probe(contract): WordNet 4⁴ fold makes HHTL adjacency semantic — 2.47 hops the 16-ary router cannot address - #875
Conversation
… gates green
Tests the operator's claim that "wordnet makes CLAM HHTL spacial
activation via 4^4", and the message before it: "4^4 would be a sparse
adjacent if the sentence refers to out of bounds meaning".
WordNet removes the hard part. A codebook must DISCOVER ancestry with
k-means and hope it survives; WordNet's @ hypernym relation IS ground
truth, so the address does not discover structure, it encodes it. The
question sharpens to: does folding real ancestry into a fixed 4-ary
depth-4 address preserve semantic distance well enough to be a search
prior, and does 4-ary buy anything the shipped 16-ary NiblePath cannot
express?
Measured on WordNet 3.1 (82,192 noun synsets, root=entity, 65,292
leaves, depth 19):
W1 ancestry-by-construction real +0.4938 vs shuffled -0.0356
W2 monotone ladder 15.78 > 12.76 > 11.15 > 8.69 > 7.05
(spread 8.73 hops, strictly decreasing)
W3 spatial activation band 0.855 vs random 0.568 = 1.50x,
under the 0.95 cover guard (twin-tested)
W4 sub-nibble structure nibble sees ONE bucket (10.55); 4-ary
splits it 11.15 vs 8.69 = 2.47 hops
W5 fold balance 256/256 cells, occ 29/255/1270
W4 is the headline: the 16-ary router expresses two levels per byte and
is blind inside a top nibble; the 4-ary address splits that population
by 2.47 WordNet hops. That is what makes a graded escalation band
implementable (3 siblings x 4 levels = 12 of 256 cells, 4.7%) where
16-ary offers 15 siblings or nothing.
Two first-run failures are kept as part of the finding: the first fold
reproduced the le-contract's own "lacking proper bucket rollover ...
saturates silently" warning (one cell held 15,769 leaves against a
median of 20) and needed an explicit granularity knob; and the W1
falsifier could not falsify, because a cell-label permutation is a
bijection so same-cell pairs survive it and scored +0.645 on identity
alone while looking like validation.
Division of labour recorded per operator ruling: the fold is the
ADDRESS, CLAM is the established CALCULATOR (not re-derived here),
HHTL+helix residue is the named alternative, and this probe's LCA walk
is an ORACLE for scoring only, never a runtime path.
Boundaries: structure probe, not a codec probe; first-parent-only turns
the DAG into a tree so a polysemous concept gets one address (visible in
the worked example, where 'dog' lands in WordNet's informal-term-for-a-
man sense); path length ignores information content; and 4-ary is
measured BETTER, not CHEAPER — no traversal-cost benchmark ran.
Corpus is local-only and gitignored; the generator is committed.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01K3RyLEbuNSHxxB3NTTrGki
|
Warning Review limit reached
Next review available in: 9 seconds Enable usage-based reviews in Billing to review now. Otherwise, wait until the next included review is available. How can I continue?After more reviews become available, a review can be triggered using the To avoid repeated limits, reduce automatic review volume by pausing incremental auto-reviews earlier, using label-based review opt-in, excluding WIP or generated PR titles, or requesting reviews manually when the PR is ready. If your team needs uninterrupted high-volume reviews, an organization admin can enable usage-based reviews. How do review limits work?CodeRabbit enforces per-developer PR review limits for each