From 9064377740bbaa8b67bfdc0931061d09dbf94450 Mon Sep 17 00:00:00 2001 From: Davide Cappellini Date: Tue, 22 Sep 2026 21:56:06 +0200 Subject: [PATCH] `s` settled from OUR code (higher = LONGER clauses), and it cannot rescue the gun === TASK 1: THE DIRECTION QUESTION, ANSWERED WITH A DEMONSTRATION === I told the user `s` controls clause length but refused to claim the DIRECTION, because I had seen it described both ways. It is now read out of our own code - one site per core, in the Type I branch of `tmLearnDir` (`guns/tm_pattern.nim:258`, `guns/tsetlin.nim:224`, `tm_diag/tm_core.nim:110`): if pol * d > 0.0: if lits[lit] == 1: if cOut == 1: if rand < (s-1)/s: st += 1 # toward Include, w.p. (s-1)/s else: if rand < 1/s: st -= 1 # toward Exclude, w.p. 1/s else: if rand < 1/s: st -= 1 # toward Exclude, w.p. 1/s => **HIGHER `s` GIVES LONGER CLAUSES.** The include step runs w.p. (s-1)/s (rising with s); both exclude steps run w.p. 1/s (falling with s). DEMONSTRATED (49-bit draft, planted 2-literal rule, 3000 train / 1500 eval): s=1.0 len 1.51 acc 100% s=2.0 len 1.82 acc 100% s=5.0 len 3.02 acc 100% s=1.5 len 1.42 acc 100% s=3.0 len 2.19 acc 100% s=10 len 4.04 acc 99.7% s=20 len 5.17 acc 94.5% WHY s=1.0 DEGENERATES: (s-1)/s = 0 so the include step NEVER fires while 1/s = 1 so BOTH exclude steps always fire - Type I can only remove literals, so a clause can grow only through the Type II penalty. (On random labels that leaves 43/120 non-empty clauses vs 120/120 at s>=3.) USABLE RANGE ~[1.5, 5]. tm_pattern uses 3.0; tsetlin uses 1.5. === TASK 4: WOULD `s` HELP THE SHIPPED GUN? NO - MEASURED === Recompiling the offline driver with -d:TM_S_DEF= (source untouched) retrains the gun end to end: s mean len verdict warm acc margin vs majority 1.5 15.82 too long 32.22% -2.03pp 2.0 14.99 too long 34.03% -0.22pp 3.0* 19.17 too long 35.72% +1.48pp (*shipped) 5.0 20.47 too long 34.72% +0.47pp Lowering `s` shrinks the clauses and makes accuracy WORSE; raising it pads them and also loses. The shipped 3.0 is the best of the four, and **no value comes near the healthy 3-8 band.** Combined with the settledness finding, the shape is consistent with "NO CONSISTENT SHORT RULE EXISTS in this representation/target". So the bottleneck is the SIGNAL - now confirmed from a THIRD independent angle (settledness, churn trend, and clause shape). This is the measurement behind the decision not to spend effort sweeping N or s. === TASK 3: AN HONEST CORRECTION TO MY OWN HYPOTHESIS === I predicted that random labels would produce `too long` clauses (the TM padding). MEASURED: on this encoding noise reads as **short / `collapsed`** (mean 1.88, median 2.0, acc 33.3%) - the TM FAILS TO COMMIT rather than padding. So "too long" is not the noise signature, which means the shipped gun's 19.17 mean is not explained by label noise. Worth knowing. Adds diagnostic group 8: the clause-shape checker - full length distribution (min/median/p10/p90/std), per-polarity and per-class breakdowns, a `clauseShapeVerdict` against a parameterised healthy band (default 3-8), per-BLOCK length contributions, and clause coverage (mean firing clauses, effectiveClauses = participation ratio, top3Share). `healthLine` now appends `shape= ()`. Validation: `test_tm_clause_shape` 66 checks. A planted 2-literal rule reads `healthy` with the literals