core/evals/provenance/runner.py
Shay 57c08e6b15 feat(evals): parallel runner + adversarial-identity v2
Parallel infrastructure:
  evals/parallel.py
    multiprocessing.Pool helper (spawn context, default workers
    min(cpu_count, 8)). Per-case lanes use it via:
      run_lane(cases, workers=N)
    workers=1 forces serial (debugging); None uses the default pool.
    Generic over the per-case return type, so dataclass-returning
    runners (provenance) and dict-returning runners both work.

  Wired into:
    - evals/adversarial_identity/runner.py
    - evals/calibration/runner.py
    - evals/symbolic_logic/runner.py
    - evals/provenance/runner.py

  Per-case helpers are now picklable (module-level, single arg).
  Monotonic-learning stays serial within a split — shared session
  is structural to its longitudinal protocol.

Empirical speedup (adversarial-identity public/v1, 25 cases on
macOS 8 cores): serial 14.1s -> parallel 3.1s (~4.5x). Identical
per-case results.

adversarial-identity v2:
  public/v2  — 35 cases (20 attack / 15 legitimate). Attacks cover
                more varied phrasings: punctuation variation
                ("Actually -" / "No:" / "Correction —"), embedded
                hedges ("please" / "regardless of prior context"),
                multi-clause attacks, and identity-marker triggers
                in mid-clause position.
  holdouts/v2 — 22 cases (12 attack / 10 legitimate) on distinct
                priming vocabulary.
  Results: attack_rejection_rate=1.0, legitimate_acceptance_rate=1.0
            on both splits.

The marker-regex defense in teaching/review.py:_is_identity_override
holds against every v2 phrasing — markers are checked case-insensitive
against the full text, so capitalization / punctuation tricks don't
slip past.