organization. Most developers receive the normal plan review availability. For paid Pro and Pro+ PR reviews, CodeRabbit uses adaptive limits for sustained high-volume activity. When a developer's recent PR review activity reaches the 95th percentile or higher among CodeRabbit users, additional reviews become available more gradually as earlier reviews age out of the rolling window. Please refer docs for additional details. Review details⚙️ Run configurationConfiguration used: Organization UI Review profile: CHILL Plan: Pro Plus Run ID: 📒 Files selected for processing (4)
Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out. Comment |
|
@coderabbitai review Requesting the review while this is still draft deliberately — in this repo merges have repeatedly outrun reviews, so the findings are worth more now than after the merge button. Two places worth the most scrutiny, since they are where a probe like this usually goes wrong:
Also flagging what the probe deliberately does not claim: 4-ary is measured better, not cheaper — no traversal-cost benchmark ran, and the Generated by Claude Code |
🧩 Analysis chain🏁 Script executed: #!/bin/bash
set -euo pipefail
printf '%s\n' '=== Changed files ==='
git diff --name-status "$(git merge-base HEAD origin/main)" HEAD 2>/dev/null || git diff --name-status HEAD~1 HEAD
printf '%s\n' '=== Probe outline ==='
ast-grep outline crates/lance-graph-contract/examples/probe_wordnet_44_activation.rs --items all --view expanded 2>/dev/null || true
printf '%s\n' '=== Generator outline ==='
ast-grep outline crates/lance-graph-planner/examples/data/wordnet/build_isa_tree.py --items all --view expanded 2>/dev/null || true
printf '%s\n' '=== W1/W3 markers and gate logic ==='
rg -n -C 8 -e '\bW1\b' -e '\bW3\b' -e 'shuffle|shuffl|same.cell|same_cell|correl|Pearson|random|0\.95|1\.5|band' \
crates/lance-graph-contract/examples/probe_wordnet_44_activation.rs
printf '%s\n' '=== Probe source with line numbers ==='
cat -n crates/lance-graph-contract/examples/probe_wordnet_44_activation.rs
printf '%s\n' '=== Generator source with line numbers ==='
cat -n crates/lance-graph-planner/examples/data/wordnet/build_isa_tree.py
printf '%s\n' '=== Relevant diff ==='
git diff --find-renames "$(git merge-base HEAD origin/main)" HEAD -- \
crates/lance-graph-contract/examples/probe_wordnet_44_activation.rs \
crates/lance-graph-planner/examples/data/wordnet/build_isa_tree.py \
EPIPHANIES.md \
exec-runs/wordnet-44-activation.md 2>/dev/null || trueLength of output: 84620 🏁 Script executed: #!/bin/bash
set -euo pipefail
printf '%s\n' '=== WordNet execution report ==='
cat -n .claude/board/exec-runs/wordnet-44-activation.md
printf '%s\n' '=== WordNet board entry ==='
rg -n -C 12 -i 'wordnet|4.ary|4\^4|spatial.activation|W1|W3|inert|twin' \
.claude/board/EPIPHANIES.md
printf '%s\n' '=== Existing probe conventions for sampled nearest-neighbour metrics ==='
rg -n -C 8 --glob '*.rs' \
-e 'nearest.*neighbou?r' -e 'recall' -e 'candidate pool' -e 'random.*baseline' \
crates
printf '%s\n' '=== Ignore-policy evidence for generated WordNet artifacts ==='
rg -n -C 3 -i 'wordnet|w[no]db|isa_tree' .gitignore crates .claude 2>/dev/null || true
printf '%s\n' '=== Changed markdown diff ==='
git diff --find-renames "$(git merge-base HEAD origin/main)" HEAD -- \
.claude/board/EPIPHANIES.md \
.claude/board/exec-runs/wordnet-44-activation.mdLength of output: 200537
I found two W3 issues:
✅ Action performedReview finished.
