Three pieces, all read-only or additive — no shard mutation, no
schema-version bump:
1. shard_for_document(document_root, M) in arborist/document.py.
Pure function: int(document_root[:8], 16) % M. 22 tests cover
determinism, range-bounds, near-uniform distribution (±5pp at
N=20k), and seven lock-in fixtures so peers will disagree
loudly if anyone changes the formula.
2. corpus_shard_count meta field + get/set helpers in store.py.
Lives in the existing key/value meta table; SCHEMA_VERSION
stays at v9.8.0 (the DDL doesn't change and source_root is
layout-independent, so cache records survive a reshard).
Legacy shards (without the field) return None; reshard tool
populates it on every target shard at migration time.
3. Pre-migration snapshot captured to
bench/results/pre-migration-snapshot.json:
docs 3,468,392 (3,468,226 globally unique)
chunks 6,235,764
edges 90,593,537
audit 3,468,403
This is the reference set post-reshard row counts must match.
4. Audit-event extraction script writes all 3.47M events from
all 4 shards to /tmp/audit-events.ndjson (2.0 GB) for the
Option-A canonical-chain consolidation step. Verifies chain
integrity on extract — all 4 source chains report 0 breaks.
5. Fixed a wrong chunk count in docs/corpus-history.md
(had ~3.54M/shard; actual is ~1.56M/shard) and added the
edge-count column (~22.6M/shard, 90.6M total). 6.24M chunks
total, not 14.12M.
Tests: 29 new pass (22 routing + 7 meta). No existing tests
touched.
Pinned the canonical shard count at M = 4 based on real-Wikipedia
ingest + query benchmark (bench/shard_count_sweep.py). Captured the
"when does SQLite stop being the right substrate" decision tree so
future operators know what bench would justify a fork or replacement.
Bench numbers (Wikipedia 2003 cur dump, 2000 docs, 4 cells of
M ∈ {1, 2, 4, 8}, 50 FTS queries per cell):
M chunks/s q_p50_ms q_p99_ms attach_ms
1 3,648 0.05 0.20 1.61
2 5,649 0.03 0.19 3.24
4 6,405 0.07 0.30 9.00
8 6,959 0.03 0.28 9.65
Key observations:
- M=1→M=2 is the biggest ingest win (+55%). Most gain happens there.
- M=2→M=4 is +13%. M=4→M=8 is only +9% — diminishing returns.
- Real wikitext canonicalization is per-worker Python CPU bound, not
SQLite-writer-lock bound. More shards don't unlock more CPU.
- Query p50/p99 is flat across M within noise (50 queries small).
- ATTACH cost grows linearly: 1.6 / 3.2 / 9.0 / 9.7 ms.
Why M=4 specifically:
- Captures 92% of peak ingest throughput (6,405 / 6,959).
- 6 ATTACH slots free under SQLite's 10 ceiling for aux DBs
(qa.db, snapshots.db, selfmodel-chain.db, crawl_*.db, future
mesh_*.db) — comfortable headroom. M=8 leaves only 2 slots.
- Mobile-tolerable: phone NAND attach is 5-10x slower than NVMe;
M=4 = 45-90 ms cold start (instant), M=8 = 50-100 ms (sluggish
with no headroom).
- Matches fox's current 4-shard layout = cheapest migration.
Decision tree for when SQLite stops being right (full text in
ticket §"When the SQLite-default substrate stops being right"):
A. ATTACH ceiling pressure (auxiliary DBs grow past 5) → bench
forked SQLite with SQLITE_MAX_ATTACHED=125, M ∈ {16, 32, 64};
if attach cost stays linear past M=10, fork viable but pays
permanent "no longer stock sqlite3" tax.
B. Ingest hits >10k chunks/s sustained ceiling → first tune
page_size / WAL checkpoint / mmap_size / synchronous. If
tuning gets 2-5x, stay on SQLite. If still ceiling-limited,
candidates: DuckDB (columnar, MVCC, FTS), libmdbx (B+tree no
FTS; we'd build it). In-house DB rejected without specific
failure of those.
C. Federation needs multi-writer-same-shard → SQLite writer-lock
serializes peers, becomes federation bottleneck. First try
leader-election (single-writer-per-shard with WAL replication
to followers). If true multi-writer required, SQLite is wrong;
candidates: FoundationDB, CRDT-on-KV-store. DuckDB does NOT
solve this — its MVCC is single-process.
Honest verdict: for current arborist workload (single-writer-per-
shard, read-mostly federation), stock python3 sqlite3 is the right
substrate. None of A/B/C are close to firing. The bench discipline
exists to know what to measure when something changes.
bench/results/shard-count-sweep-2026-05-26T16-20-48Z.csv (synthetic
baseline) + 2026-05-26T16-31-34Z.csv (real Wikipedia) committed as
the load-bearing measurement for the M=4 choice.
v1 of the same-model substrate-delta harness's non-jagged metric.
For one corpus title, surface-perturb its question (numeral / accent /
hyphen / honorific / amp / brit) preserving the referent, then ask
whether retrieval surfaces the SAME target for canonical vs perturbed
phrasing. J_norm = XOR disagreement rate @k (lower = less jagged);
graded mean |Δrank| catches rank instability the binary metric misses.
Pure query --dry-run: no LLM, no verifier, no judge, no n=3 noise, no
5pp floor — the recall_at_k discipline. Reuses recall_at_k.probe +
mine_questions._surface_variant. Feeds #000012 ForkScore
ΔJaggednessReduction. A-vs-C answer-quality arm already exists under
#000057 (control_ab/control_sweep) — not rebuilt. Curvature + LLM-arm
jaggedness delta remain open (ticket §8).
make bench-jaggedness JAGGED_LIMIT=40 JAGGED_K=8
Ran diagnose_deflection and diagnose_coherence on the 20+20 NEG
fixtures from step 3 parts A+B, plus all 808 STRICT.
Result: 0/40 NEG fire on diagnose_deflection (because the fixtures
are token-overlap-correct by construction — the question's subject
appears in the answer; that's exactly the failure mode #000052 was
built to catch beyond lexical). 0/40 on diagnose_coherence (the
fixtures are well-formed sentences).
So:
UNION (relevance OR deflection) = relevance alone (no lift)
INTERSECT (relevance AND deflection) = 0/40
Multi-signal combination doesn't help on these failure shapes.
But the architectural finding is positive in a different way: the
sidecars and the relevance reranker cover NON-OVERLAPPING failure
shapes cleanly:
- diagnose_coherence: structural breakage (word-salad, vacuous,
phrase-component-reuse). Owns the 'incoherent answer' slice.
- diagnose_deflection: token-overlap mismatch (subject anchor not
in answer). Owns the 'wholesale topic drift' slice.
- bge-large relevance: semantic aboutness mismatch despite shared
tokens. Owns the 'topic-collision / mis-cite / on-topic-but-
not-answering' slice — what §3.2 was built for.
Each signal owns its own slice; combining is redundant on these
cases. That's the architectural validation of the §3.1 + §3.2 +
existing-lexical-sidecars split as orthogonal, not overlapping.
Manifest runtime_viability.as_multi_signal_factor updated from
'viable' to 'TESTED — does not lift; sidecars are complementary
not combinatorial on these shapes'. Cleaned up stale
nli-shadow-grid-n1-minilm.json.
Also cleaned up a stale nli-shadow-grid JSON.
The §3.2 arc is now complete:
step 1 candidate-bench (overclaim, contrived data)
step 2 real STRICT FP (over-pessimistic 'NOT VIABLE')
step 3A real-context deflection (positive reversal, 55% at fp=0)
step 3B real-context mis-cite (100% at fp=0, motivating failure covered)
step 3C multi-signal combination (no lift; clean architectural split)
The recommended operating point holds: bge-reranker-large + cleaned
+ θ ≤ -2.42 → 100% mis-cite, 55% deflect, 0% STRICT FP.
Hand-inspection of the bottom-15 STRICT-fires from the raw §3.2.2 step 2
sweep showed claim-lattice overlay markup ([E\d+ | title | hash: '…'])
depressing scores on correct concise answers (the 6× Henry-VIII case),
while true-positive deflections (broad-question / narrow-answer like
'winners of all major sports?' → just-one-sport) remained correctly
low-scored. So the noise FP class is the bracket metadata; cleaning it
should reduce FP without losing true-positive signal.
Built clean_for_relevance() in arborist/qa/relevance/shadow.py — strips
[E\d+ | ... ] blocks + trailing '...']' tails. Baked into
ShadowRelevance.check_question_answer / check_claim_source by default
(opt out with clean_input=False). relevance_shadow_sweep.py applies it
to inputs before _score_batch (opt out with --no-clean).
