core/evals/multi_sentence_response/runner.py
Shay 07fefb923c feat(evals): articulate/disclosure/unarticulate partition
Tightens the multi_sentence_response lane predicates so OOV
invitations and refusal disclosures can no longer be counted as
articulate capability.  Three new metrics partition the case space:

  articulate_sentence_rate  - >=2 sentences AND grounded in
                              {pack, teaching}.  Real capability.
  disclosure_sentence_rate  - >=2 sentences AND grounded in
                              {oov, refusal, none}.  Structural
                              multi-sentence from disclosure templates.
  unarticulate_rate         - <2 sentences regardless of source.

The three sum to 1.0 (modulo rounding) by construction.  The
doctrine-correct headline is now ``articulate_sentence_rate``;
``multi_sentence_rate`` is kept as a continuity metric only.

2 new tests pin: (a) the three-way partition is total and disjoint
(articulate + disclosure + unarticulate == 1.0); (b) OOV/refusal
disclosure surfaces contribute to disclosure_sentence_rate but
never to articulate_sentence_rate.

Live A/B on 21 cases under the new partition:

  flag off: articulate=0.0952, disclosure=0.0476, unarticulate=0.8571
  flag on : articulate=0.8571, disclosure=0.0476, unarticulate=0.0952

Planner lift is +76pp on articulate.  Disclosure stays flat across
the flag (the planner gate correctly leaves disclosure surfaces
alone).  The remaining 9.5pp unarticulate flag-on is the genuine
miss list (walkthrough + compound prompts) that the next two
landings will target.

contract.md updated to make articulate_sentence_rate the headline
and to document the partition explicitly.

cognition eval byte-identical: public 100/100/91.7/100.
smoke suite 67/67.
2026-05-19 12:13:44 -07:00