|
…rate the guard Review (#875) found two defects in W3. Both confirmed; neither is fixed by moving a threshold. 1. MISLABELLED METRIC. The candidate pool was sampled with replacement and the nearest 32 taken without dedup, so one synset could occupy several slots — a weighted sampled-entry recall, not "recall of the 32 nearest neighbours" as documented. Both arms shared the bias so the ratio survived, but the label was false. Pool is now deduplicated. 2. THE TWIN GATE WAS ARITHMETICALLY MIS-SPECIFIED. Both arms credited the anchor's own cell, inflating the baseline to 0.621, which caps the achievable ratio at 1/0.621 = 1.61 while the fire-half demanded > 1.5 and the cover guard demanded < 0.95 — a 0.018-wide window. The original 1.50x pass was luck, not evidence; after dedup it read 1.44x and correctly failed. The fix is to measure the actual claim. Sparse adjacency is about references landing OUT OF CELL; a neighbour already in the home cell needs no band to reach it. Restricting to out-of-cell neighbours: band 0.754 vs random 0.052 = 14.43x The null validates itself — random-12 scores 5.2%, against 12/256 = 4.7%, the cells' exact share of the codebook. Home-cell inflation was MASKING the effect, not creating it. The all-neighbour figure is retained as SECONDARY and labelled saturating (0.895 vs 0.621 = 1.44x). The < 0.95 cover guard is now calibrated rather than asserted: the same measurement against a deliberately coarse 2-level address (6+1 of 16 cells) yields 0.998, which the guard rejects. That satisfies the workspace inertness rule — a threshold that never bites is decoration. Lesson recorded for the next twin gate: verify the two halves are mutually satisfiable by more than a hair BEFORE running, by computing the maximum achievable value of the fire statistic under the silent guard. Gates: 5/5 green. clippy -D warnings exit 0; fmt applied. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01K3RyLEbuNSHxxB3NTTrGki
|
Both W3 findings confirmed and fixed in 1. Mislabelled metric — confirmed exactly as described. Pool sampled with replacement, nearest-32 taken without dedup, so one synset could fill several slots. Both arms shared the bias so the ratio survived, but "recall of the 32 nearest neighbours" was a false label. Pool is now deduplicated. 2. The guard wasn't just uncalibrated — the twin gate was arithmetically mis-specified. Chasing your point exposed it: both arms credited the anchor's own cell, inflating the baseline to 0.621. That caps the achievable ratio at The fix was to measure the actual claim. Sparse adjacency is about references landing out of cell (the operator's framing: "if the sentence refers to out of bounds meaning") — a neighbour already in the home cell needs no adjacency band to reach it, and crediting it to both arms is what compressed the statistic. Out-of-cell only:
The null validates itself: random-12 scores 5.2 % against 12/256 = 4.7 %, the cells' exact share of the codebook. Home-cell inflation was masking the effect, not creating it. 3. The Recorded in both board files, including the generalized lesson: before running a twin gate, compute the maximum achievable value of the fire statistic under the silent guard — if the window is a hair wide, the gate is mis-specified regardless of which side it lands on. Thanks for the W1 verification too — agreed that 5/5 gates green, clippy Generated by Claude Code |
|
@coderabbitai review Re-requesting on Generated by Claude Code |
|
Tip For best results, initiate chat on the files or code changes. 🧩 Analysis chain🏁 Script executed: #!/bin/bash
set -euo pipefail
base="$(git merge-base HEAD origin/main 2>/dev/null || git rev-parse HEAD~1)"
printf '%s\n' '=== Files changed by PR ==='
git diff --name-status "$base" HEAD
printf '%s\n' '=== Files changed in requested follow-up commit ==='
git show --name-status --format='commit %H%n%s' 9739154
probe="$(fd -t f '^probe_wordnet_44_activation\.rs$' . | head -n 1)"
if [ -z "$probe" ]; then
echo 'Probe source not found.' >&2
exit 1
fi
printf '\n=== Probe outline ===\n'
ast-grep outline "$probe" --items all --view expanded