Re-ran the full 6-model sweep on the 808-cell pooled STRICT with
cleaning:
bge-reranker-large 21.8% → 18.6% (-3.2)
MiniLM-L-4-v2 15.0% → 9.4% (-5.6 pts, -37% rel)
MiniLM-L-6-v2 11.6% → 9.0% (-2.6)
MiniLM-L-12-v2 11.0% → 8.0% (-3.0)
bge-reranker-base 9.5% → 9.0% (-0.5)
MiniLM-L-2-v2 4.5% → 1.5% (-3.0 pts, -67% rel)
Universal improvement, every model better. Big surprise: MiniLM-L-2-v2
— the model that FAILED the candidate-bench separability (margin
-1.97, declared 'capacity floor') — has the LOWEST real-traffic FP
rate at its own cb θ (1.5%). Because L-2's compressed score range
gives it a low cb θ which few real STRICT pairs score below.
SEVENTH instance of 'candidate-bench doesn't predict real-traffic'.
Runtime-veto verdict UNCHANGED — still not viable; smallest fp=0 θ on
real STRICT is below the cb NEG max for every model, so at any
runtime-safe θ the catch on cb NEG is 0/12. But cleaning is now FREE
improvement for any soft-signal / advisory / contrastive use of the
relevance score. Hand-inspected bottom-10 post-cleaning confirms true-
positive deflection signal preserved.
All 6 swept rerankers (bge-large/base, MiniLM-L-2/L-4/L-6/L-12)
false-fire on 4.5–21.8% of real bench-qa STRICT at their
candidate-bench fp=0 θ. The smallest θ that yields fp=0 on real
STRICT is BELOW the candidate-bench NEG max for every model —
meaning at the runtime-safe θ, catch on the 12 candidate-bench NEG
= 0/12 across the board.
Structural reason: real bench-qa STRICT answers have a much wider
score distribution (bge-large STRICT: min -2.20, p10 +2.99, p50
+5.80, p90 +7.54) than the tight contrived candidate-bench POS
band. The bottom 10% of legitimate STRICT score below where the
candidate-bench NEG cases sat. Distributions overlap heavily; no
threshold separates them.
This is the §3.2 mirror of #000049 §7 #27's recall-side walk-back —
clean candidate-bench → fails real-pipeline gate. Same diagnosis:
lexical-candidate selection + cross-encoder scoring + hard threshold
doesn't survive real-pipeline heterogeneity.
Verdict: relevance reranker CANNOT be promoted to a runtime
demotion-only veto on this design. demote_below_score stays null;
manifest gains runtime_viability block documenting the negative
result + the still-viable advisory soft-signal uses (render-tail,
multi-signal advisory, contrastive Q→A vs claim→source delta).
bench/scripts/relevance_shadow_sweep.py + bench/results/
relevance-shadow-sweep-pooled808.json committed. Production
verifier unchanged; advisory only; no audit_mode effect.
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.
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.
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_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.
First grid (MiniLM, n=1 bench-qa 89 STRICT cells): the §7 #22 'fails the
gate' verdict was config-specific (defaults k=6/θc=0.5/θe=0.9) — the grid
finds k=1/θc=0.93/θe=0.5 → catch 8/12 falsification-hard (incl. BOTH
recombination fixtures hard-003 Mercury + hard-005 Einstein), 0/89 STRICT
FP, 0/26 synthetic-legit FP; 18/28 synthetic recombination. So the
lexical-candidate approach is NOT a dead end. (k=1 = no haystack;
θe=0.5 tighter than the clean-set 0.9.) Full multi-model + n=3
characterization next.
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.
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.
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.)
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.
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.
Closes the three open Phase-1b items of #000025; every §10 closure
criterion is now met, so the ticket flips to closed.
§10.14 — ForkScore threshold-calibration handoff to #000012.
bench/scripts/fivef_threshold_calibration.py (make bench-5f-threshold-
calibration) runs the canonical 5S/5T/5F packs + the 5F live packs and
reports baseline rates, observability granularity (1/n), and fork_score
verdicts on the parent vs synthetic child perturbations →
bench/results/5f-threshold-calibration-2026-05-11.md. Findings written
into ticket-000012 §8: keep SIGNAL_FLOOR / HARD_REGRESSION_FLOOR at
0.05; the small 5S packs (syntax n=10, semantics n=8) are coarser than
the floors so any regression there trips hard-reject (intended zero-
tolerance); the 5x averaging dilution in _delta_*; ceiling saturation
(every pack at 1.0 -> delta-rate terms <= 0). No constant change
shipped. 6 tests in tests/test_fivef_threshold_calibration.py.
§10.13 — feedback latency / efficiency on real workload.
run_feedback_loop now computes feedback_latency (listed in §5.5 since
Phase 1a, never implemented) — wall-clock seconds to apply a live
chain against its temp shard, surfaced per-task
(feedback_latency_seconds) + battery (feedback_latency_mean_seconds,
feedback_live_task_count). For live chains feedback_efficiency's cost
denominator switched from len(chain) (count of requested ops) to the
persisted footprint _persisted_cost = audit-event rows the chain
actually wrote + their body bytes / 1e6. Embedded chains keep
len(chain) and report feedback_latency_seconds = None. Latency is a
wall-clock field (run-to-run variable, like BatteryResult.timestamp)
and is not a fork_score input. 3 tests in tests/test_bench_batteries.py.
§10.11 — real selfmodel finetuning chains.
bench/scripts/selfmodel_chain_snapshot.py (make bench-5f-selfmodel-
snapshot) appends one chained SelfModel snapshot per run to a
persistent shard (~/.arborist/shards/selfmodel-chain.db, override via
ARBORIST_SELFMODEL_CHAIN_DB) with one CapabilityClaim per sub-battery
(metric = "5S-syntax" etc., measured_value = that pack's rate,
eval_digest = the pack's fixture digest, threshold = SIGNAL_FLOOR).
snapshot() auto-parents, so each snapshot is a distinct root and the
lineage grows by one per run. run_finetuning gains a third dispatch
mode — shard-chain (gated on a task's selfmodel_shard key) — via
_chain_finetuning_measure: reads the two most-recent snapshots
(latest() = child, its parent_selfmodel_root = parent) and measures
improvement on target_capability between them. This is the real
lineage replacing Phase-1a's synthetic parent->child pairs; the
chained delta reflects genuine cross-run drift (0.0 today — the
embedded packs are at ceiling). Operator pack
bench/fixtures/5f/finetuning-shardchain-v1.jsonl (6 tasks) + make
bench-5f-finetuning-shardchain; not in `make bench-5f`, `make test`,
or a fresh checkout (a missing/too-short chain fails honestly). The
real chain shard was bootstrapped 2-deep on 2026-05-11; make
chain-check-shards reports 0 breaks on it (and all other shards).
10 tests in tests/test_selfmodel_chain.py.
Full suite: 2311 passed, 28 skipped.
trigger_1_branch_density in bench/prometheus_sigma_trigger_probe.py read
the fork_score_branches table via branch_set_density() instead of the
"density check not yet implemented" stub. It groups rows by branch_set_id,
fires when the most-recently-recorded checkpoint carries >= 4 branches
(BRANCH_DENSITY_FLOOR), and surfaces n_checkpoints / latest_density /
max_density / n_checkpoints_clearing_floor in the markdown report so
section 12's "regularly" qualifier stays visible. Density sums across
shards per branch_set_id.
With no branch sets persisted yet the probe reports "table present but
empty across shards" (data_available True, fires False) rather than a
false negative. Re-ran the probe against the live shards:
bench/results/prometheus-sigma-triggers-2026-05-11.md.
6 new tests in tests/test_prometheus_trigger_probe.py (probe loaded via
importlib): no-table to no-data, empty-table to data-available-no-fire,
latest-checkpoint->=4 to fires, earlier-dense-but-latest-sparse to no-fire,
density-sums-across-shards, report-renders-density-lines.
Updated #000012 Phase 1c landing receipt, #000037 section 12 Trigger 1
note, and the TICKETS.md index rows for both. Pure measurement: no
mutation, no LLM call, no schema change.
The dry-run was still using safe_weights(); Phase 1.c shipped
sweep_weights() (γ_5f=1.5, λ_capital_cost=0.25,
ν_witness_divergence=0.5) precisely for sleep-sweep economics. Swap
both call sites in prometheus_sigma_sweep_dryrun.py and label the
report Weight profile: sweep (§15.4).
