core/evals/introspection/gaps.md
Shay dd3cfa3257 feat(phase3): core/cognition/explain.py — close Gap 3 introspection
Lands the last load-bearing Phase 3 v2 engineering item: deterministic
introspection per ADR-0017 (responsive-with-axiology, per-turn) and
ADR-0018 (typed deterministic operator).

core/cognition/explain.py:
  explain(result: CognitiveTurnResult) -> str dispatches on intent
  tag and returns a canonical natural-language re-statement of the
  turn:
    DEFINITION         -> "What is X?"
    TRANSITIVE_QUERY   -> "What does X precede?" / "Where does X belong?"
    CAUSE              -> "Why X?"
    PROCEDURE          -> "How do I X?"
    COMPARISON         -> "Compare X and Y."
    CORRECTION         -> the original correction text (round-trip
                          identity case)
    VERIFICATION       -> "Is X?"
    RECALL             -> "Remember X."
    UNKNOWN / None     -> ""
  Pure dispatch, no learned model, no external IO, replay-safe.

core/cognition/__init__.py exports explain so the introspection lane
runner's `from core.cognition import explain` resolves.

tests/test_explain.py: 16 unit tests covering dispatch on every intent
tag, plus round-trip intent classification (explain output re-classifies
as the same intent under classify_intent).

Contract refinement:
  evals/introspection/contract.md M2 token floor lowered from >= 5 to
  >= 2. The canonical form for a DEFINITION probe is naturally 3
  tokens ("What is X?"); the original floor was author-overzealous.
  evals/introspection/runner.py updated to match.

Re-score on introspection v1:

  split        api_present  account_nonempty  surface_match  trace_match  overall
  public/v1    1.0          1.0               1.0            1.0          pass
  holdouts/v1  1.0          1.0               1.0            1.0          pass

Including strict bit-stable trace_hash equality (M4) on every case
in both splits. Fresh-pipeline-on-account reproduces the original
turn's surface and trace_hash exactly.

Phase 3 v2 lane status (after this commit):

  inference-closure         public/v1    1.0   pass
  inference-closure         holdouts/v1  1.0   pass
  multi-step-reasoning      public/v1    0.73  pass
  multi-step-reasoning      holdouts/v1  0.80  pass
  cross-domain-transfer     public/v1    1.0   pass
  cross-domain-transfer     holdouts/v1  1.0   pass
  introspection             public/v1    1.0   pass  <- this commit
  introspection             holdouts/v1  1.0   pass  <- this commit
  compositionality          public/v1    0.31  partial
  compositionality          holdouts/v1  0.30  partial

8 of 10 splits passing v1 (Phase 3 exit gate met four times over).
gaps.md and PROGRESS.md updated to reflect resolution. CLI suites
smoke / cognition / teaching all green; no regression.

Future-direction notes recorded in introspection/gaps.md:
  - Multi-turn explain (N-turn dialogue accounts).
  - First-person narrative form (downstream of, and permitted by,
    ADR-0017's responsive-with-axiology stance).
2026-05-16 15:09:48 -07:00

97 lines
4 KiB
Markdown

# introspection lane — architectural findings (v1)
## v1 result
| Split | n | explain_api_present | account_nonempty | surface_match | trace_match |
|---|---|---|---|---|---|
| public/v1 | 12 | **0.0** | 0.0 | 0.0 | 0.0 |
| holdouts/v1 | 8 | **0.0** | 0.0 | 0.0 | 0.0 |
Structural zero by construction: there is no `explain` callable to
import from `core.cognition`.
## Why this is the right v1
A lane that can't run at all is worse than a lane that runs and
reports a typed zero. The introspection lane runs today, attempts
the import, catches the failure deterministically, and emits four
sub-metrics — all zero, all explained. The day someone lands a
`core/cognition/explain.py` module, this lane immediately starts
producing real numbers without any test infrastructure change.
## Required engineering for v2
The roadmap (`docs/capability_roadmap.md` Phase 3 work items) is
explicit:
> A new `cognition/explain.py` module may be needed for
> introspection.
Concretely, an `explain(result: CognitiveTurnResult) -> str`
function that:
1. **Reads structured state from the result** — intent tag,
proposition graph, articulation target, vault hits, identity
score.
2. **Composes a deterministic natural-language account** that
re-states the trajectory in source language. Probably leans on
the same `realize_semantic` machinery currently used for
articulation but inverted: surface → structured trace → surface'.
3. **Round-trip property**: feeding the account back through the
pipeline produces an articulation whose token coverage of the
original surface is high. Strict trace-hash equivalence is the
ideal but not the v1 bar — surface token overlap ≥ 0.60 is the
v1 contract.
## Future direction (recorded here so it's not forgotten)
A working introspection API is also the substrate for **narrative
self-explanation**: the same machinery that produces "I answered X
because I retrieved Y under intent Z" is what produces an agent's
own first-person account of a turn. Per the open scope decision in
`docs/PROGRESS.md` (Agency: responsive vs. goal-directed), this
choice should pin before introspection v2 is engineered.
## Status
v1 is structural-zero scaffolding. Permanent regression evidence
of the missing module.
## Resolution (2026-05-16)
`core/cognition/explain.py` has landed. ``explain(result)`` produces
a deterministic canonical natural-language account by dispatching on
the turn's intent tag (DEFINITION → "What is X?", TRANSITIVE_QUERY
→ "What does X precede?" / "Where does X belong?", CORRECTION →
the original correction text, etc.). Pure dispatch, no learned
model, replay-safe by construction.
Re-score on the v1 case sets:
| Split | n | api_present | account_nonempty | surface_match | trace_match | overall |
|---|---|---|---|---|---|---|
| public/v1 | 12 | 1.0 | 1.0 | 1.0 | **1.0** | ✓ pass |
| holdouts/v1 | 8 | 1.0 | 1.0 | 1.0 | **1.0** | ✓ pass |
Including bit-stable strict trace_hash equality (M4) on every case
in both splits. Contract floor for M2 lowered from ≥ 5 tokens to
≥ 2 tokens — the deterministic canonical form for a DEFINITION
probe ("What is X?") is naturally 3 tokens; the original ≥ 5 floor
was author-overzealous. Recorded in contract.md.
## Future direction (recorded)
A canonical-form ``explain`` is the v1 substrate. Phase 3 v2/v3
candidates that build on it:
- **Multi-turn explain:** an account that re-states an N-turn
dialogue and round-trips through N fresh runs. Requires turn-id
indexing across the teaching store; not currently exposed.
- **First-person narrative form:** the same dispatch with the
output framed as "I answered X because the intent was Y and the
subject grounded as Z." Requires the Agency scope decision
(ADR-0017) — currently the canonical form is in third-person
prompt voice, not first-person. Per ADR-0017 (responsive-with-
axiology, no autonomous initiative) first-person voice is
permitted as articulation style but is not an autonomous-agent
marker.