printf '\n=== Follow-up diff: W3 probe ===\n'
git show --find-renames --format= 9739154 -- "$probe"
printf '\n=== Current W3-related source ===\n'
rg -n -C 12 -i \
-e '\bW3\b' -e 'out.of.cell|out_of_cell|home.cell|same.cell|same_cell' \
-e 'random.*band|band.*random|candidate|dedup|unique|NEAR|0\.95|1\.5' \
"$probe"
printf '\n=== Recorded W3 results and calibration ===\n'
rg -n -C 10 -i \
-e '\bW3\b' -e 'out.of.cell|random.12|0\.052|0\.754|0\.998|0\.95|coarse|cover|dedup' \
.claude/board 2>/dev/null || trueLength of output: 200153 🏁 Script executed: #!/bin/bash
set -euo pipefail
probe="$(fd -t f '^probe_wordnet_44_activation\.rs$' . | head -n 1)"
echo "=== Probe: $probe ==="
echo '=== Focused follow-up diff ==='
git show --format= --unified=40 9739154 -- "$probe"
echo '=== W3-related line locations ==='
rg -n -i \
-e '\bW3\b' -e 'out.of.cell|out_of_cell|home.cell|same.cell|same_cell' \
-e 'random.*band|band.*random|candidate|dedup|unique|nearest|coarse|cover' \
"$probe"
echo '=== Source windows around all W3-related matches ==='
rg -n -i \
-e '\bW3\b' -e 'out.of.cell|out_of_cell|home.cell|same.cell|same_cell' \
-e 'random.*band|band.*random|candidate|dedup|unique|nearest|coarse|cover' \
"$probe" \
| cut -d: -f1 \
| sort -nu \
| awk '
{ lo=$1-10; if (lo<1) lo=1; hi=$1+16; print lo "," hi "p" }
' \
| while read -r range; do
echo "--- $range ---"
sed -n "$range" "$probe"
done
echo '=== Compact recorded W3 result excerpts ==='
rg -l -i -g '*.md' \
-e '0\.754' -e '14\.43' -e '0\.998' -e 'out-of-cell' \
.claude/board 2>/dev/null \
| while read -r file; do
echo "--- $file ---"
rg -n -C 6 -i -e '0\.754' -e '14\.43' -e '0\.998' -e 'out-of-cell' "$file"
doneLength of output: 134157
One documentation mismatch remains: the module-level W3 description at Lines 58-63 still calls the metric recall of all 32 nearest neighbours in the band. It should explicitly identify out-of-cell recall as the primary statistic and label the all-neighbour value as the secondary saturating diagnostic, matching Lines 598-604 and the recorded result. [review] ✏️ Learnings added
You are interacting with an AI system. |
Review follow-up on #875: the module-level W3 text still described "recall of the anchor's 32 nearest neighbours inside the 12-cell band", which is the SECONDARY saturating diagnostic, not the primary statistic the gate now computes. A doc-comment claim that does not match behaviour is a defect by this repo's own rule, not a cosmetic mismatch. Module docs now state: - PRIMARY = out-of-cell recall, with the reason (a neighbour already in the home cell needs no adjacency band) and the self-checking null (random 12 of the remaining 255 ~ 4.7%). - SECONDARY = all-neighbour recall, explicitly labelled the WRONG statistic, kept only for comparability in the record. - The cover guard is calibrated by a coarse 2-level arm it must reject. - The specification check this gate failed once: compute the maximum achievable fire statistic UNDER the silent guard before running. Here 1/0.62 ~ 1.61 against a >1.5 bar and a <0.95 bar left an 0.018-wide window. No behaviour change. 5/5 gates green; clippy -D warnings exit 0; fmt applied. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01K3RyLEbuNSHxxB3NTTrGki
|
Fixed in The module W3 section now states:
Also noting your bound, since it resolves something I'd left implicit: per anchor, all-neighbour recall ≥ out-of-cell recall, so No behaviour change. 5/5 gates green, clippy Generated by Claude Code |
Self-caught while replying to review. The two quantities are different
and both appear in the record:
12/255 ~ 4.71% the NULL for out-of-cell recall (the home cell is
excluded from the draw, so 255 candidates remain)
12/256 ~ 4.69% the band's share of the WHOLE codebook
Both round to 4.7%, which is exactly why the conflation is easy and why
it is worth pinning: the observed random-12 score of 5.2% is evidence
only against the correct null. Same class of imprecision as the metric
mislabelling this probe already corrected — the number was right, the
thing it was a number OF was stated wrong.