§15.4 adds the sweep profile alongside §15.1 safe / §15.2
conservative / §15.3 exploratory so the named-profile registry has
a single doc source of truth and the dry-run report's §15.4
reference resolves.
§22 dry-run-runs table (3 rows: initial uniform-τ-1d under safe;
per-mode τ under safe; per-mode τ under sweep) replaces the prior
inline narrative count. Includes a Phase-3-design observation: the
sweep profile flattens the softmax (DEFERRED 271 → 369; REJECT
205 → 110; ACCEPT 1 → 0; MARGINAL 2 → 3) because reducing
λ_capital_cost + ν_witness_divergence shrinks the gap between
high-Δ5F and low-Δ5F branches, so fewer branches reach Kelly's
p_i > 0.5 floor. Falsification-fixture proposal count stayed flat
(447 → 449) — the §13 step 11 emission path is upstream of decision
labeling, so sweep mode preserves information-gathering value while
concentrating action-taking on the high-confidence tail. Phase 3
ticket #000045 §2 deliberately leaves softmax temperature un-pinned
because of this trade-off.
Per Finding 3, a uniform τ_qa=7d filtered out every recent
CANONICAL_PROJECTION row (the π* graduations from #000027/#000030/
#000032 are all younger than 7d), so the sweep saw zero high-value
kernel-only work. Splitting τ_qa by audit_mode lets the cheap kernel
re-probe path (CP) run on a short cycle while the expensive LLM
re-witness path (STRICT/HYBRID/UNGROUNDED) keeps the long cycle.
bench/scripts/prometheus_sigma_sweep_dryrun.py: new build_tau_by_mode()
helper + per-mode CASE in iter_target_a_candidates; sweep_target_a now
takes the dict instead of a single seconds value. New CLI flag
--tau-qa-cp-days (default 1d); --tau-qa-days now scopes to LLM-witness
modes only (default 7d). Report renders the per-mode τ table in the
header and marks Findings 2 and 3 RESOLVED with their landing commits.
Makefile: PROMETHEUS_SWEEP_TAU_DAYS bumped to 7 (was 1, the prior
Finding-3 workaround); new PROMETHEUS_SWEEP_TAU_CP_DAYS=1 makevar.
bench/results/prometheus-sigma-sweep-dryrun-2026-05-10.md: regenerated
under the new defaults — 9 CP rows surface alongside 1,904 LLM-
witness candidates → 1,913 total Target A candidates → 1 ACCEPT,
2 MARGINAL, 205 REJECT, 271 DEFERRED chunks; 2 cache_drift vetoes
preserved end-to-end.
docs/tickets/ticket-000037 §22: Findings 2 + 3 marked RESOLVED with
landing-commit references; total dry-run cost line updated to the
new measurements (37.6 ms / 3,913 branches / 9.6 µs per branch).
Re-running `make prometheus-trigger-probe` after today's controller
landings shows the divergence-variance trigger has crossed both
thresholds:
Trigger 2 — divergence variance
Sample count: 37 (N_min = 30 ✓)
Mean: 0.7568, σ: 0.435
σ/mean ratio: 0.5748 (> 0.5 threshold)
Absolute σ: 0.435 (> 0.1 threshold)
Same-day morning probe (commit baseline) had only 16 samples and
did not fire; the additional witness-sweep / dry-run / harvest
activity through the afternoon brought sample count above N_min.
Agreement-label distribution across all shards:
KERNEL-LLM-DIVERGED 22
KERNEL-LLM-AGREE 6
LLM-DIVERGED 6
STRICT-WITNESSED 3
Trigger 1 (branch density) and Trigger 3 (witness cost share) did
NOT fire. Per §12 a single trigger firing is sufficient for Phase 1
gating — and Phase 1 has already landed. This commit captures the
empirical evidence that Phase 1 was on the right side of the gate.
Trigger 1 remains structurally blocked on #000012 Phase 1c
(fork_score_branches sibling table); that's the natural next move
if anyone wants to surface multi-branch consensus signals.
Closes four gaps in the initial Phase 1 landing (commit f625cac):
1. §14 row 4 — Hermes-saturation guard. `hermes_utilization` was on
the input contract but never consumed. Now: `utilization >=
budget` → DEFERRED with HERMES_SATURATED note + advisory event
carrying (utilization, budget) for Phase 2 audit. Three new
tests (saturation-equal, saturation-overflow, headroom-exists).
2. §13 step 11 — falsification-fixture proposal emission.
New `FalsificationFixtureProposal` dataclass; controller now
emits one per branch whose `witness_divergence >=
falsification_divergence_threshold` (weight-tunable, default
0.5). Emission fires BEFORE the all-vetoed cascade so vetoed-
AND-diverged branches still surface as 5F-fixture candidates
per #000025. Four new tests (high-divergence emits, low-
divergence stays silent, threshold is weight-tunable, vetoed-
diverged emits anyway).
3. Entropy + memory gating moved from module-level constants
(`H_LOW=0.3`, `H_HIGH=0.7`, `KAPPA_MEMORY=0.5`) to weight
fields (`h_low`, `h_high`, `kappa_memory`). Constants stay as
back-compat exports; defaults match exactly so byte-identical
behavior when neither override fires. Three new tests
(h_low/h_high tunable, kappa_memory tunable, back-compat
match).
4. Veto-class cascade hardening. Added explicit tests for the
`replay_window_unbounded` → ESCALATE path and the
`soft_hash_signal` → QUARANTINE path (§6 + §14 documented but
previously untested). Plus tests for the §6 fail-loud
priority ordering: ESCALATE > QUARANTINE > REJECT when mixed.
Dry-run regenerated against ~/.arborist/shards corpus:
Target A at τ_qa=1d surfaces **576 FalsificationFixture
proposals** from witness_divergence >= 0.5 (24% of swept
candidates). These rows are now an actionable funnel for 5F
fixture mining under #000025 §3.
Module: arborist/substrate/prometheus.py (+117 LOC, 899 total)
Tests: tests/test_prometheus.py (+196 LOC, 22 → 36 tests)
Dryrun: bench/scripts/prometheus_sigma_sweep_dryrun.py +
bench/results/prometheus-sigma-sweep-dryrun-2026-05-10.md
track the new falsification_proposals_total counter.
Full suite: 2326 passed, 37 skipped (+14 net from the new tests).
Phases 1 (controller) and 2 (sibling-table audit writes) landed in
prior commits. This commit adds the Phase 3 dry-run simulator
instead of the actual sleep-sweep scheduler, since Phase 3's value
is mostly in what we'd learn from running it — and the dry-run
captures those findings without committing to a scheduler design
prematurely.
bench/scripts/prometheus_sigma_sweep_dryrun.py — read-only
simulator that classifies §3 Target A (providence_cache) + Target B
(documents) sweep candidates, synthesizes ControllerBranches from
real shard data, runs the Phase 1 controller, reports decision
distribution + Phase-3-design findings. No LLM calls, no
mutations.
make prometheus-sweep-dryrun — produces a dated markdown report
at bench/results/prometheus-sigma-sweep-dryrun-YYYY-MM-DD.md.
Five findings surfaced by three dry-run iterations against the
live ~/.arborist/shards corpus (3.5M docs + 2839 providence_cache
rows) — captured in ticket §22:
1. chunk-size dominates Kelly threshold (must = Hermes
concurrency, not candidate pool)
2. flat capital_cost blocks every allocation (split kernel-cost
vs LLM-cost on the contract)
3. τ_qa=7d filters every CANONICAL_PROJECTION row (all 29 are
<7d old; need per-audit-mode τ)
4. Target B canonical-shape detection is the real headline
(~152K candidates extrapolated; controller correctly returns
MARGINAL on shape-match chunks)
5. quarantined rows correctly veto via cache_drift hard-veto
Mean per-branch controller latency in dry-run: 12.5 µs at
chunk_size=4. Phase 3's actual bottleneck is the witness fan-out
(Hermes calls), not the controller itself.
Ticket #000037 status flipped to in-progress with Phases 0+1+2
landed; Phase 3 scheduler remains future work but is informed by
the five findings.
