core/evals/conversational_thread_coherence/runner.py
Shay e06fda5b8b feat(runtime+evals): warm-path pack grounding + three long-span lanes
Step 1 — warm_grounding_stability targeted patch
- chat/runtime.py:_maybe_pack_grounded_surface accepts allow_warm=True;
  warm path invokes it after articulation and overrides
  response_surface / articulation / grounding_source when pack-grounded
  or teaching-grounded.
- CAUSE / VERIFICATION without a teaching chain on warm path emits the
  unknown-domain disclosure (matches cold-path discovery-signal doctrine
  — no fabricated vault content).
- warmed_session_consistency public lane: warm_grounding_stability
  0.0 → 1.0, grounding_match_rate 1.0, telemetry_consistency 1.0.
- Cognition lane byte-identical (public 100/100/91.7/100, holdout
  100/100/83.3/100).  Full suite 2294 passed.

Step 2 — three new red eval lanes (measurement substrate)
- conversational_thread_coherence: 6 cases / 45 turns; per-turn
  no_placeholder / not_walk_fragment / length / is_grounded predicates
  + per-case topic_anchor and no_topic_drift.  Baseline: grounded
  0.93, topic_anchor 0.50, no_topic_drift 0.83.
- multi_sentence_response: 15 cases over Explain/Tell/Describe/Walk/
  Example/Essay shapes; predicates sentence_count >= 2, non-fragment,
  connective_present, subject_named.  Baseline: multi_sentence 0.53,
  connective 0.10 — biggest architectural gap.
- self_consistency_over_time: 7 cases; same probe at multiple turn
  indices with unrelated fillers interleaved.  Baseline: byte_identical
  0.86 (one CAUSE-no-chain disclosure drifts under accumulation).

All three lanes deterministic, lexical-predicate-only — no LLM judge,
no embedding similarity.  Red-on-creation by design.  See
notes/long_span_fluency_baseline_2026-05-19.md.
2026-05-19 08:26:38 -07:00

