End-to-end gap-close from the Phase 1 extractor. Five interlocking
fixes; live-verified that `what is a CPU?` → "Central processing unit"
at #1, `what is a GPU?` → "Graphics processing unit" at #1
EVIDENCE-WARRANTED 1/1; Mount Kilimanjaro / Soviet Union queries
unchanged (no regression).
(a) `synonym_expand` over-cap path is now rank-and-truncate by
source-frequency (descending) instead of hard-skip. CPU has 13
legitimate homonym expansions across the corpus; the prior
MAX_NEIGHBORS_PER_TOKEN=8 cap contributed *zero* expansion → no
canonical-article surfacing. Now: keep the 8 dominant by per-(token,
target) source-root count via _load_neighbor_source_freq.
(b) `_search_titles` orders by FTS5 bm25 ASC instead of LENGTH(title)
ASC on the FTS5-MATCH path. The length-asc tie-break was correct for
the 2026-05-02 "Back to the Future" LIKE-substring case but
counter-productive on FTS5 (tokenized; no substring junk; length-asc
preferred "Unit" / "Unite" / "B unit" over "Graphics processing
unit"). LIKE fallback keeps length-asc since the substring issue
persists there.
(c) `accept_tokens` (feeds title-search, core-keyword, title-rerank)
uses the synonym-expanded set instead of qtokens-only. The Phase 1
expansion existed but was only used in the FTS5 OR-fallback; satellite
articles saturated the budget before the canonical article entered.
(d) ARCHITECTURAL: `synonym_expand_strict()` (new — high-trust
evidence-kind subset: manual + manual_legacy + acronym_parens,
**excludes** link_reciprocity) for use in the multiplicative
`_rerank_by_title_purity` and as the source for `accept_tokens`.
Reciprocal-wikilink edges express *topical adjacency*, not synonymy
(a `Dinosaurs` page reciprocally links to `Curious George Brigade` →
edge that should not amplify retrieval); a multiplicative ranker over
them blows up. Strict view preserves the acronym-parens surfacing
(those edges ARE the phrase=expansion identity) while keeping
link-reciprocity to additive retrieval-route boosts via the broad
synonym_expand (still wired to `or_synonym_pool` for the FTS5 OR
fallback).
(e) Extractor regex tightened `[A-Z]{2,6}` → `[A-Z]{3,6}` and purged
~21K 2-letter acronym edges from shards. 2-letter acronyms (AI/ML/OS/
US/UK/IT/PC/TV) homonym-collide too often with common 2-letter QUERY
tokens like `go`/`is`/`am` — without this, "why did the dinosaurs go
extinct?" pulled Curious George Brigade via GO-acronym edges. The
high-value acronyms (CPU/GPU/RAM/DNA/FBI/WHO/…) all clear 3 chars.
Also: CLAUDE.md gains a "prefer existing ticket; only split for
Dav1d-review audience" discipline note (saved as feedback memory) —
this work is itself an example: would have been #000055 + #000056 +
#000057 under the prior pattern; instead extends #000054.
Suite: 2531 passed (no regression). bench-qa in flight separately.
All 4 manifest candidates clean-separate (12/12 NEG catch at 0/14 POS FP),
so 'reranker discriminates Zionist-entity-style mis-cites' is a property
of MS-MARCO-trained rerankers as a class — not the specific L-6 I picked
first. Ranked by separation margin (per §7 #18: separation beats raw):
ms-marco-electra-base +5.349 ← new primary
BAAI/bge-reranker-base +3.076
ms-marco-MiniLM-L-6-v2 +3.053 (previous primary, demoted)
ms-marco-MiniLM-L-12-v2 +2.663 (worst — deeper ≠ better)
Manifest primary moved to electra-base for the cushion. demote_below_score
STAYS null — clean candidate-bench thresholds don't predict real-pipeline
behavior (the §7 #18→#27 history is 6 verdict flips on the NLI side);
§3.2.2 step 2 (real-traffic shadow sweep on pooled bench-qa STRICT) is
what sets it. Expect a walk-back. Reranker still doesn't catch the
Kilimanjaro/Mount-Kenya recombination (aboutness ≠ truth-of-attribution;
that's #000049 territory). fox's bench-maxing correction applied.
arborist/qa/relevance/ — manifest pins cross-encoder/ms-marco-MiniLM-L-6-v2
(~80MB, Apache) as primary; alternates: L-12, BAAI/bge-reranker-base,
ms-marco-electra-base. demote_below_score=null on purpose — the
#000049 §7 #18→#27 discipline (proved 6× that clean-eval thresholds
don't transfer to bench-qa data) requires the threshold to be set by a
shadow sweep against pooled real STRICT, not by a literature number.
ShadowRelevance class mirrors ShadowNLI (lazy [nli]-extra import, cuda
auto-detect via ARBORIST_RELEVANCE_DEVICE or ARBORIST_NLI_DEVICE,
batched _score_batch, graceful degrade-to-available=False). Two surface
methods: check_question_answer (deflection / Q-A drift) and
check_claim_source (topic-collision mis-cite). 13 tests.
Sanity on the motivating field case (Zionist entity): ON-topic +9.96
vs OFF-topic -9.04 → 18-pt margin. Mona Lisa Q→A deflection: on +10.45
vs deflect +3.56 → ~7-pt margin. The model CLEANLY discriminates the
failure modes #000052 §1 named. It does NOT catch the
recombination-where-the-different-entity-clause-also-mentions-the-target
case (Kilimanjaro/Mount Kenya) — and that's the right architectural
split: aboutness (#000052 §3.2) and entailment (#000049 NLI) are
orthogonal axes; the Kilimanjaro recombination case needs the semantic
candidate selector (#000050/#000051 vec hybrid).
Remaining: build candidate-bench eval (~20-30 deflection + mis-cite
fixtures), shadow-sweep θ over pooled bench-qa STRICT (expect another
walk-back per the #000049 lesson), recall-side realism check, then
fox+dav1d sign-off. Still SHADOW; production verifier unchanged.
test_concepts_extract.py grew 20 -> 28 with the acronym_parens
extractor's 8 tests; the AUTOCOUNT discipline (CLAUDE.md) fires the
doc-counts regression on stale numeric claims.
`arborist/concepts/extract.py:acronym_parens_synonym` — new
corpus-agnostic extractor. Scans each doc's lead chunk (first 4000
chars) for `<Multi-Word Phrase> (ACRO)` where the all-caps acronym's
letters strictly match the content-word initials of the phrase, in
order, after function-word filtering. Emits bidirectional synonym
edges between the lowercased acronym and each ≥3-char content token
of the phrase, evidence_kind="acronym_parens", anchored to that doc's
document_root. Idempotent like link_reciprocity_synonym.
Why this complements link_reciprocity: Wikipedia represents
abbreviation→expansion as a one-way *redirect* (CPU →
Central processing unit), which the ingest does not record as an
edge — so the existing reciprocal-link extractor never learned the
relation. The relation IS in body text by near-universal convention
("Central processing unit (CPU) is..."), which this extractor reads.
Corpus-agnostic: HTML, blogs, textbooks benefit equally.
Conservative: strict 1:1 acronym-to-atom match (rejects HTTP-shape,
where letters land mid-word), function words filtered, repeated
definitions deduped per doc, ≥3-char target floor. 8 new tests
covering CPU bidirectional emit, RAM idempotency, FBI function-word
filter, HTTP length-mismatch reject, XYZ initial-mismatch reject,
ROM hyphenated-word handling, per-doc dedupe, registry presence.
Retrieval-side only — synonym edges reshape FTS5 candidate selection
via synonym_expand at query time, never enter audit_mode / cache_key
/ audit_event_hash. No governance hash bump, no cache invalidation.
Closes#000050 §2a's CPU/GPU abbreviation rows *upstream* of vec;
the Orwell-shape conceptual-allusion row remains the genuine #000050
justification. Operational follow-up (not code): run on each shard
via `arborist concepts derive --extractor acronym_parens` (CLI
surface itself is aspirational in docstrings; extractors are called
programmatically today). Next ID 000054 -> 000055.