Full QA-quality sweep, 75 questions × 3 modes × n=3 = 675 cells
against Hermes-3 8B at concurrency=4. No mode regressed past
the 5pp signal floor:
quote 53.8% STRICT vs 54% baseline = -0pp (stable)
claim_lattice_pointer 23.1% STRICT vs 20% baseline = +3pp (within floor)
claim_lattice (JSON) 46.7% STRICT vs 42% baseline = +5pp (at floor — marginal positive)
JSON-mode +5pp is right at the noise threshold per
docs/bench-maxing.md — could be the 2026-05-10 substrate work
(92/92 warrant chains + 18 textbook ingests + cascade tuning +
Phase 3 wiring) translating to retrieval-quality lift, OR
sample variance. Follow-up bench in 1-2 days disambiguates.
Format-collapse 0/225 in every mode. Latency stable
(10.6-11.1s median). Lattice-mode directive coverage 100%.
Phase 3 per-claim warrant-chain tail still dormant on this
fixture (rescue-only by design; no warrant-shape question
retrieves a chain-backed chunk while failing lexical anchor —
parallel shift's same finding). Source-level warrants tail
fires on claim-pack-targeted retrievals as documented in
phase3-live-validation-2026-05-10.md.
Substrate + Phase 3 sprint shipped clean.
Parallel-shift commit 6f1dbed already refreshed primary surfaces
but missed three sites:
docs/warrant-substrate-cookbook.md 54 → 74
docs/tickets/ticket-000031-...md 40 → 74 (Phase 2.5 narrative)
bench/results/full-warrant-resolution-...md 40 → 74 (×2 sites)
Live count from ~/.arborist/shards/000.db is 74 citation_aliases
+ 13 term_aliases. All four edits add the "grew 40 → 54 → 74
across the day" trail so future readers see the trajectory rather
than a stale point-in-time number.
Three-iteration investigation following c5bc53f's "Phase 3 is a rescue
mechanism, not a default path" finding. Question: can a paraphrase-
eliciting fixture actually drive Phase 3 to fire on real questions?
Key reading from arborist/qa/warrant.py:507 ("warrant_check"):
the function vacuous-passes (returns (True, []) without any anchor
extraction) unless the question shape is one of:
- relation (proper-noun anchors)
- date (4-digit year in claim)
- entity-list
- count-shape
- why-cause-shape
If none of those shapes apply, Phase 3 has nothing to suppress.
Three fixture iterations, all probed before benching:
| iter | strategy | chain hits |
|------|-----------------------------------|------------|
| v1 | "What is X?" / "Define X" | 12/13 |
| v2 | "How would you use X to ..." | 1/10 |
| v2' | "X: explain in plain English" | 2/13 |
| v3 | "Who proposed X?" / "Why does X" | 1/8 |
| v3' | "X: who first stated it?" | 1/8 |
Pattern: title-leading retrieval works (high title-token boost on
chain chunks) but those questions are definitional — no warrant
shape, no Phase 3 trigger. Warrant-shape questions ("who" / "when"
/ "why") trigger warrant_check but retrieval pulls toward
Wikipedia's prose articles instead of the terse claim-pack chunks.
Structural reason: claim-pack chunks (the 92 chain-root documents)
are deliberately terse — they state axioms / theorems precisely
with mathematical formalism but contain no biographical, historical,
or explanatory prose. So warrant-shape questions retrieve from
Wikipedia, where the prose context matches the question shape.
Phase 3's data condition (chain-backed chunk cited for a
warrant-shape claim) is empirically unreachable through normal
retrieval pathways on this corpus shape.
Verdict (`bench/results/phase3-warrant-chain-paraphrase-
investigation-2026-05-10.md`): Phase 3's runtime fire rate is
structurally bounded near zero on the current corpus. Mechanism
is correct, non-regressing, and unit-tested
(tests/test_warrant_chain.py). The lexical layer is strong enough
that the rescue isn't needed at measurable rates.
Recommendation: don't engineer Phase 3 fires via fixture design.
Track fire rate as a corpus-evolution signal. If claim-pack
content gains prose context or the corpus mix changes, Phase 3
will start firing organically and the audit-line tail
(`· warrant proven via chain ×N`) will surface it.
This closes the three-thread investigation cleanly. v1 + v2 + v3
fixtures + reports all retained as the empirical trail.
Followup to dee6623 ("bench: #000031 Phase 3 A/B finds mechanism
dormant on warrant-targeted fixture") — the initial A/B reported
Phase 3 dormant on a 7-question fixture where only 2/7 questions
retrieved any chain-root sources. Open question: was the dormancy
a fixture inadequacy or a structural property of Phase 3?
This commit answers it. Aggressive fixture (13 questions using
exact claim-pack title phrasing — "What is the Pigeonhole
Principle?", "Define the Axiom of Extensionality", etc.; titles
sampled directly from the 92-document chain-root set on shard
000.db) drives retrieval to chain-root chunks on **12/13
questions** (24 chain sources retrieved total).
A/B vs a57b194~1, n=3, 117 cells per condition:
cache_key parity: 39/39 cells match
Phase 3 fire rate (warrant_proven_idxs): 0 / 117 cells
STRICT-rate: 85/117 → 85/117 (Δ +0)
per-cell movement (LLM dice): 4 ↑, 4 ↓
Phase 3 fires zero times even though retrieval does land on chain
roots on 92% of questions. The structural reason now reads as:
Phase 3 is a rescue mechanism, not a default path. It fires only
when (a) a cited chunk has warrant-resolver chain backing AND
(b) the lexical warrant_check would otherwise fail. On
well-formed axiom questions, the LLM's answer contains the
technical terminology that the lexical anchor heuristic looks
for — so warrant_check passes via the cheap lexical path and
Phase 3 has nothing to suppress.
Two readings of this finding:
1. **Positive for the lexical layer.** It's strong enough on
clean axiom questions that the chain-fallback isn't needed.
The mechanism that lands warrants on cheap lexical evidence
continues to do most of the work.
2. **Phase 3 calibration needs a different fixture shape.** To
exercise the rescue path empirically, a future fixture would
need to elicit lexical-fail-but-chain-pass claims —
paraphrase-heavy questions, disputed-terminology topics,
questions where the LLM naturally drifts from chunk vocabulary.
Mechanism is correct + non-regressing (cache_keys identical → no
retrieval movement; 0/117 fire rate confirms defensive-only
behavior). No further changes warranted to the implementation.
C — separate finding (refactor sweep verification): a57b194 has
zero v7/v8 patterns. Final exhaustive sweep across arborist/ +
tests/ shows only the intentional historical-note comments in
arborist/cli.py (Was \`arborist v8 score\`...) and
arborist/substrate/__init__.py (explaining why v-dirs were
retired). Refactor cleanup is complete.
B — separate finding (Makefile bench-fork-score smoke):
make bench-fork-baseline + make bench-fork-score runs end-to-end
under the renamed CLI. Verdict MARGINAL (score 0.0 for self-
comparison), exit 0. Yesterday's Makefile:430 fix verified live.
End-to-end smoke against the live shard cluster (~/.arborist/shards
+ ~/.arborist/crawl) using the parallel-shift probe fixture
bench/qa_questions_warrant_chain_probe.txt (7 questions explicitly
targeting the 92 claim-pack-warrant-resolved records).
Result: 6 of 7 questions reach EVIDENCE-WARRANTED on the four-rung
ladder. Q1 ("axiom of line incidence in Hilbert geometry") shows
the source-level "warrants: 5 proven" tail correctly counting all
5 cited claim-pack records with derivation chains. Q5 (De Morgan)
falls to ANCHOR-WARRANTED-PARTIAL on mixed Wikipedia/claim-pack
retrieval; Q7 (pigeonhole) hits TITLE_MISMATCH because retrieval
surfaced a Wikipedia article that shares no token with "Bogart".
Both Q5 and Q7 are correct fail-closed behavior given retrieval
choices, not Phase 3 issues.
What this validates:
- warrant_chain_lookup loads live 92 core_roots
- warrant_proven_claim_idxs threads through verdict -> result
dict -> JSON output (was None on cache hits, [] on fresh
misses — correct cache invariance preserved)
- source-level "warrants: N proven" tail fires
- EVIDENCE-WARRANTED rung promotes via existing
_ladder_rung_for_lattice logic without render-layer changes
- empty-set fail-closed default preserves backward compat
What's NOT validated end-to-end (still unit-test-only):
- Per-claim "· warrant proven via chain ×N" tail. Fires only
when warrant_check would have demoted AND chain present.
Narrow regime; covered by test_verify_claim_lattice_suppresses
_warrant_missing_with_chain.
Picks up the parallel-shift's probe fixture (was untracked); now
committed alongside the validation journal.
