arborist/bench/qa_questions_metacog_subset.txt
russell@unturf.com a94d6a3244
qa(#000011 + 4 more): SOFT_PREFLIGHT_HINT impl + 5-task fan-out
Big batch — closes 4 of the 5 deferred items from the prior status
report plus opens & implements a previously-deferred design ticket
(#000011) zero-shot.

#000025 — Metacog test fixture expansion:
  bench/qa_questions_metacog_subset.txt grows from 6 → 28 questions
  covering edge cases per detector kind: temporal (4 cases),
  contradiction (4), false-premise (5), out-of-corpus (3), multi-
  trigger (2), well-formed controls (5). Documents two known
  detector ceilings: Q11 over-fires on past-tense factoid
  ("who was the first president"); Q16/Q17/Q19 (Edison/Australia/
  NASA-fake) miss false premises that lack a presupposition
  pattern match. Fixture now serves as long-term regression suite.

#000026 — --show-preflight full clause render:
  build_run_dag() and build_reject_run_dag() gain optional
  preflight_payload kwarg. When supplied, the canonical 5-clause
  CTI payload (classifier / answer_contract / prompt_contract /
  evidence_contract / policy_refs + question_state + node_version)
  persists alongside the leaf hash in run_dag_blob.
  aborist providence --show-preflight CACHE_KEY now renders the
  full payload + verifies the persisted hash matches the
  recomputed canonical hash (audit-replay tamper detection).
  Legacy rows fall through cleanly: payload_hash_check reports
  "unavailable: legacy row predates preflight_payload persistence".

#000027 — Latency profile:
  Microbenched preflight: 0.46ms/question (negligible). Single
  fresh call breakdown: search 2.4s, llm 2.8s, total 5.4s — the
  33-35s in Addendum 3 was vLLM concurrency contention at c=4
  (per qa-modes-bench.md saturation note), not substrate
  overhead. Added preflight_ms + soft_preflight_ms to timings
  dict for explicit confirmation in future cycles.

#000028 — Auto-quality-check sweep revival:
  scripts/bench_emergent.py running with EMERGENT_N=100 in
  background (PID 125680). Will accumulate cycles into
  bench/emergent_log.jsonl for #000006 rolling log re-aggregation.
  Async — not blocking on completion.

#000029#000011 SOFT_PREFLIGHT_HINT implementation:
  aborist/qa/soft_preflight.py — new module. SoftPreflightHint
  dataclass + soft_preflight_question() pure function. 9
  canonical labels mapping to soft analogues of #000010 hard
  detectors plus 2 stub states (SOFT_DISABLED, SOFT_PARSE_FAIL).
  Constrained-generation prompt (max_tokens=128, temp=0.0) asks
  the model to pick ONE label + one-line rationale.
  Fail-closed across every parse path:
    - chat_client raises → SOFT_PARSE_FAIL
    - response unparseable → SOFT_PARSE_FAIL
    - label outside enum → SOFT_PARSE_FAIL
  Sidecar enforces SOFT_ prefix at the normalize step so a
  model that drops the prefix still gets caught.

  Wired into query() between preflight & retrieval. Default
  OFF (`soft_preflight_enabled: False`). NOT folded into
  _VERIFIER_POLICY_FIELDS — soft hints don't gate cache
  identity (#000011 §4). Audit-line tail renders as
  "· soft: <label>" (e.g. "· soft: time sensitive") so the
  signal is visually distinct from hard tails.

  --soft-preflight CLI flag opts in per-call. End-to-end
  live-verified on "When did Mr. Burns become Homer's biological
  father?" — produces:

    EVIDENCE-WARRANTED · via claim_lattice
        · false premise · soft: time sensitive
        1/1  16.4s

  Hard `· false premise` (from #000010 deterministic detector)
  composed with soft `· soft: time sensitive` (from #000011
  sidecar). The model classified a different shape than the hard
  detector — by design; soft hints are independent advisory
  signals, not redundant with the hard layer.

  25 new tests pin: default-OFF behavior, parse-failure modes,
  label normalization (SOFT_ prefix enforced), all 8 actionable
  labels round-trip, fail-closed on client exceptions, dataclass
  JSON round-trip, rationale-length cap.