recovered exactly; per-block correctly names the planted blocks (WALLS 43.0%, BULLETS 28.1%) and buries an irrelevant block (3.4%, below its uniform 8.3% share); random labels read `collapsed`. REAL READING, shipped gun: mean 19.17 / median 16.00 / p90 44.80 / max 57, 173 non-empty of 200, 27 empty => **`too long`**; coverage firing/sample 48.53 (24.3%), effectiveClauses 97.85/200, top3Share 5.3% (voting NOT concentrated); per-block is diffuse with no dominator, EXCEPT **UNUSED 6.4%** - the always-true negations of the never-written bits 38/39 acting as FREE PADDING, the same bug the kit found earlier now visible as clause bloat. Guards: test_tm_clause_shape 66 (new), test_tm_diag 48, test_tm_automata_diag 55, diag_synthetic 17, diag_automata_validation 11, test_gun_harness 39, test_vbullet_metric 11, test_power_selection 3, test_adaptive_radar 41, test_tfil_ring_weights 24, test_power_policy 26, test_ram_decision 40, test_rack_membership 48, test_selector_tiebreak 19, test_tm_pattern_registration 20, test_vbullet_admit_gate 12. acceptance_offline_vs_online not run (needs a live battle; no tm_diag dependency). --- common_libs/tests/diag_tm_pattern_offline.nim | 9 + common_libs/tests/test_tm_clause_shape.nim | 304 ++++++++++++++++ common_libs/tm_diag/README.md | 160 ++++++++- common_libs/tm_diag/diagnostics.nim | 324 +++++++++++++++++- 4 files changed, 779 insertions(+), 18 deletions(-) create mode 100644 common_libs/tests/test_tm_clause_shape.nim diff --git a/common_libs/tests/diag_tm_pattern_offline.nim b/common_libs/tests/diag_tm_pattern_offline.nim index 95aaeb7..b372fa8 100644 --- a/common_libs/tests/diag_tm_pattern_offline.nim +++ b/common_libs/tests/diag_tm_pattern_offline.nim @@ -177,6 +177,14 @@ proc main() = echo &"# class{cls} votes={c.votes:<7} len={c.length} {c.text}" echo &"# -> necessary literals: {spec.describeClause(necessaryLiterals(m, bestSamples, cls), cls)}" + # ── CLAUSE-SHAPE checker (Task 2/4): distribution + healthy-band verdict ── + # The default healthy band is 3-8 literals for a ~50-bit problem; the shipped + # tm_pattern is 40 bits, so the same band is used and the verdict is explicit. + let sd = clauseShapeDiagnostics(m, bestSamples, spec) + echo "\n## CLAUSE-SHAPE readout (healthy band 3-8 literals)" + echo formatClauseShapeReport(sd) + echo "# SHAPE VERDICT: ", sd.verdict + # ── AUTOMATA-LEVEL metrics on the shipped GF head (Task 5) ── # settledness / diversity / histogram / per-input confidence / disagreement # are read DIRECTLY off the exported teams; churn is measured on a tm_core @@ -189,6 +197,7 @@ proc main() = ad.histogram = stateHistogram(m, 9) ad.inputConfidence = perInputConfidence(m, spec, bestSamples, 0.5) ad.disagreement = voteDisagreement(m, bestSamples) + ad.shape = clauseShapeDiagnostics(m, bestSamples, spec) let churnN = min(bestSamples.len, 10000) var churnSamples = newSeq[DiagSample](churnN) for i in 0..