Docs-only; no behaviour change.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01K3RyLEbuNSHxxB3NTTrGki
![chatgpt-codex-connector[bot]](https://avatars.githubusercontent.com/in/1144995?s=60&v=4){kind=small}
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 105049c76b
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
…as inert Codex P1 on #875: the candidate pool was 20,000 of 65,292 addressed leaves (31%), so "the 32 nearest WordNet neighbours" actually meant "the 32 nearest within a sample" — an omitted leaf can be strictly closer than every sampled one, moving both the distance cutoff and which cells the neighbours occupy. Same class of defect as the earlier dedup finding: the label claimed more than the measurement. Fixed by scoring all 65,292 candidates (~20M tree walks, seconds in release) rather than relabelling the caveat. That also makes dedup moot — `addressed` is distinct by construction. Result strengthens: band 0.763 vs random 0.031 = 24.71x (was 14.43x on the sample) with the null tightening to 0.031 against the 12/255 ~ 0.047 expectation. The full-corpus run then exposed that the twin gate was STILL incoherent, the third instance of the same mistake in this probe: the fire half had moved to out-of-cell recall while the cover guard stayed on the all-neighbour statistic that had just been declared the wrong one — and that secondary had drifted to 0.941, nine thousandths from tripping a guard it should never have been gating. Both halves must measure the same quantity on the same basis. Recomputed like-for-like, the cover guard is INERT and is dropped: on the out-of-cell basis a coarse 2-level address scores 0.840, not >0.95, so the threshold separated nothing and only the degenerate band-is-the-whole- codebook case would trip it. It is structurally unnecessary for the primary as well — a band fixed at 4.7% of the codebook with no home-cell credit cannot be secretly dense; the saturation worry was only ever real for the secondary statistic. The honest silent half was already present: the random arm. Same statistic, same basis, same machinery, wrong cells, 0.031. Keeping a second non-firing threshold beside it would have been the exact defect this probe exists to catch (a threshold that never bites is decoration). Recorded generally: three defects in this probe were mis-specified FALSIFIERS and zero were wrong measurements. A broken falsifier fails silently — it reports PASS. Gates 5/5 green; clippy -D warnings exit 0; fmt applied. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01K3RyLEbuNSHxxB3NTTrGki
…uard claim Two defects in one doc block, both self-caught while reading the review diff: 1. `band_of`'s doc comment had been orphaned onto `band_generic` by an earlier insertion, so one function carried two descriptions and the other carried none. 2. The surviving text still sold the cover guard as "a threshold shown to discriminate" — a claim deleted one commit earlier, when the guard was measured on the primary out-of-cell basis, found inert (coarse scores 0.840, not >0.95), and dropped. A doc-comment claim that outlives the behaviour it describes is the same defect class this probe keeps finding. `band_generic` now documents what it actually is (the band at arbitrary depth) and what the coarse arm is now FOR (reported granularity context, not a guard). `band_of` gets its own one-line doc back. No behaviour change. 5/5 gates green; clippy -D warnings exit 0. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01K3RyLEbuNSHxxB3NTTrGki
2 participants
Tests the operator's claim that "wordnet makes CLAM HHTL spacial activation via 4^4", and the message before it: "4^4 would be a sparse adjacent if the sentence refers to out of bounds meaning".
Why WordNet changes the question
A codebook has to discover ancestry with k-means and hope it survives — that is
PROBE-CODEBOOK-44, and its real-data leg is capped by a Base17 fold ceiling (ρ≈0.26 on single words) independent of the codebook. WordNet's@hypernym relation is ground-truth ancestry, so the address does not discover structure, it encodes it. No embeddings, no k-means, no API key on the critical path. The question sharpens to: given real taxonomic ancestry, does a fixed 4-ary depth-4 fold preserve enough semantic distance to be a search prior — and does 4-ary buy anything the shipped 16-aryNiblePathcannot express?Measured — WordNet 3.1, 82,192 noun synsets, root
entity, 65,292 leaves, depth 19W4 is the headline — it answers "not sure if that's a thing in our substrate" with a number.
NiblePath(FAN_OUT = 16) expresses two levels per byte and is blind inside a top nibble; the 4-ary address splits that same population by 2.47 WordNet hops. That is what makes a graded escalation band implementable: 3 siblings × 4 levels = 12 of 256 cells (4.7 %), where 16-ary offers 15 siblings or nothing.Worked example:
dog@ cell0x5a(01|01|10|10) → nearest in-bandfoster-brother(3), macho(3), man(3), mother's_son(3), sirrah(3), Adam(4); best distance outside the band = 4 hops.Every defect found was in a FALSIFIER, never in a measurement
Six review rounds (codex ×3, CodeRabbit ×3, plus self-caught). The measurements held up every time; the machinery meant to be capable of killing them did not. A broken falsifier fails silently — it reports PASS. That is the most portable thing in this PR.