Built bench/fixtures/5f/recombination-realcontext-v1.jsonl: 25 hand-built
recombination claims (deliberate re-attributions within real bench-qa
30 KB Wikipedia haystacks the pipeline produced — Mount Kenya into a
Kilimanjaro answer, Galileo onto all 63 Jupiter moons, Joshua Reynolds
onto the Mona Lisa, etc.). Ran the full 7-model × full-grid sweep vs
the 808 pooled STRICT. Result:
- bart-large-mnli / k=12 / max / θc=0.999 → 12/25 = 0.48 catch · 0/808 FP
(the real-haystack fp=0 leader)
- deberta-large-mnli / k=3 / margin / θc=0.995 → 6/25 = 0.24 (§7 #26's
'settled' config — 28/28 synthetic, 0.24 real-haystack: 4× over-estimate)
- roberta-large 0.12, MiniLM 0.08, deberta-base 0.04
So §7 #26's 'boundary closed' walks back to 'boundary PARTIALLY closed'
on real haystacks. The bottleneck is architectural: top-k by token
overlap misses the contradicting clause when it shares few subject-area
tokens with the answer (e.g. the Mount Kenya clause only shares 'Kenya'
with a Kilimanjaro claim — ranked low, NLI never sees it). Threshold
tuning doesn't lift the ceiling; a SEMANTIC candidate selector
(vec-driven, sibling of #000050/#000051's hybrid retrieval) does.
bart's pareto above fp=0: fp=0.011 catch=0.52, fp=0.057 catch=0.84,
fp=0.068 catch=0.92 — permissive operating points are on the menu if
fox+dav1d sign off. recommended_operating_point updated to
bart-large-mnli/k=12/max/θc=0.999; deberta-large/margin kept as the
synthetic-eval reference. Sixth meta-lesson instance: clean synthetic
eval doesn't predict bench-qa precision OR recall — neither contrived
dataset axis is load-bearing, only the real pipeline shape is.
Production verifier unchanged; falsification-hard stays 10/12. Still
SHADOW; runtime promotion fox+dav1d-decides.
(a) Mined the pooled n=1+3+5 bench-qa runs (808 distinct STRICT answers)
for natural recombinations at lowered θc≥0.7 → 37 would-fires, ALL
token-collision FPs on inspection (Mount Kenya pulled into a Kilimanjaro
answer, Dalí into da Vinci, Donovan into Superman). ZERO genuine
recombination errors — the boundary is theoretical-in-practice; the
failure mode is the candidate selector (top-k by token overlap) pulling
different-entity same-subject-area clauses.
(b) Re-ran the grid against the 808-cell pooled STRICT set: the §7 #25
'max@0.96' was itself a small-sample artifact — the n=5 444-cell set
lacked the high-confidence spurious hits the pooled set has. On 808
cells θc goes back to ~0.995, and at θc=0.995 only agg=margin still
catches 28/28 (max gets 27/28). microsoft/deberta-large-mnli / k=3 /
agg=margin / θc=0.995 → 28/28 synthetic recombinations · 0/808 pooled
real STRICT FP · 0/26 synthetic legit — the ONLY config in the
7-model×full-grid sweep that hits 1.0/0.0 on 808 cells, held at n=3
too. recommended_operating_point reverted to margin@0.995.
Realistic next check: ~20-30 hand-built synthetic-recombination-vs-
real-bench-qa-context fixtures (real haystack, deliberate re-attribution).
Still SHADOW; runtime promotion fox+dav1d-decides. Production verifier
unchanged; falsification-hard stays 10/12.
Meta-lesson instance five: a bigger sample can vindicate a config a
smaller one made look unnecessary — re-confirm the config choice (not
just the threshold) each time the denominator grows.
Enumerates the concrete query-words-share-zero-tokens-with-target-title
cases the §2 gate's "semantic-allusion fixtures" must include, as a
running list: Orwell→Eastasia (genuine conceptual allusion — the case
that justifies the vec layer), "what is a CPU?"→Central processing
unit, "what is a GPU?"→Graphics processing unit (abbreviation→expansion
subclass — also fixable upstream by a concepts/ synonym edge; bench
records which fix closes each row). Field cases 2026-05-13, fox.
#000053 fixed the verifier's separate acronym blind spot but not this
retrieval gap.
ARBORIST_NLI_SHADOW=1 make bench-qa BENCH_QA_N=5 → 1125 cells, 444 real
STRICT (5x n=1, 1.6x n=3). Re-ran the 7-model mega-grid: the lexical-
candidate NLI veto robustly clears the §7 #12 gate with
microsoft/deberta-large-mnli / k=2 / agg=max / guard=max_entail /
θc≈0.96 / θe=0.9 → 28/28 synthetic recombinations (incl. both fixtures)
· 0/444 real STRICT FP · 0/26 synthetic legit FP, ~4 pts θc headroom;
roberta-large-mnli equally good (k=2/max/θc=0.95). Resolved: agg=max +
max_entail guard is the robust score-shape (§7 #24's margin win was a
sample tie); the large checkpoints (~350-400M deberta-large-mnli /
roberta-large-mnli) hit 1.0/0.0, bart-large (similar size) only ~0.71,
deberta-base at the cliff (27/28 here, 11/28 at n=3), small models
(MiniLM-82M, deberta-v3-small) cap at ~0.82 — so §7 #18's 'MiniLM is
the cost-pick' is OVERTURNED by the proper-n evidence; k=2 consistent
winner; int8-ONNX costs ~1 catch. recommended_operating_point updated.
Remaining: a bench-qa-derived recombination set (the one check not
done); n=9 if dav1d wants more; still SHADOW; runtime promotion is
fox+dav1d-decides. Production verifier unchanged; falsification-hard
stays 10/12. This run is the worked example behind CLAUDE.md's new
bench-maxing line.
`arborist.qa.evidence._content_tokens` dropped every token under 4
chars, so a short all-caps acronym (CPU, GPU, DNA, FBI, USB…) never
registered as a content token — which defeated Rule 8
(_claim_title_overlap / TITLE_MISMATCH), the subject-tokens-absent
check (Rule 9), the bare-name-claim guard, and spotlight-excerpt token
selection whenever a question/claim's topic IS an acronym. The field
case: `what is a CPU?` cited to the "CPU design" article tripped
TITLE_MISMATCH even though claim and title both contain "CPU".
Fix: keep a token if it's an all-caps 2-3-char alpha run in the source
text; everything else unchanged. The change only ever ADDS tokens, so
TITLE_MISMATCH / SUBJECT_TOKENS_ABSENT / BARE_NAME_CLAIM can only stop
firing, never start — monotone toward fewer spurious demotes; no
STRICT→non-STRICT transition is possible from it.
Versioned: `content_token_rules: "v2-acronym-aware"` added to
runner.DEFAULT_POLICY + query.DEFAULT_QUERY_POLICY +
keys._VERIFIER_POLICY_FIELDS → folds into verifier_policy_hash, prior
cache records orphan on lookup (by design; same discipline as
base_version / hyphen_fold_v1). Does NOT touch the retrieval
abbreviation→expansion gap (CPU→Central processing unit — #000050 vec
hybrid / concepts/ synonym edges; the root cause of the satellite
retrieval). 8 new tests; full suite green (2502); bench-qa-smoke clean.
Next ID 000053 -> 000054.
ARBORIST_NLI_SHADOW=1 make bench-qa BENCH_QA_N=3 → 275 real STRICT cells
(3x the n=1 sample). Re-ran the expanded grid (7 aggregations incl.
margin = max_clause(p_contra - p_entail), paired-entail guard variant,
θc to 0.999, --extra-models) over all manifest models + 4 extra
xsmall→large (microsoft/deberta-large-mnli, roberta-large-mnli,
nli-deberta-v3-small, deberta-v3-xsmall), synth-28 recombination vs the
275 STRICT cells. Result: the §7 #23 deberta-base/k=2/θc=0.99 config
does NOT survive — it catches only 11/28 at θc=0.995 (which the larger
STRICT sample forces). BUT the broader sweep found the config that does:
microsoft/deberta-large-mnli / k=3 / agg=margin / θc=0.995 → 28/28
synthetic recombinations (incl. both fixtures) + 8/12 falsification-hard,
0/275 real STRICT FP, 0/26 synthetic legit FP — a passing config at
proper n. Findings: margin is the right score-shape (single threshold,
folds the guard in); the specific checkpoint matters more than param
count (deberta-large-mnli wins clean, deberta-base collapses,
roberta/bart ~0.71-0.75 — no 'bigger is better' law). recommended_operating_point
updated. Still SHADOW; runtime promotion needs a bigger STRICT sample +
a bigger recombination set + fox/dav1d sign-off. Production verifier
unchanged; falsification-hard stays 10/12.
