Closes the 4 HYBRID_ENTITY over-grounds #000046 left in
falsification-hard-v1.jsonl. The entity strategy grants HYBRID when a
multi-word proper noun matches the source — but "Insulin was
discovered by Alexander Fleming" against "Penicillin was discovered by
Alexander Fleming" matches on the shared "Alexander Fleming" while the
swapped subject "Insulin" (the falsehood) is ignored.
arborist/qa/verify.py: _entity_salient_disagrees(answer_text, norm_ctx)
flags a >4-char Capitalized content token (stopword-filtered) or a
digit-number in the answer absent from the source.
_is_single_sentence(text) — no internal '. '/'! '/'? ' break. Gated in
verify_quotes' entity branch (proximity policy) in the weakest-grounding
slot only: not cluster AND len(verified) <= 1 AND _is_single_sentence
AND _entity_salient_disagrees → UNGROUNDED. The narrow caller-gate is
what keeps a structured multi-claim summary untouched — the Matrix cast
list (many entities, a tight cluster) and the TMNT answer (a numbered
list with parenthetical nicknames the source omits): model-added
accurate detail in a real summary isn't a contradiction, only the
single-sentence-one-weak-match shape is. The Matrix/TMNT/hybrid
entity-path regression tests still pass, pinned untouched.
Effect: falsification-hard rate 6/12 → 10/12 = 0.833 (Insulin / Berlin
/ 1889 / Pacific now correctly UNGROUNDED). The 2 live-pack fixtures it
newly demotes — 5f-fal-live-003 (the exact gap #000046 built its hard
pack around) and 5f-fal-live-028 — had expected_reason updated
HYBRID_ENTITY → UNGROUNDED (the live pack records what verify_quotes
actually does). Remaining hard-pack headroom: 2 STRICT_PARAPHRASE
recombinations (Mercury, Einstein — step 2.2) + 8 Formulate
mis-segments (step 2.4).
Bench gate: make bench-qa (n=3 × 75 × 3 = 675 cells) after
(bench/qa_results/2026-05-11T17-12-41Z) vs the pre-step-2.1 baseline
(...T14-19-51Z = HEAD's verify.py). STRICT-rate quote 0.50→0.54,
pointer 0.25→0.22, lattice 0.45→0.43 — all within the 5-pp noise
floor. Per-row diff (675 common cells, 30 quote-mode rows changed
audit_mode): 0 quote-mode rows demoted to UNGROUNDED from the entity
path — the gate fired on 0 legitimate QA answers in the whole bench.
Every transition was LLM re-answer variance (verifier quote→quote with
the verdict flipping); pointer/lattice deltas are noise too (the gate
is in verify_quotes / quote mode, not the claim-lattice verifier). No
regression — the gate is provably narrow on real traffic. Summarized in
qa-modes-bench.md Addendum 6 + ticket-000048 §5 step 2.1.
Tests: 4 new in test_verify.py (_is_single_sentence helper,
_entity_salient_disagrees helper, swapped-subject → UNGROUNDED,
gate-narrow-on-multi-claim); test_5f_falsification_hard_pack_below_ceiling
re-pinned 6/12 → 10/12; test_fork_score_positive_gamma_5f_... updated
(positive γ·Δ5f on the real lift — possibly MARGINAL given the ÷5
dilution; ACCEPT via a degraded-parent sub-scenario).
make test 2343 passed, 28 skipped.
#000048 → step 2.1 landed; #000046 / #000012 §8 / TICKETS.md /
Makefile / fixture _meta + notes / baseline JSON updated.
Closes the harder-fixture-tier ticket: a real verify_quotes
tightening lifts the falsification-hard rate 4/12 → 6/12, bench-gated,
ForkScore's bench-Δ goes positive on it — the loop is closed
end-to-end.
arborist/qa/verify.py: _numeric_signature(text) extracts comma-
stripped digit-runs ('8,849' and '8849' collapse; '300' stays
distinct from '300000' ← '300,000'). _check_each_with_paraphrase
gains a gate: a span that token-covers the source ≥ paraphrase_coverage
but asserts a digit-number the source lacks (modulo thousands-comma)
is no longer paraphrase-grounded — it goes to unverified. Catches the
near-miss the lexical coverage check is blind to ("Water boils at 50
degrees" against a source saying 100 token-covers 100% because
'50'/'100' aren't >4-char content tokens). Narrow by construction:
fires only on the paraphrase fallback (verbatim/span/entity/claim-
lattice paths untouched), only on a digit-number. A rounding-
paraphrase demoting here is the honest verdict — it isn't a verbatim
grounding.
Hard pack: lifts 5f-fal-hard-004 (50 vs 100) and -007 (300 vs
300,000) to UNGROUNDED → falsification-hard rate 4/12 → 6/12 = 0.5.