ARITYroots before balancing gave one cell 15,769 leaves against a median of 20 — "lacking proper bucket rollover … saturates silently". Fixed with an explicit granularity knob (ARITY·24). Arity alone does not balance a fold of a real taxonomy.>1.5and the cover guard demanded<0.95. An 0.018-wide window; the original 1.50× pass was luck. After dedup it read 1.44× and correctly failed.The fix for #3–#6 was never to lower a bar — it was to measure the actual claim. Sparse adjacency is about references landing out of cell; a neighbour already in the home cell needs no band to reach it.
The null validates itself: random-12 scores 0.031 against an expected 12/255 ≈ 0.047. Home-cell inflation was masking the effect, not creating it. The honest can-stay-silent half is the random arm itself — same statistic, same basis, same machinery, wrong cells, nothing found.
Generalized checks recorded for the next twin gate: (a) before running, compute the maximum achievable value of the fire statistic under the silent guard — a hair-wide window is mis-specified regardless of which side it lands on; (b) both halves must measure the same quantity on the same basis.
Division of labour (operator-fixed, same session)
"and CLAM to calculate, that's established" / "alternative is using HHTL+ helix residue":
Nothing here should be read as "the substrate computes distance by walking to an LCA." The LCA walk exists for the same reason tesseract-rs oracles link libtesseract: ground truth to score against, never to ship.
Boundaries (do not cite past these)
doglands in the informal-term-for-a-man sense, not the animal. A real limit of any fixed-arity address — reported, not hidden. Both arms see the same tree, so it cannot flatter the result.RouteAction's four variants (Skip/Attend/Compose/Escalate = exactly 2 bits) matching a rung's width is an observed consonance, not a wired mechanism.Contents
crates/lance-graph-contract/examples/probe_wordnet_44_activation.rs— the probe (zero-dep crate, 5 gates).crates/lance-graph-planner/examples/data/wordnet/build_isa_tree.py— corpus generator, beside itsbuild_wordnet_rail.pysibling. Generator committed, corpus gitignored + local-only, per this directory's convention.EPIPHANIES.mdprepend (E-WORDNET-MAKES-THE-4-ARY-ADDRESS-SEMANTIC-1) +exec-runs/wordnet-44-activation.md.Additive only — no core/contract type change.
Verification: probe ALL GATES GREEN;
cargo clippy -p lance-graph-contract --example probe_wordnet_44_activation -- -D warnings→ exit 0;cargo fmt -p lance-graph-contractapplied. Scoped-pper repo practice; the workspace-wide sweep isTD-WORKSPACE-FMT-DRIFT(measured 1,094 hunks / 64 files / 9 crates, none in these files).Queued next
PROBE-CLAM-VS-HELIX-RESIDUE— same 4⁴-addressed corpus, both calculators scored against this oracle. The seam already exists:helix::CurveRuler::from_hhtl(path, depth)turns the address into the place,index(k) = (start + 4·k) mod 17regenerates the interior, and onlyResidueEdge/Signed360are stored. Hypothesis: a finer address shifts work from stored residue to deterministic place, making W4's 2.47 hops the granularity a residue no longer carries. Two pre-registered failure modes: (a) the address enters the ruler only throughplace mod 17, so a 256-cell address may funnel to the same 17 start offsets and residue would not move for a reason unrelated to 4⁴; (b)gcd(4,17)=1buys coverage, not low discrepancy — stride 4's first four steps are0,4,8,12(one column, clumped) where the φ stride over 17 is 11 (17/φ = 10.51), which matters because HHTL terminates early and only short prefixes are ever consumed.🤖 Generated with Claude Code
https://claude.ai/code/session_01K3RyLEbuNSHxxB3NTTrGki