Meta-lesson sharpened twice: clean eval ≠ bench-qa precision (§7 #18→#20),
default config ≠ best config (§7 #22→#23), small FP sample ≠ large FP
rate (§7 #23→#24).
The §7 #22 'fails the gate' was the verdict for the DEFAULT config
(k=6/θc=0.5/θe=0.9, tuned on the clean synthetic set), not the
approach. A {model × candidate-cap k × aggregation × θc × θe} grid
sweep (bench/scripts/nli_shadow_grid.py — NLI runs once per
(model,record) over the top-12 candidate clauses, the k/agg/θ grid is
then arithmetic on cached scores; ~10s on the 4090 for 4 models) finds
clean passing configs: on the 89 real STRICT cells (n=1 bench-qa),
deberta-base-184M / k=2 / agg=max / θc=0.99 / θe=0.9 → 27/28 synthetic
recombinations caught (incl. both 5f-fal-hard fixtures), 0/89 STRICT
FP. MiniLM-82M passes too (24/28 · 0/89). Model science: 184M > 82M >
407M for fp=0 recombination recall; int8-ONNX costs ~1 catch vs fp32.
Caveats: FP side is n=1 (BENCH_QA_N=3 run in flight); recall is on the
synthetic set; flipping to a runtime demotion-only veto is fox-decides
(then nli_policy_hash folds into governance_policy_hash per §7 #2).
Manifest active defaults stay k=6/θc=0.5; recommended_operating_point
(deberta-base, k=2, agg=max, θc=0.99, θe=0.9) documented in the
manifest. 3 grid result JSONs committed. Standing lesson: neither the
clean synthetic eval NOR the default config predicts bench-qa precision
— you have to sweep. Production verifier unchanged; falsification-hard
stays 10/12.
bootstrap-nli-only: a venv + [nli] only — no [dev] extras (no crawler/
hessian/vec/sympy). On a CUDA host PyPI's torch wheel is the CUDA build,
so ShadowNLI auto-detects cuda and bench-nli-shadow / export-nli-onnx /
the candidate bench all run on the GPU with no further wiring. Verified
on the ai box (RTX 4090): torch 2.11+cu130, cuda True; 82M cross-encoder
batched ≈ 0.09 ms/pair (512 pairs in 0.047s — ~350x onnx-int8-cpu,
~1300x torch-cpu-batch1); 62-record synthetic shadow sweep in 2.4s wall;
24/24 tests pass. The GPU only accelerates the NLI half — make bench-qa
(Hermes LLM) still runs wherever the shard corpus is.
Speedup (§3 plan): ShadowNLI._nli_batch batches forwards
(ARBORIST_NLI_BATCH=64); device auto-detect (ARBORIST_NLI_DEVICE, else
cuda-if-available); auto-prefer an ONNX export — bench/scripts/export_nli_onnx.py
/ make export-nli-onnx exports + int8-dynamic-quantizes the pinned
checkpoint into ~/.arborist/models/nli/<ver>/onnx/ (operator state, NOT
committed), _ensure_loaded loads model_quantized.onnx via
optimum.onnxruntime (backend onnx-int8), falls back to torch silently.
torch-cpu-batch1 ~120ms/pair → onnx-int8-cpu-batched ~32ms/pair (~4x);
seconds on a 4090. optimum[onnxruntime] added to the [nli] extra; 24
tests.
Gate-item-4 verdict at proper n: ARBORIST_NLI_SHADOW=1 make bench-qa
BENCH_QA_N=1 → 223 cells (89 STRICT / 90 HYBRID / 44 UNGROUNDED; also
surfaced + fixed a lone-surrogate bug). Shadow sweep over those: NLI-as-
runtime-veto on STRICT has ~26% FP at θc 0.5, ~8% at θc 0.90, ~0% only
at θc 0.99 — and θc 0.99 gives up most recombination recall (hard
synthetic recombinations bottom out ~0.76). FAILS the §7 #12 gate on
this design. Only untried path that might pass: a Phase-3 runtime hook
running NLI on the verifier's actual matched clauses (1-3), not
top-6-by-overlap. Until then: runtime NLI demotion stays off; the 2
fixtures stay permanent boundary markers; θc stays 0.5. Production
verifier unchanged; falsification-hard stays 10/12.
Per-sentence shape check (no model) emitting kind ∈
{phrase_component_reuse, circular, vacuous, ok, empty}:
- circular: subject content-tokens ⊆ predicate's and the predicate
leads with a subject token ("Water is water").
- phrase_component_reuse: subject quotes a phrase, predicate reuses
one of that phrase's own tokens as a bare "the/a/an <token>"
referent — the 2026-05-12 field case ("the phrase 'Zionist entity'
is used as the entity"), a token collision the verifier +
deflection + title-relevance all pass and NLI returns neutral on.
Copulas inside a quoted span are skipped so 'war is peace' doesn't
break the subject/predicate split.
- vacuous: predicate is only placeholder hypernyms + filler ("X is
a thing").
Conservative — no full token-salad parsing; legit definitions pass ok.
Surfaced in inspect_cache_key + the `arborist inspect` human view
(· incoherent: <kind>). Advisory only — never writes providence_cache
/ audit_events / run_dag_root; demote-only verifier hook deliberately
not wired. 9 tests; full suite green (2500 passed).
Doc-only scaffold. Two more read-only/demote-only/never-in-proof-path
sidecars joining the diagnose_deflection family: (1) diagnose_coherence
— word-salad/circular/vacuous answers; lexical, no model; the near-term
win. (2) diagnose_relevance — semantic 'aboutness' (does the answer
address the question / is each claim about its cited source?); today's
checks are lexical and a token collision defeats them; a small
aboutness/reranker model (NOT NLI) under #000049 §7's discipline cage
verbatim; gated on evidence, travels with #000049's model question.
Motivating field case: the 'Zionist entity' claim_lattice query —
incoherent token-collision recombination NLI can't catch (returns
neutral) and both lexical relevance checks waved through. Flags an
upstream retrieval (polysemy/title-soup) root-cause ticket, not scoped
here. Next ID 000052 → 000053. #000049 sibling
candidate_clauses() — NLI now runs only on the top-N source clauses by
content-token overlap with the answer claim (max_candidate_clauses=6),
not the whole context; records n_candidate_clauses / best_clause_overlap
/ recombination_risk. Synthetic sweep unchanged (28/28 recombination,
0/26 legit FP, mean 1.45 candidate clauses/record). Real-traffic smoke
re-run: STRICT would-demote 30% → 20%, overall 47% → 33% — better, not
fixed; recombination-risk split doesn't separate either. Residual STRICT
false-contras at ~0.83-0.92 → θc would need ≈ 0.90 (vs the clean-set
0.5); at θc=0.90 the data in hand gives 27/28 synthetic recall, 0/26
legit FP, 0/10 smoke STRICT FP — but n=10 is too small to set on.
Next: a fuller ARBORIST_NLI_SHADOW=1 bench-qa run → sweep θc on hundreds
of STRICT cells → confirm → set it. θc stays 0.5; runtime NLI demotion
stays off. Production verifier unchanged; falsification-hard stays 10/12.
Live hook: ARBORIST_NLI_SHADOW=1 makes query() surface the verifier-input
text (gated off-by-default, never a cache_key/governance/audit_mode
input); qa_sweep.py carries it + the answer into bench rows; the shadow
sweep reads them and buckets by audit_mode. ARBORIST_NLI_SHADOW=1 make
bench-qa-smoke (15 cells) → the naive 'NLI on every context clause'
scaffold has a ~30% would-demote rate on STRICT answers — a haystack /
multiple-comparisons artifact (real contexts segment into 100-336
clauses; max-over-all almost always finds a tangential clause the model
reads as contradicting; a paraphrased STRICT answer often isn't verbatim-
entailed by any single clause so the entailment guard doesn't rescue it).
Lesson: the §7 #5 'candidate source clauses' + recombination-risk gating
is load-bearing, not optional. Do NOT enable runtime NLI demotion on the
current scaffold; next step is the candidate-clause restriction, then
re-run, then gate item 4 is meaningful. Production verifier unchanged;
falsification-hard stays 10/12.
arborist/qa/nli/ — SHADOW ONLY (never an audit_mode input; manifest not
yet in governance_policy_hash per §7 #2). manifest.json pins
cross-encoder/nli-MiniLM2-L6-H768 @ a fixed HF revision + the
bench-validated θc 0.5/θe 0.9 + 2 alternates + the Phase-3 TODO;
shadow.py = ShadowNLI/shadow_check (lazy transformers+torch behind a new
[nli] extra, clauses() segmenter, the §7 #5 clause-level Demote()
decision, degrades to available=False when [nli] absent);
bench/scripts/nli_shadow_sweep.py + make bootstrap-nli / bench-nli-shadow
(the gate-item-4 instrument); 16 tests.