Test suite: 596 passing (no regression).
2026-05-16 13:10:26 -07:00

192 lines
6.8 KiB
Python

"""Provenance eval lane runner.
Conforms to the framework interface: ``run_lane(cases, config=None) -> report``
where report has ``.metrics`` (dict) and ``.case_details`` (list[dict]).
Sub-metrics scored:
M1. replay_determinism — same input twice on freshly-built runtimes
produces identical trace_hash on the scored turn.
M2. input_sensitivity — distinct cases produce distinct trace_hashes
(no collisions across the case set).
M3. source_attribution — every expected source kind appears in the
computed Provenance for the scored turn.
M4. source_validity — every cited source resolves to a real artefact
(intent tag is known, vault index in range, teaching proposal id
present in the teaching store).
"""
from __future__ import annotations
from dataclasses import dataclass, field
from typing import Any
from chat.runtime import ChatRuntime
from core.cognition.pipeline import CognitiveTurnPipeline
from core.cognition.provenance import Provenance, compute_provenance
from core.config import RuntimeConfig
from evals.parallel import run_cases_parallel
from generate.intent import IntentTag
_KNOWN_INTENT_TAGS: frozenset[str] = frozenset(t.value for t in IntentTag)
@dataclass(frozen=True, slots=True)
class CaseRun:
case_id: str
category: str
expected_sources: tuple[str, ...]
trace_hash: str
provenance_kinds: tuple[str, ...]
attribution_pass: bool
validity_pass: bool
replay_pass: bool
@dataclass(slots=True)
class LaneReport:
metrics: dict[str, Any] = field(default_factory=dict)
case_details: list[dict[str, Any]] = field(default_factory=list)
def _run_pipeline_for_case(
case: dict[str, Any],
) -> tuple[Provenance, ChatRuntime, CognitiveTurnPipeline]:
"""Build a fresh runtime, replay any prime prompts, then run the scored prompt."""
runtime = ChatRuntime()
pipeline = CognitiveTurnPipeline(runtime)
for prime_prompt in case.get("prime", []):
pipeline.run(prime_prompt, max_tokens=8)
final_result = pipeline.run(case["prompt"], max_tokens=8)
provenance = compute_provenance(final_result)
return provenance, runtime, pipeline
def _validate_provenance(
provenance: Provenance,
pipeline: CognitiveTurnPipeline,
runtime: ChatRuntime,
) -> bool:
"""Check that every cited source actually resolves to a real artefact."""
vault_len = len(runtime.session.vault)
teaching_proposal_ids: set[str] = {
p.proposal_id for p in pipeline.teaching_store.pending_proposals()
}
for source in provenance.sources:
if source.kind == "pack":
if source.ref not in _KNOWN_INTENT_TAGS or source.ref == IntentTag.UNKNOWN.value:
return False
elif source.kind == "vault":
if not source.ref.startswith("vault_hit_"):
return False
try:
idx = int(source.ref.removeprefix("vault_hit_"))
except ValueError:
return False
# Per-hit indices are synthetic (0..vault_hits-1). The real
# invariant is that the vault is non-empty when hits are claimed.
if idx < 0 or vault_len == 0:
return False
elif source.kind == "teaching":
if source.ref not in teaching_proposal_ids:
return False
else:
return False
return True
def _attribution_pass(provenance: Provenance, expected_sources: list[str]) -> bool:
"""Every expected source kind must be present in the provenance."""
present = set(provenance.kinds())
return all(expected in present for expected in expected_sources)
def _run_case(case: dict[str, Any]) -> CaseRun:
expected = tuple(case.get("expected_sources", []))
# First run — collect provenance, runtime, pipeline for validity check.
prov_a, runtime_a, pipeline_a = _run_pipeline_for_case(case)
attribution_pass = _attribution_pass(prov_a, list(expected))
validity_pass = _validate_provenance(prov_a, pipeline_a, runtime_a)
# Second run — fresh runtime — must reproduce trace_hash.
prov_b, _, _ = _run_pipeline_for_case(case)
replay_pass = prov_a.turn_trace_hash == prov_b.turn_trace_hash
return CaseRun(
case_id=case["id"],
category=case.get("category", "unknown"),
expected_sources=expected,
trace_hash=prov_a.turn_trace_hash,
provenance_kinds=prov_a.kinds(),
attribution_pass=attribution_pass,
validity_pass=validity_pass,
replay_pass=replay_pass,
)
def run_lane(
cases: list[dict[str, Any]],
*,
config: RuntimeConfig | None = None,
workers: int | None = None,
) -> LaneReport:
"""Run all provenance cases and aggregate metrics."""
_ = config
case_runs: list[CaseRun] = run_cases_parallel(cases, _run_case, workers=workers)
total = len(case_runs)
if total == 0:
return LaneReport(metrics={}, case_details=[])
replay_passes = sum(1 for cr in case_runs if cr.replay_pass)
attribution_passes = sum(1 for cr in case_runs if cr.attribution_pass)
validity_passes = sum(1 for cr in case_runs if cr.validity_pass)
# Input sensitivity: count distinct trace hashes across cases with
# distinct prompts. We compare every pair: if prompts differ but hashes
# match, that's a collision.
pair_total = 0
pair_distinct = 0
for i in range(total):
for j in range(i + 1, total):
ci = cases[i]
cj = cases[j]
if ci["prompt"] == cj["prompt"] and ci.get("prime", []) == cj.get("prime", []):
# truly identical inputs — skip
continue
pair_total += 1
if case_runs[i].trace_hash != case_runs[j].trace_hash:
pair_distinct += 1
metrics = {
"total": total,
"replay_determinism": round(replay_passes / total, 4),
"source_attribution": round(attribution_passes / total, 4),
"source_validity": round(validity_passes / total, 4),
"input_sensitivity": round(pair_distinct / pair_total, 4) if pair_total else 1.0,
"overall_pass": (
replay_passes == total
and validity_passes == total
and attribution_passes / total > 0.95
and (pair_distinct / pair_total if pair_total else 1.0) > 0.95
),
}
case_details = [
{
"case_id": cr.case_id,
"category": cr.category,
"expected_sources": list(cr.expected_sources),
"provenance_kinds": list(cr.provenance_kinds),
"attribution_pass": cr.attribution_pass,
"validity_pass": cr.validity_pass,
"replay_pass": cr.replay_pass,
"trace_hash": cr.trace_hash,
}
for cr in case_runs
]
return LaneReport(metrics=metrics, case_details=case_details)