= half=2) + result.teams[0][2 * nl + 6] = 1 + +proc testClauseSummaryShape() = + let m = shapeModel() + let spec = tinySpec() + let samples = @[makeSample(8, @[1, 1, 1, 0, 1, 1, 1, 0], 0)] + let summ = clauseSummary(clauseInfo(m, samples, spec)) + check "summary nonEmpty = 3", summ.nonEmpty == 3 + check "summary empty = 5", summ.emptyClauses == 5 + check "summary mean length = 2.0", abs(summ.meanLength - 2.0) < 1e-9 + check "summary median = 2.0", abs(summ.medianLength - 2.0) < 1e-9 + check "summary min = 1", summ.minLength == 1 + check "summary max = 3", summ.maxLength == 3 + check "summary lengthHist", summ.lengthHist == @[0, 1, 1, 1] + check "summary posMean = (3+1)/2", abs(summ.posMeanLength - 2.0) < 1e-9 + check "summary negMean = 2", abs(summ.negMeanLength - 2.0) < 1e-9 + check "summary per-class mean = 2.0", + abs(summ.perClassMeanLength[0] - 2.0) < 1e-9 + check "summary emptyFraction = 5/8", + abs(summ.emptyFraction - 5.0 / 8.0) < 1e-9 + +# ── healthy-band verdict ───────────────────────────────────────────────────── + +proc verdictOf(mean: float, n = 4): string = + var s = ClauseSummary() + s.nonEmpty = n + s.totalClauses = 8 + s.meanLength = mean + clauseShapeVerdict(s) + +proc testShapeVerdict() = + check "mean 1.5 -> collapsed", verdictOf(1.5) == "collapsed" + check "mean 2.5 -> short", verdictOf(2.5) == "short" + check "mean 3.0 -> healthy (band low edge)", verdictOf(3.0) == "healthy" + check "mean 8.0 -> healthy (band high edge)", verdictOf(8.0) == "healthy" + check "mean 12.0 -> too long", verdictOf(12.0) == "too long" + check "mean 19.17 -> too long (the shipped gun)", verdictOf(19.17) == "too long" + check "no non-empty clauses -> n/a", verdictOf(0.0, 0) == "n/a" + # custom band is honoured + var s = ClauseSummary() + s.nonEmpty = 4 + s.totalClauses = 8 + s.meanLength = 2.5 + check "custom band 2-6 makes mean 2.5 healthy", + clauseShapeVerdict(s, 2.0, 6.0, 1.0) == "healthy" + check "custom collapsedMax makes mean 2.5 collapsed", + clauseShapeVerdict(s, 2.0, 6.0, 3.0) == "collapsed" + +proc testLengthHistText() = + let t = lengthHistText(@[0, 3, 1]) + check "lengthHistText skips empty bins", "len 0" notin t + check "lengthHistText renders len 1", "len 1: 3" in t + check "lengthHistText renders len 2", "len 2: 1" in t + +# ── per-block contribution ─────────────────────────────────────────────────── + +proc testBlockContribution() = + let m = shapeModel() + let spec = tinySpec() + let blocks = blockLengthContributions(m, spec) + check "one entry per block", blocks.len == 2 + # block0 (bits 0-3): 2 in clause0 + 1 in clause1 = 3 over 3 clauses + check "block0 total = 3", blocks[0].totalLits == 3 + check "block0 mean/clause = 1.0", abs(blocks[0].meanPerClause - 1.0) < 1e-9 + check "block0 clausesUsing = 2", blocks[0].clausesUsing == 2 + check "block0 max per clause = 2", blocks[0].perClauseMax == 2 + # block1 (bits 4-7): 1 in clause0 + 2 in neg clause2 = 3 over 3 clauses + check "block1 total = 3", blocks[1].totalLits == 3 + check "block1 mean/clause = 1.0", abs(blocks[1].meanPerClause - 1.0) < 1e-9 + check "shares sum to 1", abs(blocks[0].share + blocks[1].share - 1.0) < 1e-9 + # skipEmpty = false averages over all 8 clauses instead of 3 + let blocksAll = blockLengthContributions(m, spec, skipEmpty = false) + check "skipEmpty=false dilutes the mean", + abs(blocksAll[0].meanPerClause - 3.0 / 8.0) < 1e-9 + +# ── coverage / concentration ───────────────────────────────────────────────── + +proc testCoverage() = + ## 2 classes x 2 clauses = 4 clauses. class0/clause0 = {+b0.0} fires on the + ## all-ones sample; nothing else fires. So 1 of 4 clauses fires. + var m = newMachine(8, 2, 2, 64, 3.0, 1) + let nl = m.nLiterals + m.teams[0][0 * nl + 0] = 1 + let samples = @[ + makeSample(8, @[1, 