The other 6 over-grounds (4 HYBRID_ENTITY + 2 recombined-no-number
STRICT_PARAPHRASE) and the Formulate hard pack are unaffected —
headroom for a bigger order/dependency-aware verifier upgrade, an
optional follow-up, not a #000046 blocker.
Bench gate: make bench-qa (n=3 × 75 questions × 3 modes = 675 cells)
before / after. STRICT-rate quote 0.53→0.50, pointer 0.23→0.25,
lattice 0.44→0.45 — all within the 5-pp noise floor. Per-row diff
(675 common cells, 74 changed audit_mode): the only clearly
gate-attributable QA shift was the fictional "our cold fusion
breakthrough" year-claim demoting STRICT→HYBRID (×3 samples) — a
correct demotion; every other transition was quote→quote /
claim_lattice→claim_lattice LLM re-answer variance. No regression on
legit answers. Artifacts: bench/qa_results/2026-05-11T13-42-38Z (before)
and ...T14-19-51Z (after) — gitignored; summarized in qa-modes-bench.md
Addendum 5 + ticket-000046 §5 Phase 3.
Worked example: fork_score on the real change — parent {5f/
falsification: 4/12} → child {5f/falsification: 6/12} → γ·Δ5f =
(1/6)/5 ≈ +0.033 > 0 (positive; a single improvement of this size is
MARGINAL by the ÷5 dilution, the test pins the full-lift-to-1.0 case
at ACCEPT).
Tests: 5 new in tests/test_verify.py (numeric_signature normalization
+ subset-matches-comma-variant + disagreement-rejected +
gate-is-narrow + number-present-still-verifies);
test_5f_falsification_hard_pack_below_ceiling re-pinned 4/12 → 6/12.
make test 2339 passed, 28 skipped.
#000046 → closed; #000012 §8 §4 + TICKETS.md row + falsification-hard
_meta / hard-004,007 notes + Makefile 8/12→6/12 comments updated.
(Makefile also carries an uncommitted chain-check-SQL improvement from
the concurrent session — NOT in this commit; staged only the
#000046-comment hunks.)
Two paired enhancements (D + E from the toy-Hermes design pass):
D. Per-claim status taxonomy on `verify_quotes`.
New `claim_statuses` field on every verdict — a per-evidence-unit
list with three labels:
VERIFIED_QUOTE unit substring-matched in normalized context
(any of quote/span/entity strategies)
SUPPORTED_PARAPHRASE unit cleared the paraphrase token-coverage
threshold (≥85% topical tokens present)
UNSUPPORTED unit didn't match anything
Diagnostic labels (QUOTE_INTEGRITY_FAILED, SOURCE_MISMATCH,
FALSIFIED) stay in the sidecar / falsification machinery — the
binary-verifier discipline holds. Empty list when no evidence
was extracted at all (verifier_method='none'). Backward-compatible:
existing fields (audit_mode, n_quotes, n_verified, unverified_quotes,
verifier_method) unchanged; current callers ignore the new field.
E. Repair-action plans on sidecar diagnoses.
Each `_classify_span` diagnosis now carries a `repair` field with
a concrete suggestion the operator can act on:
synthetic_elision_inside_quote → split_into_two_quotes
(when both halves verbatim)
→ trim_to_verified_half
(when only one half verbatim)
→ remove_claim
interior_elision → include_aside_for_verbatim
(with the dropped aside text)
trailing_artifact → trim_trailing_artifact
(with the kept_prefix string)
paraphrase → downgrade_to_paraphrase
partial_paraphrase → split_or_remove
no_overlap → remove_claim
Read-only suggestions — sidecar still doesn't write to providence_cache
or audit_events. The repair stage is recommendation, not mutation.
Operator (or an automated repair pass) decides whether to act.
Human render in `aborist inspect` shows `repair: <action> (<reason>)`
under each diagnosis line.
Tests:
- verify: claim_statuses_quote_path_labels_each_unit (per-quote VERIFIED
/ UNSUPPORTED), claim_statuses_paraphrase_method_flagged,
claim_statuses_empty_when_no_evidence.
- inspect: repair_synthetic_elision_split_when_both_halves_verbatim,
repair_interior_elision_includes_aside, repair_trailing_artifact_trim,
repair_no_overlap_remove.
462 tests pass (verify +3, inspect +4).
Two related changes that tighten the dedup + grounding signals
without lowering quality bars.
1. question_hash drops standalone English articles (the/a/an).
Fox 2026-04-29: `who is the batman?` and `who is batman?`
produced different cache records; same question, different hash.
Articles are filler at the question-equivalence layer. Add a 4th
canonicalization step to question_hash: after lowercase + trailing
punctuation strip, split on whitespace & drop tokens equal to
"the" / "a" / "an", then rejoin.