First sweep (116 records — 5f-falsification packs + the arborist-nli-bench
eval sets): 28/28 synth recombination demoted, 0/26 FP on legit summaries,
0/9 fires on already-STRICT_SPAN records, 25/50 on UNGROUNDED (the
contradiction half; quiet on non-sequiturs). Gate items 1/2/3/5/6 clear
on available data; item 4 — shadow FP rate on a real live-bench-qa
sample — remains the open measurement. Production verifier unchanged;
falsification-hard stays 10/12.
fox asked: JIT-trained, gossip-pulled, or repo-embedded?
Added §2.4 — three shapes, two dead ends:
- JIT-trained per deployment: ❌ — GPU-hours over MNLI/SNLI/ANLI, not
a per-deployment step; non-deterministic across runs → different
audit_mode → breaks the verifier's reproducibility invariant.
"Training" is a one-time act by someone; the distributed thing is
the resulting fixed blob, not the recipe.
- Embedded in the git repo as data: ❌ — a few-hundred-MB binary in
.git blows the standing "fresh clone = python3.12 + venv + sqlite3"
invariant; Git LFS adds a dep and still bloats; the repo is source
+ tiny fixtures + docs, not a model registry.
- One-time obtained, content-addressed, fetched on demand, mesh-
distributable: ✓ — the repo carries a tiny manifest
(arborist/qa/nli/manifest.json: {checkpoint_sha256, source_url,
license, nli_model_version}); the weights live under
~/.arborist/models/nli/<hash>/, fetched on first use (make
fetch-nli or auto-fetch on first --enable-nli query) with the
sha256 verified fail-closed against the manifest; nli_model_version
folds into governance_policy_hash (a swap stales the cache, like
chunking_version / canonicalization_version), keeping audit_mode a
deterministic function of (answer, source, policy, pinned hash);
and once arborist's mesh grows blob-sync, a deployment with peers
pulls the checkpoint from a peer (AXFR-style, like ClouDNS slaves
pulling a zone, like a shard rehydrating from snapshot.db) instead
of the origin URL — content-addressing makes peer-pull and
origin-pull interchangeable, the mesh is the optimization not the
canonical source, cold start falls back to origin. Source: a
published off-the-shelf NLI checkpoint (MIT/Apache, recorded) or a
one-time fox/blackops-trained-and-published one.
Bottom line: not JIT-trained, not repo-embedded weights — a
content-addressed hash-pinned checkpoint, manifest in the repo,
weights under ~/.arborist/, fetched on demand (origin or peer), hash
in governance_policy_hash. Precedents cited: the textbook manifest
(pointers + licenses not texts), the [vec] extra (fastembed's
bge-small fetched not committed — #000039), ~/.arborist/ operator-
state, docs/mesh.md + snapshot rehydrate. All presupposes the §2.2/§5
discipline call came back "yes" (is a fixed NLI model allowed in the
audit_mode path?); if "no", none of it builds. Also wired §2.2 con
(b)/(c), §3, §5, §6 to point at §2.4. Doc-only.
bge-small-en-v1.5 batched on a 4090 ≈ 10^3-10^4 chunks/s → full
6.24M-chunk corpus in minutes. Drop a CUDA Embedder (fastembed
CUDAExecutionProvider, or sentence-transformers device=cuda) into
default_embedder()'s pluggable callable; int8 quant stays CPU-side
post-embed; CUDA stack lives only on the producer box. Determinism
note: GPU backfill isn't byte-identical to CPU — irrelevant, soft
data, pin model revision + recipe not output bytes.
#000051: gossip the embedding backfill over mesh — backfill once on any
CPU box, publish a vecpack (leaf_hash-keyed, soft data, cheap structural
gate, never proof path), peers pull + bulk-load. The laptop never runs
the transformer; this is the mechanism behind whitepaper §1's "the
embedding pass runs off the device". Supplies #000050's backfill prereq
as a distributable artifact.
#000050: fold the Dav1dPrometheus-review high-value items into Phase-2
scope — accept-path-5 (vec hits clear the title gate via span-level
warrant, not similarity score, so the gate doesn't drop the semantic
candidates vec exists for), six vec config fields into
governance_policy_hash + a cache-write guard until wired, run-DAG
records the vec stage, four-condition bench (A/B/C/D, C-beats-D) on
semantic-allusion + curated + adversarial-semantic-neighbor fixtures.
Backfill prereq now routes through #000051.
Next ID 000051 -> 000052.
The #000012 top-of-file Status line still said "Phase 1c (branch-set
persistence) remains proposed-not-opened" — but Phase 1c landed
2026-05-10 (d53115e: fork_score_branches table + persist_branch_score
/ branch_set_density + the default-off CLI flags), the #000037 §12
Trigger 1 probe was wired to it 2026-05-11 (dbe8249), and the ticket's
own §7 Phase 1c section + the TICKETS.md index row already say
"landed". Only the header was stale.
Refreshed: Phase 1c landed (with the probe wire — measurable, not yet
fired); §8 = the ForkScore threshold-calibration handoff from #000025
§10.14 (delta_aggregator knob #000047; harder-fixture-tier line
#000046 closed / #000048 closed / #000049 open); Phase 2 (the v8
multi-validator consensus implementation) noted as gating on a
multi-validator deployment, like #000045. Doc-only.
make bench-emergent EMERGENT_N=30 — 30 fresh 3-word-triangulation
cycles appended to bench/emergent_log.jsonl.
Aggregate: 29/30 UNGROUNDED, 1/30 HYBRID, 0 STRICT — the expected
healthy null result. The triangulator generates absurd questions
("How does the depth of one's music-making abilities on a kazoo …")
with no real corpus support; the verifier ladder honestly returns
UNGROUNDED rather than hallucinating-then-STRICTing. 0 false-positive
STRICT on garbage input.
The one HYBRID (perversely/succeeds/disagreeably) is a re-run of
failure shape E ("total deflection to unrelated trivia"): the common
words pulled Hirabah / Original-righteousness chunks via FTS5; the
model quoted them verbatim (3/4 claims verify) but the answer is
total topic-drift → DEFLECTION_DETECTED + CITATION_MISMATCH → capped
at HYBRID, not STRICT. The deflection sidecar + verbatim verifier
holding on adversarial-by-construction input — nothing new, logged as
a confirmation.
No new failure shape or correctness concern. #000006 amend records
it; the 30 entries join the --print-pending queue if a future
teacher-review pass wants a closer look. Doc + log only.
Capture the positioning fox articulated: arborist's per-document ingest
is ~10-100x cheaper than building a vector-DB representation — same
SQLite substrate, different retrieval philosophy — which is the
difference between "ingest + search runs on a phone" and "the NPU is
now a hand-warmer."
New docs/lexical-first-rationale.md (positioning/architecture
reference, not a ticket): the cost asymmetry with the measured numbers
(FTS5 + SHA-256 leaf + Merkle commit + sqlite + zstd pipeline << 1 ms/
chunk vs bge-small ONNX inference ~5-30 ms/chunk, worse contended; the
query side too — a vec query embeds the query string first, an FTS5
query is B-tree lookups); the same-SQLite-different-philosophy table
(inverted index vs dense vectors + ANN; build cost; query cost;
matching; proof-bearing); the deep version of the point — arborist IS
the Merkle Providence model and that model is cheap by construction,
embeddings are a soft signal (CLAUDE.md "soft hash vs hard hash") that
never enter a proof and are the expensive bolt-on; the edge/mobile
consequence; the honest caveat (lexical-first trades the semantic
allusion gap — which is why vec is opt-in/additive, never the default,
and the embed pass is lazy/out-of-band so the heavy transformer work
runs off-device/off-peak; int8 keeps the storage tax at +6%).
Wired in: TICKETS.md "Distinction from other docs" reference list gains
the doc; #000039 §14.6 gains a "Strategic framing" pointer to it.
(A possible follow-up: fold the mobile-viability argument into the
Merkle Providence Reverse RAG whitepaper proper — noted in the doc's
references; not done here, that's a deliberate cross-repo paper edit.)