198 lines
6.3 KiB
Python

"""Conversational thread coherence eval lane runner.
Measures whether ``ChatRuntime`` maintains coherent grounding and
topic continuity across an 8-12 turn thread. Predicates are
deterministic and lexical — no LLM judge, no embedding similarity.
Framework contract: ``run_lane(cases, config=None) -> LaneReport``.
Case schema (``cases.jsonl`` line):
{
"id": "...",
"category": "...",
"turns": [
{
"prompt": "...",
"subject_lemma": "truth", # optional — for topic-anchor check
"expects_grounded": true, # default true
"anaphora_anchor_to": "truth", # optional — prior subject expected to appear
"is_replay_of_prompt_at_turn": 0 # optional — drift check
}
]
}
"""
from __future__ import annotations
import re
from dataclasses import dataclass, field
from typing import Any
from chat.runtime import ChatRuntime
_PLACEHOLDER_MARKERS = ("...", "<pending>", "<prior>", "<empty>")
_FINITE_VERB_RE = re.compile(
r"\b(is|are|was|were|has|have|had|does|do|did|will|would|can|could|"
r"should|might|may|must|shall|been|being|[a-z]+(?:es|ed|ing)s?)\b",
re.IGNORECASE,
)
def _check_no_placeholder(surface: str) -> bool:
return not any(m in surface for m in _PLACEHOLDER_MARKERS)
def _check_not_fragment(surface: str) -> bool:
tokens = [t for t in re.findall(r"[A-Za-z]+", surface)]
if len(tokens) < 4:
return False
return bool(_FINITE_VERB_RE.search(surface))
def _check_length_adequate(surface: str) -> bool:
return len(surface.strip()) >= 20
def _check_is_grounded(grounding_source: str, expects: bool) -> bool:
if not expects:
return True
return grounding_source in {"pack", "teaching", "vault", "oov", "partial"}
@dataclass(frozen=True, slots=True)
class TurnResult:
turn_index: int
prompt: str
surface: str
grounding_source: str
no_placeholder: bool
not_walk_fragment: bool
length_adequate: bool
is_grounded: bool
topic_anchor_satisfied: bool | None # None when not applicable
@dataclass(frozen=True, slots=True)
class CaseResult:
case_id: str
category: str
turn_results: tuple[TurnResult, ...]
no_topic_drift: bool
@dataclass
class LaneReport:
metrics: dict[str, Any] = field(default_factory=dict)
case_details: list[dict[str, Any]] = field(default_factory=list)
def _run_case(case: dict[str, Any]) -> CaseResult:
rt = ChatRuntime()
turns: list[TurnResult] = []
grounding_by_prompt: dict[str, list[str]] = {}
for idx, turn in enumerate(case["turns"]):
prompt = turn["prompt"]
expects_grounded = bool(turn.get("expects_grounded", True))
anaphora_anchor = turn.get("anaphora_anchor_to")
resp = rt.chat(prompt)
surface = resp.surface
grounding = resp.grounding_source or "none"
anchor_ok: bool | None = None
if anaphora_anchor:
anchor_ok = anaphora_anchor.lower() in surface.lower()
turns.append(TurnResult(
turn_index=idx,
prompt=prompt,
surface=surface,
grounding_source=grounding,
no_placeholder=_check_no_placeholder(surface),
not_walk_fragment=_check_not_fragment(surface),
length_adequate=_check_length_adequate(surface),
is_grounded=_check_is_grounded(grounding, expects_grounded),
topic_anchor_satisfied=anchor_ok,
))
grounding_by_prompt.setdefault(prompt, []).append(grounding)
# No topic drift: any prompt that repeats must produce the SAME
# grounding tier on every firing (pack/teaching once → pack/teaching
# always). Drops to `none` after a successful grounding indicate
# state corruption.
no_drift = True
for srcs in grounding_by_prompt.values():
if len(srcs) <= 1:
continue
strong = {"pack", "teaching"}
if any(s in strong for s in srcs) and any(s == "none" for s in srcs):
no_drift = False
break
return CaseResult(
case_id=case["id"],
category=case.get("category", "uncategorised"),
turn_results=tuple(turns),
no_topic_drift=no_drift,
)
def run_lane(cases: list[dict[str, Any]], config: Any = None) -> LaneReport: # noqa: ARG001
if not cases:
return LaneReport(metrics={}, case_details=[])
results = [_run_case(c) for c in cases]
total_turns = sum(len(r.turn_results) for r in results)
def _rate(pred: str) -> float:
passing = sum(1 for r in results for t in r.turn_results if getattr(t, pred))
return round(passing / total_turns, 4) if total_turns else 1.0
anchor_turns = [t for r in results for t in r.turn_results
if t.topic_anchor_satisfied is not None]
anchor_rate = (
round(sum(1 for t in anchor_turns if t.topic_anchor_satisfied) / len(anchor_turns), 4)
if anchor_turns else 1.0
)
metrics: dict[str, Any] = {
"cases": len(results),
"total_turns": total_turns,
"no_placeholder_rate": _rate("no_placeholder"),
"not_walk_fragment_rate": _rate("not_walk_fragment"),
"length_adequate_rate": _rate("length_adequate"),
"is_grounded_rate": _rate("is_grounded"),
"topic_anchor_rate": anchor_rate,
"no_topic_drift_rate": (
round(sum(1 for r in results if r.no_topic_drift) / len(results), 4)
if results else 1.0
),
}
case_details = [
{
"case_id": r.case_id,
"category": r.category,
"no_topic_drift": r.no_topic_drift,
"turns": [
{
"turn_index": t.turn_index,
"prompt": t.prompt,
"surface": t.surface,
"grounding_source": t.grounding_source,
"no_placeholder": t.no_placeholder,
"not_walk_fragment": t.not_walk_fragment,
"length_adequate": t.length_adequate,
"is_grounded": t.is_grounded,
"topic_anchor_satisfied": t.topic_anchor_satisfied,
}
for t in r.turn_results
],
}
for r in results
]
return LaneReport(metrics=metrics, case_details=case_details)