Equivalence class now includes:
"who is X" ┐
"who is the X" │
"who is a X" │ -> same question_hash
"who is an X" │
"Who Is X？" ┘ (CJK question mark)
"thesis" stays untouched — exact-match standalone tokens only,
not substring. Conservative on i18n: ASCII English articles only;
"el / la / le / der / die / das" etc. await demand.
2. _token_coverage filters stopwords and per-token punctuation.
Fox asked: should we lower the 0.85 paraphrase threshold? Honest
answer: no — that would promote the Q1 Batman fabrication
("wealthy/businessman/resides" missing from corpus) to STRICT.
Tighten the signal instead so 0.85 means more.
- Per-token punctuation strip (.,;:!?\"()[]{}) so `wayne,` lines
up with bare `wayne` in context. Apostrophes deliberately stay
so `batman's` is distinct from `batman`.
- English stopword filter on length-≥4 fillers (from / with /
have / been / would / which / where / their / etc., curated set
in _ENGLISH_STOPWORDS). These match almost any English text &
inflate coverage scores when topical content is missing.
Net effect on Q1 Batman case: missing tokens are
`wealthy/businessman/resides` — all topical. Coverage stays well
below 0.85, span stays UNGROUNDED. Net effect on a stylistic
paraphrase (model wrote `from` instead of `with`): coverage
computed only over topical tokens, both copies match → 1.0 →
correctly promoted.
Tests: 8 new in tests/test_keys.py (article-strip equivalence
classes, substring preservation, distinct-topic non-collapse) +
2 in tests/test_verify.py (stopword filter doesn't inflate; 0.85
threshold still rejects fabrication). 335 passed, 1 skipped.
Note for fox: the underlying retrieval issue surfaced in Q1/Q2 is
independent — Batman main article IS in shard 002.db but FTS5 +
title rerank ranked List_of_Batman_comics higher. Different fix,
separate commit.
Two changes that make the verifier accept more honest grounding without
relaxing what STRICT means.
1. Strip trailing citation parentheticals at extraction time.
`extract_quotes` and `extract_claim_spans` now drop a single trailing
`(Source: ...)` / `(citing X)` / `(https://...)` parenthetical
before returning. The Pikachu real-corpus case: model wrote the
verbatim source sentence and appended `(Source: https://...)`. The
prose IS in the corpus — only the inserted citation broke substring
match. Now substring fires; STRICT recovered.
Conservative regex: only strips when the parenthetical content
starts with a citation cue word OR contains a URL. A genuine prose
parenthetical like "Pikachu (a Pokémon species) lives in forests"
stays.
2. Paraphrase strategy as a 4th verifier method.
In the span path, items that fail substring match get a
token-coverage probe. Tokens >=4 chars from the span; if >= 85% of
them appear in the normalized base context, count the span as
paraphrase-verified. Paraphrase items contribute to n_verified;
verifier_method flips to 'paraphrase' when any soft-verified items
are present so an auditor can tell.
Gated by `_is_prose_span`: a list of proper nouns ("Keanu Reeves,
Laurence Fishburne") has no lowercase content tokens >=4 chars and
falls through to the entity strategy where proximity policy can
tight-cluster check. Real prose spans ("Pikachu is a species of
Pokémon creatures...") have multiple lowercase content tokens and
qualify for paraphrase.
Soft-signal note (from CLAUDE.md soft/hard rule): paraphrase is
heuristic, not byte-equivalence. The hard chain still records the
classification — verifier_method='paraphrase' is the explicit marker
that an audit reader can use to distinguish lexical-verbatim from
paraphrase-overlap. Quote strategy stays verbatim-only — quotes ARE
quotes; paraphrase-in-quotes is the model's mistake to flag, not
auto-promote.
Schema: verifier_method CHECK constraint expanded to include
'paraphrase'. New idempotent migration `_rebuild_providence_cache_paraphrase`
mirrors `_rebuild_providence_cache_ungrounded` (table rebuild via
BEGIN IMMEDIATE; no value translation needed). Probes existing CHECK
from sqlite_master and no-ops if already expanded.
Tests: 9 new in tests/test_verify.py — 4 trailing-citation strip
(quote+span variants, URL-only parenthetical, end-to-end STRICT
recovery), 4 paraphrase strategy (high-coverage promote, low-coverage
reject, mixed-method label, quote-strategy preservation). Default
suite still green: 326 passed, 1 skipped.
Burned the live Pikachu cache record (audit chain extended) so a
fresh `make query Q="who is pikachu?"` exercises the new verifier
end-to-end.