Doc-only.
fox 2026-05-12: the attribution-aware path (#000048's deferred §2.3 —
closing the 2 recombination over-grounds in falsification-hard) is its
own ticket, not a #000048 phase. So:
#000048 → closed (at 2.1 + 2.4). Steps 2.1 + 2.4 landed 2026-05-11
(12 of 16 residual items: 4 HYBRID_ENTITY over-grounds + 8 Formulate
mis-segments → formulate-hard 12/12, falsification-hard 10/12; each
bench-gated, no STRICT-rate regression). Step 2.2 (single-clause-
containment paraphrase check) attempted + reverted — recombination-
vs-summary isn't lexical (§5 "What we learned"). The 2 residual
falsification-hard fixtures (hard-003 Mercury, hard-005 Einstein)
stand as a documented marker of where the lexical verifier stops.
Header + §5 Closure + §2.3 updated; cross-refs in #000046 / #000012
§8 / TICKETS.md repointed from "#000048 §2.3" to "#000049".
#000049 opened (doc-only, awaiting go/no-go) — "Attribution-aware
grounding check (the recombination boundary)". The recombination
class needs an attribution / dependency-parse or mini-NLI check
(distinguishing "Mercury is the largest" against "Jupiter is the
largest; Mercury is the smallest" from a legit cross-sentence
summary). Options: 2.1 hand-rolled dependency-attribution heuristic
(no model, brittle — same threshold-can't-separate problem one rung
up); 2.2 small purpose-built NLI model ([nli] extra, policy-gated,
off-by-default, bench-gated — the right capability, but forces the
"is a fixed NLI model an LLM-judge?" discipline call + a model
dependency + a non-determinism surface to pin); 2.3 do nothing (the
2 fixtures are a boundary marker, no observed real-traffic harm).
Recommends 2.3 until real-traffic recombination over-grounds show up,
then 2.2 *if* fox rules a fixed NLI model is acceptable in the
proof-adjacent path; the first decision the ticket needs is that
discipline question. Next ID 000049 → 000050. #000048 follow-up.
Doc-only — no code change.
New §14.6 captures the 2026-05-11 findings:
- Measured embed rate: ~4.3 chunks/s on the contended dev box (load
~11 on 8 cores; the embed process got ~28% of one core). One wiki
shard (~1.56M chunks) at that rate ≈ 100h ≈ 4+ days; all four ≈
~16 days. The full int8 backfill was abandoned as not feasible to
brute-force there.
- The 54K-chunk partial on 000.db confirmed ~409 B/chunk apparent
→ 384 B amortized → ~2.4 GB for the full 6.24M-chunk corpus at
int8 — the deterministic number a full backfill would only
re-confirm, so finishing it bought nothing.
- On an idle healthy box (batching + all cores) bge-small does
~50-200 chunks/s → full corpus ≈ ~9-35h (the ticket's earlier
"~17h" is the optimistic end).
- Per-chunk cost breakdown: bge-small ONNX inference ~5-30 ms/chunk
dominates; the existing arborist ingest steps (chunker + SHA-256
leaf + Merkle commit + sqlite INSERTs + zstd + FTS5) are well
under 1 ms total. So adding vec multiplies ingest by ~10-100×,
entirely in the ONNX matmuls — the non-vec ingest path is
unchanged and still runs at hundreds of chunks/s.
- Implication for §14.2: this is *why* lazy-out-of-band is the
default and `arborist ingest --embed` is the opt-in. Production
guidance: a corpus-wide backfill is a one-time batch job (hours
on idle / days on contended), best run off-peak or on a dedicated
box; it does not slow ongoing ingest (which never embeds unless
--embed is passed); a GPU/accelerated embedder is a drop-in via
the pluggable Embedder callable if backfill latency matters.
Status line updated to point at §14.6. Doc-only.
What we learned from the (reverted) step-2.2 attempt, stated as a
general principle in #000048 §5 "What we learned":
A recombination ("Mercury is the largest planet …" reusing the
source's "largest planet …" with its "Mercury") and a legitimate
cross-sentence summary ("Batman, who is the alias of Bruce Wayne,
lives in Gotham City." reusing two adjacent source sentences) are
the same SHAPE to any lexical signal — both scatter the answer's
content tokens across source clauses, and the recombination's
best-single-clause coverage (4/5 = 0.8) sits ABOVE the legit
summary's (4/6 = 0.67), so no token-coverage / clause-containment /
bigram threshold separates them in the safe direction. The
discriminating thing is *attribution* — in the source, are these
tokens attached to the same subject/predicate the answer attaches
them to? — which is a dependency / NLI question, not a string
metric. That's the boundary of the deterministic, no-LLM-judge
lexical verifier: absence signals (#000046 numeric gate, #000048
step 2.1 entity gate) and structure-of-the-model's-own-output
signals (step 2.4 segmenter) are lexical and work; "the source
contradicts this pairing" is not, and proxying it with a coverage
cut trades a small contrived-fixture win for honest demotions of
real summaries — a net loss against bench-maxing's 5-pp floor.
Updated: #000048 §2.2 (the attempted idea kept as design log + the
no-threshold-separates finding), §2.3 (now framed as the only path to
the last 2 — attribution / mini-NLI, its own ticket if ever), §3
(original plan annotated with the LANDED / ATTEMPTED+REVERTED / NOT
DONE outcome), §5 (the step-2.2 receipt + the "What we learned"
subsection + the closure recommendation). Stale cross-refs fixed:
TICKETS.md #000048 + #000046 rows, #000046 ticket Headroom section,
#000012 §8 #4 — all of which said "#000048 step 2.2 closes the last
2", now corrected to "step 2.2 attempted + reverted; the 2 recombination
fixtures stand as documented residue; #000048 §2.3 is the path if
ever wanted".
Recommendation unchanged: close#000048 at 2.1+2.4 (12 of 16 residual
items closed — formulate-hard 12/12, falsification-hard 10/12). Doc-
only — no code change.
Tried a single-clause-containment check on the paraphrase path (and
the entity path's weakest slot): a span is grounded only if a single
source clause (;/. -delimited) covers >= paraphrase_coverage of its
content tokens — a recombination spreads them across clauses. On the
contrived fixtures it works (hard-003 Mercury, hard-005 Einstein →
UNGROUNDED → falsification-hard 10/12 → 12/12). But it ALSO rejects
legitimate cross-sentence summaries — the established fox-2026-04-29
Batman case ("Batman, who is the alias of Bruce Wayne, lives in
Gotham City." against "Batman is the alias of Bruce Wayne. Batman
lives in Gotham City.": spans two clauses, best clause covers 4/6 =
0.67 → demoted, but it's a correct paraphrase). And there's no
threshold that separates the two directions — the recombination
("Mercury …": best clause covers 4/5 = 0.8) sits ABOVE the legit
summary's 0.67, so any cut that rejects Mercury also rejects Batman.
Distinguishing "recombined into a different statement" from
"summarized two adjacent sentences" needs attribution / dependency
parsing or a mini-NLI model (the §2.3 territory), not a conservative
lexical proxy.
Reverted — verify.py stays at its post-step-2.4 state; falsification-
hard stays at 10/12 with hard-003 + hard-005 as documented residue.
#000048 §5 records the attempt + the no-threshold-separates argument;
recommendation: close#000048 at 2.1+2.4 (12 of 16 residual items
closed), with the 2 recombination fixtures a marker for a future
attribution-aware verifier (its own ticket). Awaiting fox's call.
Doc-only — no code change in this commit.
#000035 closed 2026-05-11 (commit 4b85eb1: v7 substrate spec into the
repo at docs/_source/merkle-agi-dag-v7.rst, § 9.10 folded in). Update
soft-hash-channel-analysis.md §9 open-question 2 from "Phase 2 lands
when v7 gains a deployment target" to resolved: φ_PRG = HMAC-SHA-512
with a 32-byte committed seed, little-endian.
Closes the 8 mis-segments #000046 left in formulate-hard-v1.jsonl.
The parser was line/bullet-only — one line ⇒ one claim — so a line
that crammed several pointered claims onto one row ("Water is wet
[E1]; fire is hot [E2]", "X happened [E1]. Y followed [E2]") became
one monolithic claim with all the pointers, and a wrapped bullet
became two.
arborist/qa/parse_claims.py: _SEGMENT_SEP_RE splits a line on ';',
sentence boundaries ('. '/'! '/'? ' then a Capital), spaced dashes
(' - '/' — '/' – '), ' and '/' or '/' because '/' although '/' since
'/' while ', inline '(N)' enumeration markers, and commas — with
'(?![^\[]*\])' so a comma inside a [E1, E2] bracket never splits it.