235 lines
8.4 KiB
Python

"""Multi-sentence response eval lane runner.
Measures whether ``ChatRuntime`` emits more than one sentence when
the prompt structurally calls for elaboration. Strips the trailing
provenance tag (``pack-grounded (...).``) before counting sentences
so the metric reflects substantive content.
Framework contract: ``run_lane(cases, config=None) -> LaneReport``.
Case schema:
{
"id": "...",
"category": "...",
"prompt": "Tell me about truth.",
"subject_lemma": "truth",
"expects_connective": true,
"priming_prompts": ["What is truth?"] # optional
}
``priming_prompts`` is an optional list run before the scored prompt
on the same ``ChatRuntime`` instance. Their responses are discarded;
only ``prompt`` is scored. Priming exists because the discourse
planner currently hooks the warm pack/teaching-grounded path (post-
vault), so a one-shot cold-start case cannot exercise it. Cases
remain backward-compatible — missing or empty ``priming_prompts``
yields the original cold-start behavior.
"""
from __future__ import annotations
import re
from dataclasses import dataclass, field
from typing import Any
from chat.runtime import ChatRuntime
_PROVENANCE_TAIL_RE = re.compile(
r"\s*(pack-grounded|teaching-grounded)\s*\([^)]+\)\.?\s*$"
)
_TRUST_DISCLOSURE_TAIL_RE = re.compile(
r"\s*No session evidence yet\.?\s*$"
)
_CONNECTIVES = (
"and", "because", "therefore", "which", "since", "also",
"furthermore", "however", "consequently", "thus", "so", "while",
"whereas", "moreover", "in turn",
)
def _strip_provenance(surface: str) -> str:
stripped = _PROVENANCE_TAIL_RE.sub("", surface).strip()
return _TRUST_DISCLOSURE_TAIL_RE.sub("", stripped).strip()
def _split_sentences(text: str) -> list[str]:
"""Split substantive sentences without treating domain dots as stops.
Pack and teaching surfaces often contain semantic-domain atoms such as
``cognition.truth`` or ``logos.core``. A raw ``period + whitespace``
splitter over-counts those atoms as sentence boundaries, especially in
older structured disclosures like ``logos.core. truth grounds ...``.
Treat a stop as sentence-final only when it is followed by whitespace and
an uppercase/digit opener, or by the end of the text. This keeps
``cognition.truth. In turn, ...`` as two sentences while preventing
lowercase domain continuations from inflating the metric.
"""
stripped = text.strip()
if not stripped:
return []
parts = re.split(r"(?<=[.!?])\s+(?=[A-Z0-9])", stripped)
return [p.strip() for p in parts if p.strip()]
def _alpha_tokens(text: str) -> int:
return len(re.findall(r"[A-Za-z]+", text))
def _has_connective(text: str) -> bool:
low = text.lower()
return any(re.search(rf"\b{re.escape(c)}\b", low) for c in _CONNECTIVES)
@dataclass(frozen=True, slots=True)
class CaseResult:
case_id: str
category: str
prompt: str
surface: str
grounding_source: str
sentence_count: int
each_sentence_long_enough: bool
connective_present: bool
grounded: bool
subject_named: bool
expects_connective: bool
primed: 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], config: Any = None) -> CaseResult:
rt = ChatRuntime(config=config) if config is not None else ChatRuntime()
# Run optional priming turns on the same runtime so the scored
# prompt executes on the warm pack/teaching path. Responses are
# discarded; only the scored prompt's response is measured.
priming = case.get("priming_prompts") or ()
primed = False
for prime in priming:
if not isinstance(prime, str) or not prime.strip():
continue
rt.chat(prime)
primed = True
resp = rt.chat(case["prompt"])
surface = resp.surface
grounding = resp.grounding_source or "none"
stripped = _strip_provenance(surface)
sentences = _split_sentences(stripped)
each_long = all(_alpha_tokens(s) >= 4 for s in sentences) if sentences else False
subj = case.get("subject_lemma", "").lower()
subj_named = (subj in surface.lower()) if subj else True
return CaseResult(
case_id=case["id"],
category=case.get("category", "uncategorised"),
prompt=case["prompt"],
surface=surface,
grounding_source=grounding,
sentence_count=len(sentences),
each_sentence_long_enough=each_long,
connective_present=_has_connective(stripped),
grounded=(grounding in {"pack", "teaching"}),
subject_named=subj_named,
expects_connective=bool(case.get("expects_connective", False)),
primed=primed,
)
def run_lane(cases: list[dict[str, Any]], config: Any = None) -> LaneReport:
if not cases:
return LaneReport(metrics={}, case_details=[])
results = [_run_case(c, config=config) for c in cases]
total = len(results)
multi = sum(1 for r in results if r.sentence_count >= 2)
non_frag = sum(1 for r in results if r.each_sentence_long_enough)
grounded = sum(1 for r in results if r.grounded)
named = sum(1 for r in results if r.subject_named)
conn_expected = [r for r in results if r.expects_connective]
conn_rate = (
round(sum(1 for r in conn_expected if r.connective_present) / len(conn_expected), 4)
if conn_expected else 1.0
)
# ``multi_sentence_rate`` historically counted any case with ≥ 2
# sentences regardless of grounding source. That admitted OOV
# teaching invitations and refusal disclosures into the headline
# capability number — fixed here by splitting into three honest
# buckets:
#
# articulate_sentence_rate — ≥2 sentences AND grounded in pack
# or teaching (real capability).
# disclosure_sentence_rate — ≥2 sentences but grounded in oov,
# refusal, none, etc. (structural
# multi-sentence from disclosure
# templates, not articulation).
# unarticulate_rate — <2 sentences regardless of source.
#
# ``multi_sentence_rate`` is retained as a continuity metric. The
# doctrine-correct headline is ``articulate_sentence_rate``.
_DISCLOSURE_SOURCES = {"oov", "refusal", "none"}
articulate = sum(
1 for r in results
if r.sentence_count >= 2
and r.grounding_source in {"pack", "teaching"}
)
disclosure = sum(
1 for r in results
if r.sentence_count >= 2
and r.grounding_source in _DISCLOSURE_SOURCES
)
unarticulate = sum(1 for r in results if r.sentence_count < 2)
metrics: dict[str, Any] = {
"cases": total,
"multi_sentence_rate": round(multi / total, 4) if total else 0.0,
"articulate_sentence_rate": round(articulate / total, 4) if total else 0.0,
"disclosure_sentence_rate": round(disclosure / total, 4) if total else 0.0,
"unarticulate_rate": round(unarticulate / total, 4) if total else 0.0,
"non_fragment_rate": round(non_frag / total, 4) if total else 0.0,
"grounded_rate": round(grounded / total, 4) if total else 0.0,
"subject_named_rate": round(named / total, 4) if total else 0.0,
"connective_present_rate": conn_rate,
}
primed_results = [r for r in results if r.primed]
metrics["primed_cases"] = len(primed_results)
if primed_results:
multi_primed = sum(1 for r in primed_results if r.sentence_count >= 2)
metrics["primed_multi_sentence_rate"] = round(
multi_primed / len(primed_results), 4
)
else:
metrics["primed_multi_sentence_rate"] = 0.0
case_details = [
{
"case_id": r.case_id,
"category": r.category,
"prompt": r.prompt,
"surface": r.surface,
"grounding_source": r.grounding_source,
"sentence_count": r.sentence_count,
"each_sentence_long_enough": r.each_sentence_long_enough,
"connective_present": r.connective_present,
"grounded": r.grounded,
"subject_named": r.subject_named,
"expects_connective": r.expects_connective,
"primed": r.primed,
}
for r in results
]
return LaneReport(metrics=metrics, case_details=case_details)