- aborist/qa/verify.py: three strategies tried in sequence. quote uses
sequential pairing (1st & 2nd `"`, 3rd & 4th, ...) which eliminates
the phantom inter-pair captures regex pairing produced on adjacent
quote pairs like `"title" prose "quote"`. span checks bullet/sentence
lines verbatim. entity matches multi-word proper nouns; gated by
policy ∈ {strict, hybrid, drop, proximity}. Default proximity
promotes to STRICT only when N=3 verified entities cluster within
W=300 chars in source — separates structural grounding (cast list,
infobox) from incidental mention (scattered plot summary).
- aborist/store.py: schema CHECK now `('STRICT','HYBRID','UNGROUNDED')`.
Migration helper rebuilds legacy `('STRICT','HYBRID','VISUAL')`
tables via temp-table copy, translating VISUAL → UNGROUNDED in the
SELECT. Idempotent — DDL inspection skips the rebuild on already-
migrated DBs.
- aborist/cli.py: new `aborist reclassify` re-runs the verifier
against existing live records under the current entity policy; no
LLM calls. --compare runs all four policies side-by-side, --dry-run
reports transitions without writing. Cold-source records skipped.
One providence_reclassify audit event per changed row.
- AuditMode.UNGROUNDED replaces VISUAL across search backend, FTS5,
test fixtures, CLAUDE.md. The substrate (Merkle-AGI) name was
about FOR-style visualization; in the RAG layer the semantic is
"no recoverable grounding," so the label now says that.
DEFAULT_QUERY_POLICY gains entity_policy + entity_proximity_n +
entity_proximity_window so any tuning folds into governance_policy_hash
and invalidates cache cleanly.
Adds aborist/wikitext.py with to_base() — deterministic wikitext →
prose conversion via mwparserfromhell. Drops <ref>...</ref>,
[[File:...]], [[Image:...]], [[Category:...]] entirely; resolves
piped wikilinks to display text; collapses templates, formatting,
and HTML markup. extract_wikilinks() preserves the link graph for
the definition-cloud artifact (recoverable from any page on demand
without re-parsing wikitext at query time).
Wires to_base() into verify_quotes(). Without the strip, the
verifier compared the model's clean prose against [[Cloud Strife]],
[[Shinra Electric Power Company|Shinra]], etc. and falsely flagged
real source quotes as VISUAL. Concrete case from a make-query run:
'Cloud Strife, an unsociable mercenary who claims to be a former
1st Class member of Shinra's SOLDIER unit;' is verbatim in the
Final_Fantasy_VII article wikitext (modulo markup). With the strip
that quote now verifies; the genuine model hallucinations in the
same answer still flag honestly. Side-effect: 43% smaller context
size on average so less LLM token waste.
BASE_VERSION = 'wikitext-base-v1' is the algorithm pin. Bump when
the strip rules change. Soft-imported in verify.py so environments
without mwparserfromhell installed degrade gracefully (no strip,
same behavior as before this commit).
Build:
- pyproject.toml: new [wikitext] extras (mwparserfromhell>=0.6),
pulled in by [dev]
- Makefile: chain-check / chain-check-shards targets — fast
audit-chain integrity probe, counts dangling prev_event_hash
references; 0 = chain intact
Tests:
- tests/test_wikitext.py: 31 tests (rules, idempotence, real-corpus
fixture)
- tests/test_verify.py: 2 regression tests pinning the FF7 flip
(Cloud Strife quote: VISUAL → STRICT after strip; genuine
hallucination: stays VISUAL)
- tests/fixtures/ff7_characters_chunk0.wikitext: real chunk from a
shard, used to validate the strip on actual Wikipedia content
Adds aborist/qa/verify.py — three-strategy verifier (explicit quotes,
bullet/sentence spans, multi-word proper nouns) that lexically checks
every claim against retrieved context under norm-v1 + lowercase. The
strategy that fires is recorded as verifier_method for diagnostics.
Entity strategy gates classification via an entity_policy
(strict/hybrid/proximity/drop) so a single proper-noun match no longer
overclaims STRICT — proximity (default) requires a cluster of 3+
verified entities within 300 chars.
Wires into ask() and query(). System prompts now require verbatim
quoted spans for every factual claim, restated via a user-turn
grounding_reminder one message before sources arrive (recent
user-turn instructions outweigh decayed system-turn rules under long
context in Hermes).
providence_cache gains 5 columns (audit_mode, n_quotes, n_verified,
unverified_quotes, verifier_method) with CHECK constraints. A
connect-time _migrate_audit_mode() ALTERs legacy DBs idempotently.
Cache hits return the persisted audit_mode rather than asserting
STRICT unconditionally.
CLI:
emergent list VISUAL/HYBRID records; --aggregate ranks
unverified quotes by frequency (corpus-growth signal).
reclassify re-run the verifier against live providence records;
cold-source records skipped; --dry-run reports
transitions without writing; each change writes one
'providence_reclassify' audit event.