Bench follow-up to a57b194 ("ticket #000031 Phase 3: warrant-chain-
aware verifier suppresses WARRANT_MISSING"): land the warrant-
targeted probe fixture + capture an A/B comparing pre-Phase 3
(a57b194~1) vs HEAD on 7 questions × 3 modes × n=3.
bench/qa_questions_warrant_chain_probe.txt (new fixture)
========================================================
7 questions targeting claim-pack-resolved topics: Hilbert geometry
(2), Russell classes (1), Cantor diagonal (1), De Morgan (1),
Boolean absorption (1), pigeonhole in Bogart (1). Picked because
their topical match against the 18 textbook substrates +
warrant-resolver derivations should — in principle — drive
retrieval to chain-root chunks where Phase 3's suppression logic
fires.
bench/results/phase3-warrant-chain-A-B-2026-05-10.md (A/B report)
=================================================================
Findings:
- **21/21 cache_keys match BEFORE/AFTER.** Phase 3 doesn't fold
into governance_policy_hash (warrant_chain_roots is a runtime
parameter, not a policy field), so retrieval + prompt + LLM
inputs are byte-identical between conditions. Any STRICT-rate
movement is real verifier signal, not LLM dice.
- **`warrant_proven_claim_idxs` is non-empty in 0 / 63 AFTER cells.**
Phase 3's suppression path never fires on this fixture.
- **STRICT-rate: 42/63 → 43/63 (+1 cell, +1.6pp).** Four cells
move (2 up, 2 down); within n=3 LLM noise.
Why dormant despite warrant-targeted topic selection:
1. Pre-bench probe via `--dry-run --json` showed only 2/7 questions
retrieve any chain-root sources — the claim-pack textbooks have
far fewer chunks than Wikipedia, so retrieval defaults to the
bulkier corpus on general queries.
2. Even on the 2 questions that DO retrieve chain roots, the LLM
cites the larger Wikipedia chunks more often.
3. On the rare cell where the LLM cites a chain-root chunk, lexical
`warrant_check` already passes — Phase 3 has nothing to suppress.
Calibration finding, not a defect:
- Phase 3 mechanism is correct (no regression on cells where it
doesn't fire; cache_keys identical → can't move retrieval).
- The data condition for Phase 3 to actually fire requires a
fixture where (a) retrieval surfaces chain-root chunks AND
(b) the LLM cites them AND (c) the cited claim's lexical
warrant_check fails.
- Future fixture would need very specific terminology that
biases retrieval AWAY from Wikipedia (exact theorem labels,
axiom names, etc.).
Same calibration shape as yesterday's progressive-AND smoke A/B
finding: a targeted change can be dormant on the fixture you're
benching against. The bench-fixture-design discipline that ticket
#000037 §12 codifies — "measured trigger, not calendar date" —
applies here too: until a fixture lands that exercises Phase 3,
the suppression mechanism's value is theoretical.
The +14-record close-out: most stuck records had substrate content
in already-ingested shards but the FTS5 cascade picked discriminator
tokens the substrate text didn't use as titles. Two surgical
additions to _build_record_query_cascade:
(1) parenthetical-phrase variant via new _phrase_from_parenthetical
helper. The discriminating term often lives INSIDE a trailing
parenthetical that _phrase_for_axiom was stripping:
"Kolmogorov's First Axiom (Non-Negativity)" -> '"non negativity"'
"Newton's First Law (Law of Inertia)" -> '"inertia"'
"Axiom of Side-Angle-Side (SAS)" -> '"sas"' (already had this via prior path)
(2) content-tokens OR-join (top 3) as soft fallback after the AND
join. Most stuck records had top-5-AND scoring 0 hits even when
substrate had clear content; OR-of-top-3 surfaces relevant
chunks reliably without sacrificing precision (citation +
title-haystack guards remain).
Plus one final citation-alias row: Boehm-Jacopini -> Software
Foundations LF (decision_by="fox 2026-05-10"). SF-LF Imp.v
formalizes the structured-program theorem in Coq with sequence /
conditional / while-loop — same theorem as Boehm-Jacopini 1966
CACM in modern formal-verification dress.
Per-pillar:
I: 12/13 -> 13/13
II: 9/10 -> 10/10
III: 11/13 -> 13/13
IV: 18/18 stable
V: 2/5 -> 5/5
VI: 3/5 -> 5/5
VII: 11/14 -> 14/14
IX: 11/14 -> 14/14
Total: 92/92 (100%)
Cumulative day-of-2026-05-10 lift: 18/92 -> 92/92 (+74 records,
+80 pp absolute). Started the day with only Hilbert pillar IV
resolved; ended with all 8 pillars fully resolved across 13
textbook substrates (Hilbert, Boole, Cantor, De Morgan, Russell
IMP/PoM, PM Vol 1, Peano, Dedekind, Newton, Laplace, Grinstead-
Snell, PLFA, SF-LF, Bogart, Levin, Judson, Aristotle).
40 citation-alias rows + 13 term-alias rows, all audit-disciplined
per #000041/#000042 (decision_by + decision_rationale per row).
All 1623 tests pass.
Under fox "work all paths" directive, four orthogonal lifts:
(1) 14 Levin alias rows for pillar VII compound citations — each
Stanley/Brualdi/Knuth ref now has BOTH Bogart AND Levin as
substitute. Lifted +2 (hockey-stick + stars-and-bars).
(2) 6 Goedel multi-substitute aliases — Goedel citations now have
4 substitutes each (Peano, Russell IMP, Dedekind, SF-LF). Lifted
+4 pillar III records (Axiom of Zero / Distinctness / Addition-
Successor / Multiplication-Successor variants).
(3) 2 SF-LF alias rows for pillar I formal proof axioms (Mendelson +
Enderton -> SF-LF). SF Logic.v has 12 hits "excluded middle",
40 "existential", 1 "double negation". Lifted +2 (Existential
Introduction, Double Negation Elimination).
(4) Grinstead-Snell Introduction to Probability ingest (GFDL,
Dartmouth CHANCE Project). Initial LibreTexts BFS crawl thin
(24 chunks nav-only at depth=2). Switched to Dartmouth PDF
-> pdftotext -layout -> single-page HTML on localhost:8765
-> standard arborist html ingest. Result: 339 chunks of
substantive probability content. Lifted +1 (Inclusion-Exclusion).
New ingest pattern useful for any PDF-only textbook.
Per-pillar:
I: 10/13 -> 12/13 (+2)
II: 9/10 stable
III: 7/13 -> 11/13 (+4 — biggest win)
IV: 18/18 stable
V: 1/5 -> 2/5 (+1)
VI: 3/5 stable
VII: 9/14 -> 11/14 (+2)
IX: 11/14 stable
Total: 77/92 (84%)
7 term aliases registered (physics Newton-vocab + arithmetic Peano-
vocab) but most didn't lift — cascade picks different tokens than
my aliases target. Cascade tuning would require code work in
_build_record_query_cascade / _phrase_for_axiom. Deferred.
Cumulative day-of-2026-05-10: 18 -> 77 / 92 (+59 records, +64 pp).
15 records still stuck. Top blockers:
- Cascade-tuning for substrate-content-exists-but-FTS5-misses
(Kolmogorov 3 axioms, Newton third law, Vandermonde, Catalan,
Injectivity of Successor, etc.) — code work
- PM *24+ not transcribed (Russell's Paradox Resolution)
- Boehm-Jacopini esoteric (no PD substitute)
Three threads landed in a single commit because they share the same
substrate (the #000037 §12 trigger probe shipped in c422216):
A — code-side stale-map sweep (arborist/qa/repair.py)
=====================================================
Walked inline TODO/FIXME/XXX markers across arborist/ + tests/. Five
hits: two false-positives (\\uXXXX in escape pattern docs), one
genuine deferral (raw_html cache in async_web_fetcher), and **two
stale TODOs in arborist/qa/repair.py** referring to "re-prompt
feedback path is future work" — even though `reprompt_repair` is
fully implemented (lines 140+ at the time of this commit), wired
through CLI `--repair-reprompts N` flag (cli.py:4574), and gated by
`policy["repair_max_reprompts"]`. Refreshed the docstring header
and the in-body comment to point at the actual function.
Same drift pattern as today's earlier ticket sweep (e84f453):
implementation lands, the TODO doesn't get refreshed, future readers
re-implement what's already there.
B — N-power follow-up to the §12 trigger probe
==============================================
c422216's first probe run reported trigger 2 (divergence variance)
at N=16, σ=0.5, ratio=0.8 — both threshold conditions would fire if
N reached the 30-sample N_min. To validate that the variance signal
holds at N≥30 (rather than vanishing on a wider sample), drove the
canonical-witness path 20 additional times via a new
bench/qa_questions_canonical_witness_npower.txt fixture
(10 arithmetic@v1 + 10 logic-kernel@v1 questions; all
canonicalize-then-LLM-witness without errors).