_segment_line keeps the split ONLY IF every resulting non-empty
segment is a well-pointered claim — a legit single claim ("The cat
is black and white [E1].", "The cast: A, B, C [E1].") is never
broken because splitting it would manufacture pointer-less prose
fragments → guard rejects; a leading colon-terminated header with no
pointer ("Two facts:", "Key points:") is allowed and dropped. Plus a
wrapped-bullet join: a continuation line (leading whitespace then a
lowercase letter, no bullet glyph) folds its text + pointers into the
previous claim.
Effect: formulate-hard rate 4/12 → 12/12 (the pack is now at ceiling
— a harder Formulate tier would re-open below-ceiling headroom; a
#000046 follow-up). Remaining #000048 headroom: 2 STRICT_PARAPHRASE
recombinations in falsification-hard (Mercury, Einstein — step 2.2).
Bench gate: make bench-qa (n=3 × 75 × 3 = 675 cells; parse_pointer_claims
feeds the 450 claim_lattice_pointer + claim_lattice cells) after
(bench/qa_results/2026-05-11T20-26-37Z) vs the pre-step-2.4 baseline
(...T17-12-41Z = HEAD's parse_claims.py). STRICT-rate quote 0.54→0.55,
pointer 0.22→0.22, lattice 0.43→0.45 — all within the 5-pp noise
floor. Per-row diff: the segmenter changed the parsed-claim count on
the SAME answer text for 7 of the 450 lattice cells (0 in
claim_lattice, 7 in claim_lattice_pointer); of those, 2 caused an
audit_mode change — both correct: a wrap-join recovered an answer's
intended structure (4 claims, 2 pointer-less wrap-fragments → HYBRID)
into 2 well-pointered claims → STRICT; and a crammed-one-line blob (1
monolithic claim, all pointers → STRICT) split into 8 claims, some
not individually verifying → HYBRID (the honest verdict — false-
positive STRICT was the corruption). Every other lattice/quote delta
is LLM re-answer variance. No regression — the segmenter's only
visible effects on real traffic are honest improvements. Summarized
in qa-modes-bench.md Addendum 7 + ticket-000048 §5 step 2.4.
Tests: 8 new in test_claim_lattice.py (semicolon/sentence/conjunction
splits; pointerless-fragment + cast-list guards; leading-colon-header
drop; wrapped-bullet join; pointer-order/multi-pointer); existing
parse_pointer_claims tests pass untouched; test_5f_formulate_hard_pack
re-pinned 4/12 → 12/12. make test 2358 passed, 28 skipped.
#000048 → steps 2.1 + 2.4 landed; #000046 / #000012 §8 / TICKETS.md /
Makefile / fixture _meta + notes updated.
Brought the Merkle-AGI v7 formal substrate spec into the repo as
docs/_source/merkle-agi-dag-v7.rst (previously referenced only as the
un-version-controlled ~/Downloads/merkle-agi-dag_v7.txt). Section
structure converted to reStructuredText; inline math kept in the
source's literal notation; added to the docs/_source/index.rst
"Substrate" toctree (also added the pre-existing merkle-agi-v8-consensus
entry that was missing from it).
Folded ticket #000035's § 9.10 + § 9.10.1 (anchor PRG map φ_PRG;
dav1d-reviewed-final, little-endian, HMAC-SHA-512 / 32-byte seed) in at
their numbered positions, after § 9.9, with a .. note:: citing the
reference implementation (arborist/substrate/anchor_prg.py). #000035 ->
closed (Phase 1 + Phase 2 both landed); #000018 §9.2 ("which PRG?")
resolves to HMAC-SHA-512 with a 32-byte committed seed. Full upstream
v7 spec revision stays exogenous; this lands the amendment into the
tracked in-repo copy where future amendments also go.
(docs/TICKETS.md also carries the in-flight #000048 index-row update
from a concurrent session.)
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.
Wire vector quantization (the §3.1 production knob): arborist embed
--quant int8 [--rebuild]. The chunk_vecs vec0 column becomes int8[384]
vs float[384] per quant; the quant folds into VEC_BACKEND_VERSION
(...-384int8-... / ...-384float32-...) and vec_meta records it per
shard. Switching quant on an existing chunk_vecs requires --rebuild
(the vec0 element type can't be altered in place — embed_documents
raises ValueError telling you to --rebuild).
int8 serialization: scale each bge component by 127 (theoretical
[-1,1] range), clamp to [-127,127], round, serialize_int8. The same
scaling on the query vector → distances comparable; cosine is
scale-invariant so the uniform x127 cancels in the ranking.
sqlite-vec v0.1.9 quirk worked around: a bare blob inserted into a
vec0 column is interpreted as float32 regardless of the column's
declared type — int8 vectors MUST be wrapped in vec_int8(...). So
the INSERT and the MATCH now wrap the blob in vec_f32(?) (float32)
or vec_int8(?) (int8) — constructor name from a fixed dict, no
injection surface. (Discovered the hard way: a bare int8 blob into
an int8[384] column → "expected int8, but a float32 vector was
provided".)
VecBackend reads the quant from the existing chunk_vecs schema (or
defaults to float32) so search uses the matching wrapper. New module
exports: QUANTS, EMBED_QUANT, vec_backend_version(quant), existing_quant.
int8 head-to-head on crawl_appliedcombinatorics_org.db (168 chunks):
- storage: float32 1,597,440 B -> int8 417,792 B = 3.8x smaller
(~4x at corpus scale where the 1024-vector blocks fill; the
~28 KB of vec0 metadata doesn't quarter, hence 3.8 not 4.0).
- recall vs the float32 baseline:
Q "how many ways to choose k things from n":
identical top-5 (Combinations, Permutations, Exercises,
Derangements, Graph Coloring).
Q "pigeonhole principle counting":
identical top-2 (Graph Coloring, Exercises); ranks 3-4 swap
Derangements <-> Permutations at Δdistance 0.002 — sub-noise.
- embed speed unchanged (~3.9 chunks/s — model-load-dominated).
Conclusion: int8 is the obvious production config (§3.1's +6%-tax
recommendation confirmed empirically). v1 default stays float32 for
max fidelity; flipping the default to int8 is a fox call.
CLI (arborist/cli.py): arborist embed --quant {float32,int8}; output
JSON gains "quant"; embed_documents ValueError → exit 2 with the
"--rebuild" hint.
tests/test_search_vec.py (9 -> 16): test_int8_quant_roundtrips
(int8[384] schema, vec_meta version, search round-trip, quant
inferred by VecBackend), test_quant_mismatch_requires_rebuild,
test_invalid_quant_rejected.
#000039 status updated. Full suite: 2343 passed, 28 skipped.
(Unrelated parallel-clone work in the tree — Makefile, arborist/qa/
verify.py, bench/fixtures/5f/*, tests/test_bench_batteries.py,
tests/test_verify.py — is #000046's hard-fixture tier, not touched.)
Tracks the headroom #000046 (closed) left: the 6 over-grounds still
in falsification-hard-v1.jsonl (4 HYBRID_ENTITY where the
entity-proximity strategy matches a shared proper noun while the
answer's other salient term is wrong; 2 STRICT_PARAPHRASE where the
false claim recombines source tokens into a different true statement —
"Mercury is the largest" — and lexical token-coverage can't tell
recombination from grounding) + the 8 mis-segments in
formulate-hard-v1.jsonl (line/bullet-only parse_pointer_claims merges
multi-claim lines / splits wrapped bullets).
Options + recommended order: 2.1 verify_quotes entity
salient-token-disagreement gate (the direct analogue of #000046's
numeric gate — narrow, lexical, bench-safe; catches the 4
HYBRID_ENTITY) → 2.4 parse_pointer_claims sentence/clause
segmentation (closes the Formulate pack) → 2.2 sequence-aware
paraphrase match (conservative threshold; catches the 2
recombinations); defer 2.3 mini-NLI (heavier; only if 2.1+2.2 leave a
residue worth a model dep). Each landing bench-gated — verify_quotes
changes run `make bench-qa` before/after, parse_pointer_claims changes
run the Formulate fixtures + a QA smoke. Doc-only proposal; no code in
this ticket.
