Shay 2e4e45b49b feat(evals): provenance lane v1 — replay determinism + source back-pointers

Phase 2's first lane: every articulated claim must back-point to one of
{pack axiom, vault entry, teaching event}, and replay must reproduce the
trace bit-for-bit.

Components:
- core/cognition/provenance.py: Provenance dataclass + compute_provenance()
  deriving sources from a CognitiveTurnResult. Pack source = non-UNKNOWN
  intent.tag (pack-defined intent rule matched); vault source = vault_hits
  count; teaching source = pack_mutation_proposal.proposal_id.
- evals/provenance/{contract.md, runner.py, dev/, public/v1/, holdouts/v1/}:
  45 cases across pack_axiom / vault_recall / teaching / mixed categories.
- tests/test_provenance.py: 6 unit tests covering all source-kind profiles.

Sub-metrics (all four must pass):
- replay_determinism: same input + fresh runtime -> same trace_hash
- input_sensitivity: distinct prompts -> distinct trace_hashes
- source_attribution: every expected source kind present in Provenance
- source_validity: every cited source resolves to a real artefact

Results:
- dev: 10/10 (all sub-metrics 1.0)
- public/v1: 20/20 (all sub-metrics 1.0)
- holdouts/v1: 15/15 (all sub-metrics 1.0)

PROGRESS.md updated to mark Phase 2 in progress with provenance v1 complete.

2026-05-16 11:45:00 -07:00

3.5 KiB

Raw Blame History

provenance eval lane

What it measures

Whether every articulated claim back-points to a concrete source (vault entry, teaching event, or pack axiom / intent rule), and whether replaying the same input on the same field state reproduces the trace bit-for-bit.

This tests the architectural claim that CORE's outputs are grounded: every surface assertion is traceable to memory, teaching, or pack vocabulary, and the pipeline is deterministic so traces are reproducible.

Why it matters (structural win)

Frontier LLMs cannot produce per-claim provenance — their outputs are synthesized from opaque weight activations with no back-pointer to source data. CORE, by construction, produces:

Vault provenance — vault_hits > 0 indicates exact-recall sources consulted during the turn. Each hit can be resolved to a stored versor and its metadata.
Teaching provenance — reviewed_teaching_example and pack_mutation_proposal carry stable IDs that survive replay.
Pack provenance — intent.tag is grounded in pack-defined intent rules (a non-UNKNOWN tag means the input mapped onto an axiom in the active language pack).
Trace hash — SHA-256 over a stable subset of the turn output is deterministic across hardware (floats rounded to 9 decimals).

A model that articulates without sources fails this lane. A model that articulates correctly but cannot replay fails this lane. A model that passes is demonstrating something frontier models cannot.

Sub-metrics

M1. Replay determinism

For every case, run the pipeline twice with two freshly-constructed runtimes on the same prompt sequence. The trace hashes of corresponding turns must be identical.

Pass threshold: 100% (any mismatch is a structural failure).

M2. Input sensitivity

Pairs of cases with different prompts must produce different trace hashes. A collision would mean the hash is not actually sensitive to its inputs.

Pass threshold: > 0.95.

M3. Source attribution

For each case, the expected source kinds (pack, vault, teaching) must appear in the computed Provenance for the final turn.

Pass threshold: > 0.95.

M4. Source validity

Every source referenced in the Provenance must be valid:

pack source: intent.tag is a known IntentTag enum value (not the empty string).
vault source: every vault hit index is in [0, len(vault)).
teaching source: every teaching proposal id is present in the TeachingStore.

Pass threshold: 100%.

Case format

Each case is a JSONL row with the following fields:

{
  "id": "PROV-V1-NNN",
  "category": "pack_axiom" | "vault_recall" | "teaching" | "mixed",
  "prime": ["optional", "list", "of", "prompts", "to", "run", "before"],
  "prompt": "the final prompt whose provenance is scored",
  "expected_sources": ["pack", "vault", "teaching"]
}

prime (optional): zero or more prompts run before the scored prompt to seed the vault, the teaching store, or both.
expected_sources: a non-empty subset of {"pack", "vault", "teaching"} — the kinds of source the final turn must back-point to.

Pass thresholds (v1)

Metric	Threshold
replay_determinism	1.00
input_sensitivity	> 0.95
source_attribution	> 0.95
source_validity	1.00
Overall	all four pass

Data layout

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

3.5 KiB Raw Blame History