Result at N=37: trigger 2 fires. Ratio 0.575 > 0.5, abs σ 0.435 > 0.10.
Variance signal is real at the floor. Captured in
bench/results/prometheus-sigma-triggers-2026-05-10-npower.md with a
loud caveat at the top: this is N-power validation, NOT a measure of
real workload pressure. The §12 phase-1 go/no-go decision should
still come from natural workload N or operator-stated need.
The prior baseline (prometheus-sigma-triggers-2026-05-10.md) stays
in place as the workload-state-at-time-of-ticket-c422216 record;
this new -npower.md report is the statistical-power follow-up.
C — ForkScore Phase 1c proposal (#000012)
=========================================
#000037 §12 Trigger 1 ("ForkScore receives ≥4 candidate branches per
checkpoint") gates on multi-branch persistence — but #000012 Phase 1a
(landed 2026-05-08) is single-validator scoring of one fork at a
time, and Phase 1b (still open) is the consensus paper. Neither
persists branch-sets. So Trigger 1 structurally cannot fire today,
which the probe correctly reports as "no fork_score branch-set table
found".
Added Phase 1c to #000012 as a doc-only proposal (no code in this
commit):
- New `fork_score_branches` table (sibling to capital_ledger; does
NOT enter audit_events.event_hash preimage).
- Optional `--branch-set <ID>` flag on `arborist v8 score`.
- One read API: `branch_set_density(conn, branch_set_id)`.
- #000037 probe wires through the new function once it lands.
Not opened as its own ticket because operator pressure for it
hasn't surfaced naturally — gates on either #000012 Phase 1b
landing OR #000030 algebra/calc kernel expansion producing
competing-kernel branches an operator wants to compare. Captured
here so a future shift doesn't re-discover the gap.
Hygiene
=======
- make test → 1623 passed, 45 skipped
- make chain-check-shards → 0 across all 7 shards
- fox's in-flight #000037 ticket modifications left untouched
Phase 0 of ticket #000037 (Prometheus-Σ recursive falsification
controller) is doc-only and gates Phase 1 on §12 measured-pressure
triggers. This commit lands the empirical-evidence harness fox needs
for the go/no-go decision, plus the §16.2 test scaffolding so the
contract is discoverable from the test runner today.
bench/prometheus_sigma_trigger_probe.py
=======================================
Read-only walk of audit_events + capital_ledger across shards.
Reports each §12 trigger:
- Trigger 1 (branch density, ≥4 branches/checkpoint): looks for a
fork_score branch-set table; reports "no data — single-validator
ForkScore Phase 1a" when absent. Trigger structurally cannot fire
until #000012 multi-branch persistence lands.
- Trigger 2 (divergence variance, N≥30 + ratio>0.5 OR abs>0.10):
aggregates providence_canonical_witness audit-event bodies,
maps each agreement_label per #000028 §1.2 to a binary
LLM-divergence score, computes mean/stddev/ratio. Defensive on
the max(mean,ε) guard from §12 itself.
- Trigger 3 (witness cost share > 0.30): aggregates capital_ledger
rows; uses `material` (kWh proxy) as the canonical compute axis;
surfaces both `material` and `financial` so fox can pick a
different form if needed. Tags the report with the caveat that
the current ledger reflects ad-hoc activity, not a controlled
#000026 sweep.
- Trigger 4 (operator mission need): n/a — operator decision.
Pure measurement, no LLM, no schema change, no mutation. Output is
a markdown report at $(PROMETHEUS_PROBE_OUT) (default
bench/results/prometheus-sigma-triggers-<utc-date>.md).
Wired via `make prometheus-trigger-probe`.
bench/results/prometheus-sigma-triggers-2026-05-10.md
=====================================================
First captured baseline. Verdict on current shards:
Trigger 1: NO (no fork_score branch-set table; #000012 Phase 1a
is single-validator)
Trigger 2: NO (16 samples; N_min=30. But mean=0.625, σ=0.5,
ratio=0.8 — both ratio AND abs floors would fire
if N reaches 30. Signal is there; just needs more
samples.)
Trigger 3: NO (witness/total material = 0.005, well under 0.30
threshold. Caveated: ledger is ad-hoc, not a
controlled #000026 sweep.)
Trigger 4: n/a (operator-stated)
Empirical answer: no §12 measured-pressure trigger has fired yet.
Phase 1 of #000037 remains paper-only unless fox invokes Trigger 4.
tests/test_prometheus_sigma.py
==============================
17 skip-stubs pinning the §16.2 acceptance contract. Each test:
- Collects today via `pytest --collect-only` so the test surface is
discoverable.
- Skips with reason "Phase 1 not landed; controller module
arborist/v9/prometheus.py absent" until that import succeeds.
- Carries a one-sentence intent line tying it back to a numbered
ticket section (§5 / §6 / §7 / §10 / §13 / §14 / §16.1 / §4.4).
When Phase 1 lands, the implementer adds the controller module
under arborist/v9/prometheus.py per §13; CONTROLLER_AVAILABLE
becomes True; each test gets its body filled in. The names, intents,
and skip-reason strings are the durable contract.
Hygiene
=======
- make test → 1623 passed, 45 skipped (was 28; +17 new skips).
- make chain-check-shards → 0 breaks across all 7 shards.
- No code touched outside the new probe + scaffolding files +
Makefile target wiring. fox's in-flight #000037 ticket
modifications left untouched.
Three orthogonal lifts under fox 2026-05-10 "all of the above"
directive:
(1) 7 medium-confidence citation aliases per #000041:
Mendelson -> Russell IMP + De Morgan FNL + Boole Laws of Thought
Enderton -> Russell IMP + De Morgan FNL
Jech -> Cantor 1915 Contributions
Landau -> Russell IMP
Each row carries partial-coverage caveat in decision_rationale.
(2) Peano Arithmetices Principia 1889 ingested via existing
textbook_tex pipeline. Source: mdnahas/Peano_Book GitHub
repo (CC-BY-SA-4.0, Verheyen+Nahas English+Latin LaTeX).
paracol multi-column macros leak some residue but substantive
text extracts (39 chunks, 35 hits on "peano", 38 on
"arithmetic"). Make target textbook-peano. Manifest entry
peano-arithmetices-principia-1889. Plus 3 substitute-citation
aliases:
Landau -> Peano (lifts 2 pillar III records)
Goedel -> Peano (2 variant strings)
(3) Title-from-author backfill on Cantor / Russell IMP / Bogart /
Judson / Peano shards. Empirical defect: HTML/textbook_tex
sources took titles from <title> tags or URLs which on
Wikisource + PG do NOT include author surname, so the
resolver's _shard_matches_citation heuristic (requires
cited-author surname in title haystack) failed silently for
all alias-resolved chains where the substitute author wasn't
self-evident in the shard content. Surgical SQL UPDATE on
documents.title fixed it. Single biggest lift in this turn:
+13 records.
Per-pillar:
I: 4/13 -> 7/13 (+3, Russell IMP / Boole / De Morgan)
II: 0/10 -> 5/10 (+5, Cantor)
III: 0/13 -> 7/13 (+7, Peano + Russell IMP)
IV: 18/18 stable
V: 0/5 stable (Kolmogorov URAA-blocked)
VI: 3/5 stable
VII: 9/14 stable
IX: 5/14 stable
Total: 54/92 (59%)
Title backfill is a one-off SQL; the proper long-term fix is in
the source ingest pipelines (HTML / textbook_tex should pull
author from manifest's `author` field at ingest time and prepend
to document.title). Tracked under #000031 Phase 1 follow-up.
Three high-confidence alias decisions per fox 2026-05-10 + 14
full-string aliases for compound pillar-VII citations + a deeper
Newton recrawl.
Lift breakdown:
pillar IV (Hilbert): 18/18 unchanged
pillar VI (Goldstein/Newton): 0/5 -> 3/5 (+3)
pillar VII (Stanley/Bogart): 0/14 -> 9/14 (+9)
pillar IX (Dummit-Foote/Judson):0/14 -> 5/14 (+5)
Citation aliases live in ~/.arborist/shards/000.db (audit table per
#000041). 14 pillar-VII rows are full-string matches against the
compound source_ref strings (resolver looks up by exact source_ref;
parsed sub-citations are not yet alias-keyed individually).