Next ID 000048 → 000049; #000046 "Headroom" section + TICKETS.md
index updated to point at #000048.
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.)
Fox: "i like both" — keep the lazy out-of-band pass as the default
AND add the eager opt-in. Plus the Phase-1 gap fix (incremental embed).
Key insight folded into the design (new ticket section 14): a chunk_id's
content is immutable in arborist — same content → same chunk_id;
different content → a NEW chunk_id (re-ingest makes a new doc_root +
new chunk_ids linked by supersedes; a chunker bump re-chunks → new
chunk_ids). So a chunk, once embedded, never needs re-embedding — the
ONLY re-embed trigger is the embedder changing (VEC_BACKEND_VERSION
bump). That makes the idempotency story clean.
arborist/search/vec.py — embed_documents() now:
- incremental=True (default): embed only chunk_ids NOT already in
chunk_vecs (chunk_id NOT IN (SELECT chunk_id FROM chunk_vecs)).
This is the after-ingest / cron / Prometheus-Sigma-sweep path —
it picks up exactly the newly-ingested chunks; re-running is a
cheap no-op once everything's embedded.
- incremental=False: re-embed every chunk with content (delete-then-
insert all) — the embedder-changed case.
- rebuild=True: DROP + recreate chunk_vecs first, then a full pass —
the clean VEC_BACKEND_VERSION-bump path (a search mid-rebuild never
mixes old- and new-model embeddings: the recreated table starts
empty and grows new-model as the pass runs). Implies non-incremental.
- Cold-evicted chunks (content NULL) still skipped; vec rows persist
and stay valid (content is identical on rehydrate).
CLI (arborist/cli.py):
- arborist embed --rebuild — the DROP+recreate+full-re-embed path
(default is incremental). Output JSON now reports "mode".
- arborist ingest --embed — eager opt-in: after the chunk+Merkle-
commit pass, incremental-embed this run's new chunks. Default
ingest does NOT embed. ingest output gains "chunks_embedded" when
--embed is set. Only surfaced when the [vec] extra is installed.
- Hoisted the _vec_ok check up to the top of build_parser so both
the ingest --embed flag and the search --backend / embed subcommand
can gate on it.
tests/test_search_vec.py (7 -> 9): test_embed_incremental_only_embeds
_new_chunks (second pass after a follow-up ingest embeds only the new
chunk; third pass is a no-op), test_embed_rebuild_re_embeds_all
(DROP+recreate+full pass; vec_meta still records the version).
Verified on crawl_appliedcombinatorics_org.db: incremental on an
already-embedded shard reports chunks_embedded=0 in ~1.8s; --rebuild
re-embeds all 168 in ~61s; semantic search after rebuild still returns
topically-correct hits ("pigeonhole principle counting" -> "AC Graph
Coloring" chunk containing "Generalized Pigeon Hole Principle").
Ticket section 14 added: the idempotency table (re-ingest / chunker
bump / cold eviction / embedder bump / superseded docs), the two
integration models (lazy default + eager opt-in; the lazy pass's
natural home is a Prometheus-Sigma unconscious-sweep task per #000037
section 3.1), the command matrix, concurrency notes, versioning.
Status line updated.
Full suite: 2339 passed, 28 skipped.
(Unrelated parallel-clone work in the working tree — Makefile,
arborist/qa/verify.py, bench/fixtures/5f/*, tests/test_bench_batteries.py,
tests/test_verify.py — is #000046's hard-fixture tier, not touched here.)
Extends the #000046 hard tier to the Formulate sub-battery.
bench/fixtures/5f/formulate-hard-v1.jsonl — 12 prose inputs that
arborist.qa.parse_claims.parse_pointer_claims SHOULD segment into a
particular claim lattice (recorded in expected_lattice). The parser
is line/bullet-based — one line ⇒ one claim, [E#] tokens attach to
it — so 8 of 12 it mis-segments: merges and-/semicolon-/dash-joined
or (1)(2)-enumerated multi-claim lines into one claim with all the
pointers, or splits a wrapped bullet into two. Those 8 fail at HEAD
on claim-count mismatch; the other 4 are well-formed bullet/numbered
lists / single claims the parser handles right. Rate at HEAD = 4/12
= 0.333, stable (parse_pointer_claims is deterministic). A
claim-lattice parser that does sentence/clause segmentation (split on
'. ', ';', subordinating conjunctions, inline enumerations) + joins
wrapped bullets lifts the rate toward 1.0 → positive γ·Δ5f for that
child fork.
make bench-5f-formulate-hard runs it (|| true past the runner's
nonzero-on-failures exit). Not in `make bench-5f` / `runner --all`.
tests/test_bench_batteries.py — test_5f_formulate_hard_pack_below_ceiling
(pins rate 4/12, source=live, the 8 fails are claim-count
mis-segments).
#000046 → "Phase 1 + Phase 2 landed"; two below-ceiling 5F subs now
exist (falsification, formulate). Closure still pending an actual
surface improvement (verify_quotes tightening — bench-gated — or
parse_pointer_claims segmentation) that lifts a rate. §5 + §6 +
TICKETS.md row updated.
(Makefile also carries an uncommitted chain-check-SQL improvement
from the concurrent #000039 session — NOT included in this commit;
staged only the bench-5f-formulate-hard hunk + the .PHONY line.)
The #000025 §10.14 calibration showed _delta_5{s,t,f} mean over a
battery's 5 subs, so a single-sub gain weighs 1/5 of face value (the
5× dilution). #000047 ships the knob to pick the aggregation, default
unchanged.
WeightSet.delta_aggregator ∈ {"mean","max","sum"} (default "mean") —
a categorical field, validated in __post_init__ against
DELTA_AGGREGATORS; from_dict takes it as a string. Default unchanged →
ScoredFork output byte-identical → no fork_score.ESTIMATOR_VERSION
bump.
fork_score._aggregate(deltas, how): mean = arithmetic mean, max =
max(0.0, max_i Δ_i), sum = Σ Δ_i; empty → 0.0. _delta_5s/_delta_5t/
_delta_5f take an aggregator arg (default "mean"); the 5F efficiency
bonus is added after the aggregated base (aggregator-independent).
fork_score passes weights.delta_aggregator. The per-sub
HARD_REGRESSION_FLOOR flags are computed before aggregation, so a
single-sub regression still forces REJECT under max/sum. The chosen
aggregator is recorded in ScoredFork.weights["delta_aggregator"] (via
WeightSet.as_dict()); fork_score_branches traceability stays via the
opaque weights_id — no schema migration.
bench/scripts/fivef_threshold_calibration.py gained §5 — runs the
#000046 below-ceiling pack (5f/falsification at 0.333) and shows the
verdict / γ·Δ5f under each aggregator; bench/results/5f-threshold-
calibration-2026-05-11.md §5 is the captured record. Default stays
"mean" — the conservative, noise-robust, regression-symmetric choice
matching docs/bench-maxing.md's per-rate floor framing; v8 picks
max/sum per-deployment.
Tests: 8 new in tests/test_fork_score.py + 1 anchor in
tests/test_fivef_threshold_calibration.py; tests/test_weights.py
as_dict field-set test updated to include delta_aggregator;
test_fork_score.py AUTOCOUNT tags (#000012 §286, warrant-substrate-
cookbook.md ×2) bumped 23 → 31.
#000047 closed; #000012 §8 §3 + TICKETS.md row updated.
Full suite: 2330 passed, 28 skipped.
Implements the optional vec backend from the #000039 doc, with the
"obvious" v1 tuning, and demonstrates it on a real corpus shard.
arborist/search/vec.py (new):
- VecBackend(SearchBackend) — ANN over chunk_vecs, UNGROUNDED hits
(same as FTS5; vec changes recall, never warrant — embeddings are
soft signal, never in the proof path).
- chunk_vecs vec0 virtual table + vec_meta — sibling tables, additive,
don't touch chunks/documents/the audit chain.
- embed_documents() — batched ingest; delete-then-insert per chunk_id
(vec0 doesn't honor INSERT-OR-REPLACE — re-inserting an existing PK
is a hard UNIQUE error), so re-runs are idempotent and content-
changed → re-embed works. Skips cold-evicted chunks (content NULL).
- Pluggable Embedder callable; default = fastembed bge-small-en-v1.5
(~130 MB ONNX, downloads on first use). load_vec_extension(conn)
toggles enable_load_extension + sqlite_vec.load.
