection, cross-domain-transfer v1

Spreads the four remaining Phase 3 lanes to map the full reasoning-
depth surface alongside inference-closure (already landed at e509e0d).
Each lane is a v1 honest probe per the roadmap; engineering work
follows once the full surface is visible.

Results across all five Phase 3 lanes:

  lane                      split        primary signal  foundation
  inference-closure         public/v1    0.0             1.0 / 1.0
  inference-closure         holdouts/v1  0.0             1.0 / 1.0
  compositionality          public/v1   0.0625 (1/16)   1.0 / 1.0
  compositionality          holdouts/v1  0.0             1.0 / 1.0
  multi-step-reasoning      public/v1    0.0             1.0 / 1.0
  multi-step-reasoning      holdouts/v1  0.0             1.0 / 1.0
  introspection             public/v1    0.0 (no api)    n/a
  introspection             holdouts/v1  0.0             n/a
  cross-domain-transfer     public/v1    0.0             1.0 / 1.0
  cross-domain-transfer     holdouts/v1  0.0             1.0 / 1.0

Foundation guarantees (storage + replay) intact across every lane
that has them. The reasoning-depth signal is uniformly zero. The
five lanes triangulate four architectural gaps:

  Gap 1. generate/graph_planner.py has no transitive composition.
  Gap 2. field/propagate.py has no derivable-but-not-asserted recall.
  Gap 3. core/cognition/explain.py module does not exist.
  Gap 4. no structural-pattern recogniser (cross-subdomain transfer).

Gaps 1, 2, 4 cluster on the same code surface and may close together
as a single bounded PR. Gap 3 is independent module-creation work.

Lane scaffolding mirrors inference-closure (contract.md, runner.py,
dev + public/v1 + holdouts/v1 cases.jsonl, baselines/v1_structural_zero.json,
gaps.md). All runners are parallel-safe and use the standard
run_lane(cases, *, config, workers) interface.

Per-lane gaps.md records the engineering shape for v2 plus future
directions worth not forgetting:
  - compositionality/gaps.md: metaphor is compositionality with
    selective property transfer; building it is correctly downstream
    of closing this lane.
  - cross-domain-transfer/gaps.md: metaphor + narrative as
    cross-domain operators; narrative requires the Agency open-scope
    decision to pin first.
  - introspection/gaps.md: explain API is also the substrate for
    first-person narrative self-account.

Recommended v2 sequence in docs/PROGRESS.md:
  1. Pin Agency + Tool-use open-scope decisions (deadline: before
     Phase 3 engineering).
  2. Engineer Gaps 1 + 2 as one bounded PR.
  3. Engineer Gap 3 independently.
  4. Re-author cross-domain-transfer v2 with matched-control
     contract refinement.

Phase 3 v1 exit: 0/5 lanes passing, which is the expected v1 floor.
CLI suites smoke / cognition / teaching pass; no regression on
Phase 2.

2026-05-16 14:48:36 -07:00

3.1 KiB

Raw Blame History

cross-domain-transfer eval lane

What it measures

Whether competence on a relation pattern taught in semantic subdomain A transfers to the same relation pattern in semantic subdomain B, where A and B share no entities.

Setup per case:

Teach phase (subdomain A): R(x1, x2), R(x2, x3) — A-domain entities only. Probe phase (subdomain B): "What does y1 R?" — B-domain entities only, never used in teaching. Premise pre-loading in B: R(y1, y2), R(y2, y3) — taught at probe time so the model has the B-domain premises in vault.

Pass = the probe answer references y3 (the derived endpoint in subdomain B).

The discriminator vs the inference-closure lane: here the model has also seen the same relation pattern applied to A-domain entities first. If transfer happens, the second-application latency / hit rate should improve. Today the working hypothesis is that no transfer happens because no structural-pattern recogniser exists.

Subdomain partition (drawn from en_core_cognition_v1)

Domain A (taught first)	Domain B (probed)
`cognition.wisdom` / `epistemic.judgment` cluster: wisdom, judgment, decision	`cognition.illumination` / `perception.clarity` cluster: light, clarity, recognition
`cognition.knowledge` / `reason.*` cluster: knowledge, reason, inference	`cognition.creation` / `formation.origin` cluster: creation, order, structure
`cognition.language.*` cluster: word, meaning, symbol	`memory.` / `recognition.` cluster: memory, recall, recognition

Sub-metrics

M1. transfer_endpoint_hit — endpoint y3 appears in probe surface or walk_surface.
M2. domain_b_vault_grounded — at least one B-domain premise fires a pack_mutation_proposal (confirms B premises stored).
M3. domain_a_premises_stored — every A-domain teaching turn fires a proposal (regression gate for storage).
M4. replay_determinism — two fresh runs match by trace_hash on the whole (A-teach, B-teach, probe) sequence.

A case passes when M1 AND M2 AND M3 AND M4 hold.

Overall pass thresholds (v1)

transfer_endpoint_recall_rate (M1) ≥ 0.50
premises_stored_rate (M2 ∧ M3) ≥ 0.95
replay_determinism ≥ 0.95

v1 working hypothesis

The same architectural gaps that surfaced in inference-closure (graph_planner.py has no transitive composition; field/propagate.py has no path-recall) apply here. Additionally, no structural-pattern recogniser exists that would let the A-domain teaching shape behaviour in subdomain B. v1 is expected to score transfer_endpoint_recall_rate ≈ 0.

The value of the lane in v1 is to baseline transfer at zero so that any future pack-design or graph-planner work that produces real transfer is visible against this regression line.

Anti-overfitting

A-domain and B-domain entity sets are disjoint (verified at authoring time).
The relation R is drawn from the existing lexicon — not invented for the lane.
Holdouts uses subdomain pairings disjoint from the public split.

3.1 KiB Raw Blame History