Newton shard rebuilt: original `crawl_en_wikisource_org_newton.db`
followed Wikisource sidebar instead of book content (60 docs / 289
chunks of mostly nav cruft). Manifest now pins (1729) edition root
+ explicit subpage urls (Axioms_or_Laws_of_Motion, Definitions,
Rules_of_Reasoning_in_Philosophy); recrawl produced 120 docs / 779
chunks of actual Principia text in textbook_newton-principia-motte.db.
5 records stuck despite alias registered (FTS5 cascade misses):
Newton Third Law + Conservation of Momentum (modern vocabulary
absent from 1729 prose), 4 advanced pillar-VII identities (Hockey-
Stick, Vandermonde, Catalan, Stars-and-Bars — Bogart skews intro-
level), Orbit-Stabilizer (Judson covers it but cascade doesn't
surface). Tractable one-by-one; deferred.
Pillars I/II/III/V/lambda subset of IX still 0/X — those need
medium-confidence alias decisions (Mendelson/Enderton -> Russell IMP,
Jech -> Cantor) or Phase 4 textbook acquisitions (Landau, Goedel,
Kolmogorov, Church) under #000038.
The bench-qa-smoke A/B against commit 416f956 surfaced a methodology
gap: the smoke fixture is structurally insensitive to retrieval-side
changes — every smoke query succeeds at full-AND chain 0 on every
shard, so cache_keys match byte-for-byte BEFORE/AFTER and any
STRICT-rate variance is pure LLM dice. Discovered when 5pp signal
floor at n=3 produced ±20pp swings with identical cache_keys.
This commit pins a fixture that ACTUALLY exercises the new code path:
- 5 PROGRESSIVE queries (full-AND fails on at least one shard;
_progressive_and_token_chains drops shortest-tokens-first). Includes
the canonical Gundremmingen case from commit 2b9d1f0.
- 2 DF_FILTER queries (every AND chain fails on at least one shard;
search lands in OR-mode where _filter_or_pool_by_df trims
high-DF tokens).
- 2 CONTROLs (full-AND succeeds on every shard; cache_keys must
remain identical BEFORE/AFTER for any retrieval-side change).
Which-chain-fires-on-which-shard was empirically verified at HEAD
against ~/.arborist/shards (4 wiki shards, 1.5M chunks each). Header
documents the chain that wins per query; re-probe if shards drift.
Baseline A/B (commit 416f956 ON vs OFF, n=3, --burn between samples,
fixture run via make bench-qa-progressive-and):
category cache_key parity STRICT-rate delta median latency delta
PROGRESSIVE 9/15 drifted +0.0pp (11/45 both) -3.6s
DF_FILTER 3/6 drifted +5.6pp (10/18 → 11/18) -3.9s
CONTROL 0/6 drifted +0.0pp (15/18 both) +0.0s
The CONTROL +0.0s latency delta is the load-bearing diagnostic —
when retrieval is byte-identical the search-side change cannot
affect total wall, and it doesn't. PROGRESSIVE / DF_FILTER cache_key
drift is the change actively redirecting retrieval, with the
expected latency win and no STRICT regression.
Future retrieval-side work (#000039 sqlite-vec backend, synonym
expansion changes, embedding rerank, hybrid scoring) should bench
against this fixture in addition to the full bench-qa scoreboard.
Hygiene:
- make test → 1623 passed, 28 skipped
- make chain-check-shards → 0 breaks across all 7 shards
Adds Cantor (Jourdain 1915 transl., PD), De Morgan First Notions of
Logic (PG #67017, PD), Russell Introduction to Mathematical Philosophy
(PG #41654, PD), and Judson Abstract Algebra Theory and Applications
(GFDL-1.3) to the textbook manifest, plus per-book make targets +
a `textbooks-base-knowledge` bulk target. Idempotent at DB layer.
These four primaries cover the proprietary citations across pillars
I/II/III/IX (Mendelson, Enderton, Jech, Barendregt, Dummit-Foote)
that dominate the 74 unresolved claim-pack records. They land as
honest "no match" under the existing resolver `_shard_matches_citation`
heuristic (cited-author surname required) — full lift gates on
fox citation-alias decisions per #000041.
Total ingest delta: 43 docs / 295 chunks across four crawl shards.
Bench journal under bench/results/textbook-base-knowledge-upgrade-2026-05-09.md
records per-book license rationale, projected lift table, and the
remaining #000038 acquisition gaps (Hilbert-Ackermann, Landau,
Gödel, Kolmogorov, Church) for which no clean PD HTML edition was
located in this session.
Smoke-test alias rationale was published with #000042, but two of the
three trailing records ("non triviality", "side-angle-side") needed
their own per-concept aliases since Hilbert 1902's vocabulary differs
in more than one place. Map of registered aliases (geometry domain):
incidence -> connection (4 records lifted)
euclidean parallel postulate -> axiom of parallels (1)
non triviality -> space axiom (1)
side angle side -> included angle (1)
Hilbert pillar IV is now the first complete warrant-chain vertical:
18/18 records bind to a primary-source surface chunk via Merkle
inclusion proof. 11 direct + 7 via-alias derivations rows. Bench
journal under bench/results/hilbert-pillar-iv-fully-resolved-2026-05-09.md
records reproducibility steps + per-alias signal data so a future
shift can extend the same pattern to Boole 1854, Newton 1729 Motte,
and Aristotle's traditional translations once #000038 lands those
surfaces.
Cluster-wide remains 18/92 (20%); the other 74 records cite textbooks
not yet ingested. Phase 4 content acquisition (#000038) gates next
lift.
Six-question Q&A sweep against the augmented shard cluster
post-#000031 Phase 3 commit (69e0a95). Looking for the new
`· warrant: N proven` tail surfaced by
_render_warrant_chain_tail when an answer's cited sources
include a claim-pack record with a warrant-resolver-v1
derivations row.
Result: 1 of 6 questions lit up the warrant tail.
Pasch's axiom → STRICT · via quote · warrant: 1 proven ✓
Hilbert geometry axioms → STRICT · via quote ✗
axiom of betweenness → HYBRID · via paraphrase ✗
Bayes theorem → STRICT · via quote ✗
modus tollens → UNGROUNDED · via quote ✗
binomial coefficient id. → HYBRID · via paraphrase ✗
1/6 matches the data substrate: 11 of 92 records have warrant
chains, all pillar-IV Hilbert axioms. The 5 unsigned cases are
honest:
- Hilbert / betweenness queries cited Wikipedia + crawled
textbook surfaces, NOT the claim-pack records — no claim-pack
record cited → no warrant chain to surface.
- Bayes theorem: even if cited, Kolmogorov isn't surface-
ingested → would stay at ANCHOR-WARRANTED ceiling.
- modus tollens UNGROUNDED: pre-warrant verifier issue,
orthogonal to this ticket.
- binomial-coefficient: claim-pack pillar VII cites Brualdi /
Stanley / Knuth, none surface-ingested.
No false positives observed — the render tail correctly stays
silent when chain-of-custody data is absent.
Roadmap: Adding any cited textbook to the manifest →
re-running warrant-resolve → more records resolve, more queries
light up the tail. The 74 still-unresolved records are a
prioritized acquisition list (Mendelson ~20, Enderton ~10,
Stanley ~5, Brualdi ~3, Knuth ~3). The substrate writes its
own roadmap.
Three streams. Two land in the repo; one lands in fox's Downloads
(existing g4 packs live there too).
Witness sweep automation
========================
`bench/scripts/witness_sweep_cron.sh` — schedulable harness that
runs `make bench-witness-sweep` + `make bench-witness-divergence`
back-to-back and (with --commit) auto-commits any new divergence
fixtures + sweep result. Fail-closed: pre-commit hook failure
aborts; non-main-branch refuses to commit.
$ bench/scripts/witness_sweep_cron.sh # dry run
$ bench/scripts/witness_sweep_cron.sh --commit # auto-commit
$ bench/scripts/witness_sweep_cron.sh --commit --push # full pipeline
Recommended cron (every 6 hours; LLM cost scales with frequency):
0 */6 * * * \
bench/scripts/witness_sweep_cron.sh --commit --branch main \
>> /var/log/arborist-witness.log 2>&1
Smoke this commit: dry-run produced 5 fresh divergences (10
cumulative across two extraction passes). The witness-sweep + 5F
extractor pipeline runs unattended now; calibration-data corpus
will grow over time.