0, 0, 0, 0, 0, 0, 0], 0), + makeSample(8, @[1, 1, 1, 1, 1, 1, 1, 1], 0), + ] + let cov = clauseCoverage(m, samples) + check "coverage totalClauses = 4", cov.totalClauses == 4 + check "coverage mean firing = 1", abs(cov.meanFiringClauses - 1.0) < 1e-9 + check "coverage fraction = 1/4", abs(cov.meanFiringFraction - 0.25) < 1e-9 + check "one firing clause", cov.firingClauseCount == 1 + check "effective clauses = 1 (single clause does all voting)", + abs(cov.effectiveClauses - 1.0) < 1e-9 + check "top3 share = 1 (all fires from one clause)", + abs(cov.top3Share - 1.0) < 1e-9 + +# ── report helpers ─────────────────────────────────────────────────────────── + +proc testReportHelpers() = + let m = shapeModel() + let spec = tinySpec() + let samples = @[makeSample(8, @[1, 1, 1, 0, 1, 1, 1, 0], 0)] + let d = clauseShapeDiagnostics(m, samples, spec) + check "shapeLine mentions shape/coverage", + "shape mean=" in shapeLine(d) and "coverage=" in shapeLine(d) + let rep = formatClauseShapeReport(d) + check "report has length histogram", "length histogram" in rep + check "report has per-block contribution", "per-block literal contribution" in rep + check "report has per-class mean", "per-class mean/median" in rep + +# ── Task 3: synthetic validation with KNOWN geometry ───────────────────────── + +const + NBits = 49 + BitA = 0 ## dist-to-nearest-wall (WALLS block) + BitB = 45 ## bullet-lateral-offset (BULLETS block) + NoiseBit = 17 + NClasses = 3 + +proc genRule(n, seed: int): seq[DiagSample] = + ## class2 = A AND B ; class1 = A AND NOT B ; class0 = NOT A. + var rng = initRand(seed) + for i in 0..= 3.0 and d.summary.meanLength <= 8.0 + # the recovered necessary literals are the 2-literal rule + var c2ok = false + let c2 = necessaryLiterals(m, ev, 2) + var pos2: set[uint8] + for l in c2: + if l < NBits: pos2.incl uint8(l) + c2ok = pos2 == {uint8(BitA), uint8(BitB)} + check "SHORT rule recovered exactly (A AND B)", c2ok + check "SHORT rule is NOT called too long", d.verdict != "too long" + + # ── per-block contribution: planted blocks dominate, irrelevant block dead ── + # Use the sparser s=3 model here: with less padding the irrelevant block is + # unambiguously negligible (below the uniform 1/12 = 8.3% share). + let btmpl = newMachine(NBits, NClasses, nClauses = 40, nStates = 64, + sValue = 3.0, seed = 1) + let bm = trainModel(btmpl, train, epochs = 25, seed = 777) + let bd = clauseShapeDiagnostics(bm, ev, spec) + let walls = blockByName(bd.blocks, "dist-to-nearest-wall") + let bullets = blockByName(bd.blocks, "bullet-lateral-offset") + let usBlock = blockByName(bd.blocks, "dist-from-us") + echo &"# per-block (s=3.0): walls(mean={walls.meanPerClause:.2f} share={walls.share*100:.1f}%) " & + &"bullets(mean={bullets.meanPerClause:.2f} share={bullets.share*100:.1f}%) " & + &"dist-from-us(mean={usBlock.meanPerClause:.2f} share={usBlock.share*100:.1f}%) " & + &"[uniform=8.3%]" + check "the WALLS block (holds the planted A) dominates", walls.share > 0.25 + check "the BULLETS block (holds the planted B) is a top contributor", + bullets.share > 0.15 + check "the irrelevant US block is negligible (below uniform share)", + usBlock.share < 0.083 + check "the planted blocks outweigh the irrelevant block by >5x", + usBlock.totalLits > 0 and + walls.totalLits + bullets.totalLits > 5 * usBlock.totalLits + # the noise MOTION bit is not singled out as a top block + check "the pure-noise turn-direction block is far smaller than WALLS", + blockByName(bd.blocks, "turn-direction").totalLits < walls.totalLits + + # ── RANDOM labels: report whatever the checker actually says ── + let rtrain = genRandom(3000, 5) + let rev = genRandom(1500, 6) + let rtmpl = newMachine(NBits, NClasses, nClauses = 40, nStates = 64, + sValue = 3.0, seed = 1) + let rm = trainModel(rtmpl, rtrain, epochs = 25, seed = 777) + let rd = clauseShapeDiagnostics(rm, rev, spec) + echo &"# RANDOM labels s=3.0: mean={rd.summary.meanLength:.2f} " & + &"median={rd.summary.medianLength:.2f} p90={rd.summary.p90Length:.2f} " & + &"max={rd.summary.maxLength} empty={rd.summary.emptyClauses}/" & + &"{rd.summary.totalClauses} verdict={rd.verdict} " & + &"acc={evalAcc(rm, rev)*100:.1f}% eff={rd.coverage.effectiveClauses:.1f} " & + &"top3={rd.coverage.top3Share*100:.1f}%" + # On THIS encoding noise does not pad clauses long; it fails to commit at all + # (short clauses, many firing, low concentration). Report and assert only that + # the verdict is a legal word and that it is not "healthy". + check "RANDOM labels are not reported healthy", + rd.verdict in ["collapsed", "short", "too long"] + check "RANDOM-label accuracy is near the majority baseline", + abs(evalAcc(rm, rev) - 1.0 / 3.0) < 0.10 + + echo "" + if failures > 0: + echo &"{failures} / {checks} check(s) FAILED" + quit(1) + echo "All ", checks, " clause-shape checks passed." diff --git a/common_libs/tm_diag/README.md b/common_libs/tm_diag/README.md index 60c9b8b..c6741f6 100644 --- a/common_libs/tm_diag/README.md +++ b/common_libs/tm_diag/README.md @@ -16,7 +16,7 @@ Everything is pure and offline — no battles, no Java, no harness. |---|---| | `feature_spec.nim` | `FeatureSpec`, `describe`, `describeClause`, `draftTMSpec()` (49-bit draft), `tmPatternSpec()` (40-bit shipped encoding) | | `tm_core.nim` | compact deterministic Granmo Table 2/3 multiclass TM (mirrors the tm_pattern core), introspectable clause layout | -| `diagnostics.nim` | the seven groups (re-exports the two above) | +| `diagnostics.nim` | the eight groups (re-exports the two above) | Import everything with: @@ -24,6 +24,65 @@ Import everything with: import tm_diag/diagnostics ``` +## Task 1b — the `s` specificity knob (direction SETTLED, from the code) + +`TM_S` / `m.sValue` is the Type I specificity knob. It appears in ONE place in +each core — the Type I feedback branch of `tmLearnDir`: + +`common_libs/guns/tm_pattern.nim` lines 258-265 (identical in +`guns/tsetlin.nim` lines 224-231 and `tm_diag/tm_core.nim` lines 110-121): + +```nim +if pol * d > 0.0: + # Type I (Table 2) collapsed to the resulting state move: + # c=1, lk=1 -> +1 (toward Include) w.p. (s-1)/s + # c=0, lk=1 -> -1 (toward Exclude) w.p. 1/s <- the missing counter-force + # lk=0 -> -1 (toward Exclude) w.p. 1/s + for lit in 0..=3`: +the specificity mechanism is switched off. + +**Practical usable range: `s` in roughly [1.5, 5].