- v1 hyperparams (VEC_BACKEND_VERSION = vec-v1-bge-small-en-v1.5-
384float32-cosine-flat): model bge-small-en-v1.5, dim 384, quant
float32 (int8/binary = the production storage knob per §3.1, not
wired in v1), metric cosine (bge outputs L2-normalized, so cosine
ranking ≡ L2 ranking), ANN flat (vec0 default), top_k 20. These
five fold into governance_policy_hash in a later phase (§6).
CLI (arborist/cli.py):
- — populate chunk_vecs
for --db; prints progress + timing.
- — semantic ANN search (errors with
an install/embed hint if [vec] missing or chunk_vecs empty).
- Both surfaced only when sqlite_vec imports (mirrors the [html] /
selectolax pattern).
pyproject.toml: [vec] optional extra (sqlite-vec>=0.1.9, fastembed>=0.4);
added to [dev]. Note: sentence-transformers is the heavier "official"
embedder path §5 names; fastembed is the lightweight ONNX one.
tests/test_search_vec.py (7 tests, skip-if-no-[vec]): deterministic
stub embedder (hash → unit vector) so the suite exercises the
sqlite-vec plumbing — ext load, schema, ingest, KNN, JOIN, Hit shape,
limit, idempotent re-embed, --limit cap, empty/unpopulated — without
the heavy fastembed model. Semantic quality is demonstrated on a
shard, not unit-tested.
Demonstrated on ~/.arborist/shards/crawl_appliedcombinatorics_org.db:
168 chunks embedded in ~37 s (mostly model load); semantic queries
return topically-correct hits — "how many ways to choose k things
from n" → top hit "AC Combinations", "binomial coefficient counting"
→ "AC Introduction" (integer-solution counting) + "AC Combinatorial
Proofs". None of the query tokens need stem-match the chunk — the
semantic-allusion-gap closure the ticket promised. chain-check on
that shard reports 0 after embedding (chunk_vecs is a sibling table).
#000039 status flipped to "in progress · Phase 1 landed"; Phase 2
(RRF hybrid fusion in query.py) gated on a ≥5pp recall-lift
measurement with no STRICT-rate regression (§8).
(Unrelated: tests/test_weights.py::test_as_dict_returns_all_eleven_fields
fails in the working tree — that's a parallel-clone in-flight change
to arborist/substrate/weights.py + its test, not touched here.)
Closes the "everything is at rate 1.0 so fork_score's bench-Δ terms
are inert" gap the #000025 §10.14 calibration surfaced — at least on
the 5F/falsification axis.
bench/fixtures/5f/falsification-hard-v1.jsonl — 12 near-misses, each a
FALSE/unsupported claim whose correct verdict is UNGROUNDED (recorded
in expected_reason). 8 of 12 are over-grounded by
arborist.qa.verify.verify_quotes at HEAD — its paraphrase
token-coverage strategy returns STRICT_PARAPHRASE, its entity-proximity
strategy returns HYBRID_ENTITY, both matching on incidental overlap
(shared entities/numbers, the same key terms stated in the opposite
direction) — so those tasks fail by design; the other 4 the verifier
handles correctly. Rate at HEAD = 4/12 = 0.333, stable (verify_quotes
is pure-lexical / deterministic). Built around the pre-documented gap
5f-fal-live-003.
make bench-5f-falsification-hard runs the pack; make
bench-fork-baseline-hard pins it to
bench/results/baseline-falsification-hard.json. Both targets `|| true`
past the runner's nonzero-on-failures exit (8 fixtures fail by design;
the JSON is still written).
tests/test_bench_batteries.py — test_5f_falsification_hard_pack_below_ceiling
(pins rate 4/12, source=live, every fixture asserts UNGROUNDED, the 8
fails are over-grounds not abstentions) +
test_fork_score_positive_gamma_5f_on_hard_falsification_improvement
(the worked example: fork_score(parent={5f/falsification: 1/3},
child={5f/falsification: 1.0}) → gamma*Delta5f ≈ +0.133 > 0, verdict
ACCEPT, no regression flags — the bench Δ-rate carrying signal it
can't carry while every canonical pack is at ceiling).
NOT in `make bench-5f` / `make bench-5s5t5f` / `make
bench-fork-baseline` / `runner --all` — the hard pack is a separate,
deliberately-failing artifact pinned on its own.
#000046 flipped to "in progress · Phase 1 landed"; closure pending an
actual verify_quotes tightening that lifts the rate (a separate,
larger task). ticket-000012 §8 §4 + TICKETS.md row updated.
Full suite: 2314 passed, 28 skipped.
The 5F threshold calibration (#000025 §10.14) surfaced two loose ends
that #000012's v8 acceptance protocol owns; opening them as proper
proposals so they don't get lost.
#000046 — Harder 5S/5T/5F fixture tier (below-ceiling baselines).
Every canonical 5S/5T/5F pack runs at rate 1.0 at HEAD, so fork_score's
α·Δ5s+β·Δ5t+γ·Δ5f terms are inert (a child can only move them ≤ 0).
The embedded packs can't be made "harder" (their evaluator is the
gold); a harder tier is necessarily a live-path tier where a real,
imperfect arborist surface produces wrong/partial output. Recommends
Option A scoped narrow: a `falsification-hard-v1.jsonl` built around
the verify_quotes entity-strategy gap already documented by
5f-fal-live-003, confirm the rate sits below 1.0 and is stable, pin it
via `make bench-fork-baseline-hard`, demonstrate a planted surface fix
lifting the rate. Status: open · awaiting go/no-go.
#000047 — ForkScore _delta_* aggregator (mean vs max vs sum). _delta_*
means over a battery's 5 sub-batteries, so a single-sub gain is worth
a fifth of face value (the 5× dilution from the §10.14 calibration:
+0.06 on one 5F sub → MARGINAL, same +0.06 on all five → ACCEPT).
Recommends Option D: parameterize delta_aggregator on WeightSet,
default "mean", record it in the breakdown + fork_score_branches; bump
ESTIMATOR_VERSION only if the default changes. Defer implementation
until #000046 lands a below-ceiling baseline so the aggregator choice
is benchable, not guessed. Status: open · awaiting go/no-go; parks
behind #000046.
TICKETS.md index rows + Next ID bump (000046 → 000048) landed in
8599ce3 (swept in by a concurrent commit; content is correct).
ticket-000012 §8 §3/§4 updated to name #000047 / #000046.
"One more iteration then close" (fox): added committed KAT-regeneration
scripts for both the T3 calculator and φ_PRG — the regen step was a
throwaway temp script before; now it's reproducible and the phi_prg
test's skipif reason ("run scripts/generate_phi_prg_kat.py") points at
a file that exists. Then closed#000036.
New scripts:
- scripts/generate_t3_bound_kat.py — regenerates
bench/fixtures/t3-bound/known-answer-tests.jsonl from a fixed 12-config
list (the §7 worked examples under max_envelope + non-default-C_B*
+ g=0 edge + explicit-b1_model pins for the other three models).
- scripts/generate_phi_prg_kat.py — regenerates
bench/fixtures/phi-prg/known-answer-tests.jsonl from a fixed 10-entry
list (placeholder/random seeds, one-bit-flip variants, block-boundary
dim_h=16/17, 4096 counter-rollover stress).
- Both verified to reproduce the committed fixture data lines byte-
for-byte (only the header comments changed, to reference the script).
Each docstring states: run after any algorithm change, then bump the
module version (CALCULATOR_VERSION / PHI_PRG_VERSION) so the fixture's
version field changes too.
Doc/test:
- test_t3_bound_calculator.py skipif reason now references the regen
script (matches the phi_prg test pattern).
- #000035 §3.3 + t3-bound.md §10.1 reference the regen scripts.
Closure (#000036):
- Status → closed · 2026-05-11 in the ticket file + TICKETS.md row.
Phase 1 + dav1d Tier-1/Tier-2 (Option B in v1) + KAT-regen tooling
all landed; all §5 acceptance criteria met; both dav1d closure
blockers cleared. Continuation: empirical C_B1/C_B2/C_B3 tightening
under #000043 (parks on v7 deployment data); landing the bound's
framing into a v7 plastic-training spec parks on that spec gaining
a deployment target; R2's architectural integrations (Merkle audit-
event commitment, SQD canonicalization, CTI clause-lattice, 5F
trigger, ForkScore security-risk) are separate tickets if wanted.
- t3-bound.md header flipped to "closed 2026-05-11".
Full suite: 2312 passed, 28 skipped.