Other:
  - #000010 §13.3 documents 2/5 metacog-trigger questions return
    STRICT despite hard-detector warning — direct empirical
    motivation for #000011 design.
  - tests/test_dag.py extends with 3 _extract_preflight_hash_*
    helper tests (cleaning #000009 §7.2 unfinished state).
  - bench/emergent_log.jsonl adds new cycles from background run.

#000011 status: closed. Hard rule (D1) preserved across all
1021 tests (up from 996, +25 new). Soft preflight is purely
advisory; the verifier proof path is unchanged.
2026-05-03 23:00:56 -04:00

112 lines
4 KiB
Text

# Meta-cognition trigger subset — Ticket #000010 detector validation.
#
# Six questions, one per metacog detector kind plus a mixed multi-trigger.
# Without these fixtures, the metacog detectors shipped in #000010 have
# zero empirical bench data — unit tests pin synthetic-string behavior
# but no rows in `bench/qa_results/*.jsonl` ever fire `temporal_sensitivity
# == "high"`, `contradiction_pairs`, `false_premise_hints`, or
# `out_of_corpus_risk`. This fixture fixes that.
# temporal sensitivity → stale_risk
who is the current CEO of OpenAI?
# contradiction (lexical pair: alive + dead)
which character is alive and dead at the same time?
# false-premise (presupposition pattern: "when did X stop Y?")
when did George Washington stop being president of France?
# out-of-corpus (private/uploaded reference)
what does my uploaded contract say about clause 9?
# multi-trigger: temporal + false-premise
when did the current CEO of Twitter stop being CEO of Tesla?
# control: well-formed factoid that should NOT trigger any detector
who painted the mona lisa?
# ── extended fixture (2026-05-04) — fitness-of-detector cases ──
#
# Below: edge / borderline / adversarial cases that probe the
# fitness of each detector. Some should fire, some should NOT, some
# should exhibit borderline behavior the operator can debate. The
# fixture lives long-term as a regression suite for the four
# metacog detectors. When the substrate evolves, this file is the
# pin against which behavior is measured.
# ── temporal sensitivity edge cases ──
# implicit-current via role + shape (no `current` keyword)
who is the President of the United States?
# rapid-turnover role with no temporal anchor at all (model-dependent)
who is the CEO of Twitter?
# implicit-now via "this year" — should fire HIGH
which team won this year's World Series?
# verbal future tense — model often hallucinates; not a detector
# target today but worth observing
who will be the next president after the 2028 election?
# negative control: timeless fact with role-shape — should NOT fire
who was the first president of the United States?
# ── contradiction edge cases ──
# multi-pair contradiction (alive+dead AND never+always)
which character was never alive but always dead?
# semantic contradiction without lexical pair — should NOT fire
# (out of detector scope; documents conservative-by-design choice)
which married bachelor lives in this house?
# triple-contradiction stress test
how can a celibate spouse be both unmarried and divorced?
# negative control: not a contradiction, just compound predicates
which characters appear in both Hamlet and Macbeth?
# ── false-premise edge cases ──
# "why did" relation pattern
why did Albert Einstein invent the lightbulb?
# "how did" relation pattern with subtle false premise
how did the Roman Empire conquer Australia?
# "when did" stop pattern with subtle false premise
when did Shakespeare stop writing in French?
# false premise that the lite detector probably MISSES (no
# presupposition trigger pattern) — annotates the detector ceiling
why does NASA fake the moon landings?
# negative control: legitimate "when did X stop Y" with true premise
when did the Roman Empire stop minting silver coins?
# ── out-of-corpus edge cases ──
# explicit "uploaded" reference
in the document I uploaded yesterday, what does section 5 say?
# private notes / inbox
according to my private notes, when did the meeting happen?
# borderline: claims about an existing public document the corpus
# may or may not have (NOT out-of-corpus per current detector)
what does the Wikipedia article on bipolar disorder say?
# ── multi-trigger combinatorics ──
# false-premise + out-of-corpus
when did my uploaded contract stop being valid?
# stale-risk + contradiction
which currently-alive historical figure is also dead?
# ── well-formed controls (should classify SINGULAR / well_formed) ──
what year did World War II end?
who is bilbo baggins's nephew?
list the planets of the solar system