#000018 — adversarial soft-hash covert-channel analysis
========================================================
`docs/soft-hash-channel-analysis.md` — formal threat model + channel-
capacity definition + cryptographic reduction + mitigation table
+ residual analysis. Doc-only deliverable per the ticket's hard
constraint.
Headline result: under threat models T1 (data poisoner) and T2
(gradient adversary), the channel is bounded by SHA-256 partial-
preimage hardness when φ (the anchor map) is independent of
parameter geometry. T3 (hyperparameter adversary) requires
M2 (per-checkpoint anchor nonce) for boundedness.
Recommended mitigation: M2 (per-checkpoint nonce) — 32 bytes added
to the snapshot manifest, no training-ergonomics regression,
closes T1+T2 channels under any φ choice and bounds T3 by
per-checkpoint window. Fall back to M1 (PRG-based φ) if M2-only
deployment surfaces structural concerns. M3 (drop anchor entirely)
stays in reserve as the strict-construction fallback.
Three open questions (§9): Hessian alignment under φ_linear,
PRG choice for φ_PRG, and explicit T3 per-window bound. Each is a
follow-up ticket.
Ticket #000018 status: closed · landed 2026-05-09 (analysis doc).
v7 § 9.10 spec amendment proposed in §7 of the analysis.
#000033 — pillar VII (combinatorics), Claude-authored
======================================================
NOT committed to the arborist repo (the existing g4-v2 packs live
in `/home/fox/Downloads/` too — that's the operator's bundle
location). Two new bundle files at:
/home/fox/Downloads/axiomsclaude-vii-v1.json (7 axioms)
/home/fox/Downloads/theoremsclaude-vii-v1.json (7 theorems)
Pillar VII covers combinatorial counting — the gap between Grok's
pillars VI and IX in the v2 packs:
axioms (7): addition principle · multiplication principle ·
pigeonhole principle · factorial definition ·
binomial coefficient definition · Pascal's rule ·
empty-set / boundary axiom
theorems (7): binomial theorem · inclusion-exclusion (counting
form) · hockey-stick identity · Vandermonde's
identity · Catalan number closed form · stars-and-
bars · strong pigeonhole
Each record in the dual-thread format the existing g4 packs use
(Δ symbolic LaTeX + ∇ verbose prose + ∇ concise + sigil + formal
language + role + status + source_reference + date + foundational
group + category + subfield). Per fox's directive: explicit
authorship metadata everywhere — `authored_by: Claude (Anthropic)
— model claude-opus-4-7`. NOT Grok-generated; no silent invention.
Each record carries `pi_star_ref: combinatorics@v1` so the kernel
binding is explicit. Theorems list `depends_on_axioms` arrays so
each theorem cites the foundation axioms it bottoms out on.
Smoke test (committed alongside):
$ arborist --db /tmp/test.db ingest --source claim_pack \\
--bundle /home/fox/Downloads/axiomsclaude-vii-v1.json \\
--bundle /home/fox/Downloads/theoremsclaude-vii-v1.json
→ 14 docs, 14 chunks, 0 cross-bundle edges
Source attributions: Stanley EC1, Brualdi Introductory
Combinatorics, Knuth TAOCP Vol 1, plus historical sources where
applicable (Pascal 1654, Vandermonde 1772, Dirichlet 1834, Catalan
1838, Feller 1950 for stars-and-bars).
Tests: 1636 passed, 37 skipped (no regressions; pillar VII
ingestion smoke covered above).
Apples-to-apples FTS5 search comparison on a single shard
(~/.arborist/shards/000.db, ~867 K Wikipedia surface docs)
before vs after ingesting the claim-pack source (#000029)
bundles.
Method: copy 000.db twice; ingest axiomsg4-v2.json +
theoremsg4-v2.json into one copy; run identical search queries
against both. FTS5-only path so retrieval-layer behavior is
isolated from QA pipeline reranking.
Results: 3/4 representative queries show claim-pack record in
top-3.
modus tollens → claim-pack at #3 (BM25 31.77)
law of excluded middle → claim-pack at #3 (related axiom
in same pillar)
associativity of addition → claim-pack at #1 — displaces
Wikipedia's general "Addition"
article entirely
Bayes theorem → no top-3 lift; Wikipedia "Bayes"
+ "Bayes rule" articles dominate
via short-doc BM25 bias
Storage tax: sub-MB. 78 chunks against 6.2M existing chunks is
below filesystem allocation granularity. Claim-pack rides
essentially for free on the existing Wikipedia substrate.
Honest verdict: claim-pack ingest earns its storage tax for
narrow technical queries where pre-distilled CORE-shape content
has title-token advantage. Doesn't help on queries Wikipedia
already covers with focused articles. The Bayes theorem case
likely shifts under the full QA pipeline's body-coverage sqrt
rerank (which counters BM25 short-doc bias) — that's a separate
larger bench worth running once the retrieval layer's other
stages are exercised.
Four-rung audit ladder still caps at ANCHOR-WARRANTED per
#000029's honest ceiling. Surface-ingest of cited textbooks
(#000031) is the path toward EVIDENCE-WARRANTED.
Three small streams:
#3 — close#000030 properly
============================
All 7 phases + Phase 1b landed across two commits (`04f3f5d`,
`abe5988`). Status header updated; ticket body now carries a phase
landing table with commit refs:
Phase 1 algebra-symbolic@v1 04f3f5d
Phase 1b algebra-symbolic-simplified@v1 04f3f5d
Phase 2 calculus-derivative@v1 04f3f5d
Phase 3 calculus-integral@v1 fox-direct
Phase 4 calculus-limit@v1 abe5988
Phase 5 calculus-series@v1 abe5988
Phase 6 linear-algebra@v1 abe5988
Phase 7 function-sampled@v1 abe5988
Plus tabular-pinned@v1 (last reserved stub) graduated in abe5988
closes the registry chapter — 15 concrete π*'s, no remaining
reserved stubs. Index updated.
#5 — composition fixtures across new SymPy π*'s
================================================
12 new tests in tests/test_pi_star_compositions.py covering pairs
that compose naturally:
- algebra-symbolic ∘ algebra-symbolic — idempotency check (running
expand twice equals expand once for any expression).
- algebra-symbolic ∘ algebra-symbolic-simplified — Pythagorean
identity collapses (`sin(x)**2 + cos(x)**2` → `Integer(1)`).
- Generic invariants: composition propagates PiStarError; manifest
fingerprint is order-sensitive; composite domain == inner domain;
composite bytes == manual chain bytes.
Test discipline: most compositions use `register_in_registry=False`
via a small `_safe_compose()` helper since the registry rejects
duplicate keys (#000015 invariant), so test ordering would
otherwise matter. Only the registration-test path uses real
compose().
#4 — end-to-end witness sweep against real shards + Hermes
===========================================================
New script `bench/scripts/witness_sweep.py`. Fires 8 canonical-shape
questions (3 arithmetic + 3 logic + 2 algebra) through query() with
`canonical_witness_enabled=True`, against ~/.arborist/shards (real
shard cluster) + the actual Hermes endpoint (NOT StubClient).
Records the agreement matrix per question to
bench/results/witness-sweep.json.
`make bench-witness-sweep` Makefile target. Honors
`ARBORIST_SHARDS_DIR`.
First real sweep (this commit, against Hermes-3-8B):
agreement label count rate
KERNEL-LLM-DIVERGED 5 62.5%
KERNEL-LLM-AGREE 3 37.5%
───────────────────────────────────────────
divergence_count 5 62.5%
wall median / max 130 ms / 1.1 s
Hermes diverged on 5/8 of the canonical-shape questions:
- said `1/10` for `0.1 + 0.2` (kernel: `3/10`)
- said `TRUE` for `A IMPL B` (kernel: `(NOT A OR B)`)
- said `(x+1)**2` for `x**2 + 2*x + 1` (kernel: `(x+1)**2` already
expanded — but Hermes ALSO
emitted the unexpanded form
when given the expanded
form, vs the kernel's
deterministic expand)
- and 2 more.
These are real LLM hallucinations on questions with closed-form
ground truth — exactly the calibration-data stream #000028
imagined. Pipeline validated end-to-end against actual hardware.
Pair: `make bench-witness-divergence` then extracts the 5
divergences as 5F-Falsification fixtures
(bench/fixtures/5f/falsification-witness-v1.jsonl, also committed).
Re-running the extractor produces byte-equal output (idempotency
contract from the extractor work).
Tests
=====
Full suite: 1636 passed, 37 skipped (was 1624; +12 composition
tests). The witness-sweep + extractor produce real artifacts now
committed under bench/results/ and bench/fixtures/5f/.