** `s=1.0` is degenerate and +must be avoided. The shipped guns sit inside the range: `tm_pattern` uses `3.0`, +`tsetlin.nim` uses `1.5`. The clause-shape checker's healthy-band verdict is the +tool that tells you whether a given `s` produced a sane geometry. + ## Task 1 — named features / clause rendering ```nim @@ -43,7 +102,7 @@ last bits (`UNUSED-38/39`) are a real bug**: `tmBuildBits` writes only 38 raw bits into an `array[TM_NBITS=40, uint8]`, so bits 38 and 39 are always 0 and their negations always 1. The kit reports them as constant dead inputs. -## Task 2 — the six groups +## Task 2 — the data / clause / feature / accuracy / curve / ablation groups All functions take a trained `TmMachine` (or an externally supplied clause set) plus `seq[DiagSample]` where `DiagSample.lits` is the pos-then-neg literal vector @@ -93,6 +152,55 @@ let rs = ablateScrambleFeature(tmplMachine, train, eval, spec, bit) input earn its bits"). Pass `baselineAcc` from `ablateBaseline` to avoid recomputing it per block. +## Task 2b / 4 — the CLAUSE-SHAPE checker (group 8) + +The automata metrics say whether the TM is settling. The clause-shape checker +says whether what it settled ON is sane: a rule needing ~19 conditions to fire +is almost certainly fitting noise. + +```nim +import tm_diag/diagnostics + +let d = clauseShapeDiagnostics(m, samples, spec, + healthyLo = 3.0, healthyHi = 8.0, + collapsedMax = 2.0) +d.summary.meanLength # 19.17 for the shipped gun +d.summary.medianLength # 16.0 +d.summary.p10Length # 4.0 +d.summary.p90Length # 44.8 +d.summary.maxLength # 57 +d.summary.lengthHist # index = length, value = count +d.summary.posMeanLength / .negMeanLength +d.summary.perClassMeanLength / .perClassMedianLength / .perClassHist +d.verdict # "healthy" | "too long" | "collapsed" | "short" | "n/a" +d.blocks # per-FeatureSpec-block literal contribution +d.coverage # firing fraction + effective clause count +echo shapeLine(d) # one line, same style as healthLine +echo formatClauseShapeReport(d) # full readout +``` + +* **length distribution** — `meanLength`, `medianLength`, `p10Length`, + `p90Length`, `minLength`, `maxLength`, `stdLength` and `lengthHist`, plus + per-polarity (`posMeanLength` / `negMeanLength`) and per-class + (`perClassMeanLength` / `perClassMedianLength` / `perClassHist`) breakdowns. + `lengthStats(lengths)` and `percentile(sorted, p)` are exported standalone. +* **healthy-band verdict** — `clauseShapeVerdict(summ, healthyLo=3, healthyHi=8, + collapsedMax=2)`: `> healthyHi` = `too long`, `< collapsedMax` = `collapsed`, + between `collapsedMax` and `healthyLo` = `short`, in band = `healthy`, no + non-empty clauses = `n/a`. The band is a PARAMETER of the call. +* **per-block contribution** — `blockLengthContributions(m, spec)`: for each + `FeatureSpec` block, the literals it contributes across clauses + (`totalLits`, `meanPerClause`, `share`, `clausesUsing`, `perClauseMax`). A + block that contributes to every clause is dominating; a block contributing + ~0 is dead (the block-level view of the dead-INPUT list). +* **coverage / concentration** — `clauseCoverage(m, samples)`: the fraction of + clauses that fire on a sample (`meanFiringFraction`), the participation-ratio + `effectiveClauses = (sum v)^2 / sum v^2` (if 3 clauses cast most votes it + reads ~3, not the configured count) and `top3Share`. + +The one-line `healthLine` now appends `shape= ()` alongside the +settledness / diversity / churn / disagreement verdicts. + ## Task 5 — the AUTOMATA level (settledness / diversity / churn / disagreement) Task 2 reads the clauses. Task 5 reads the **automata inside them**: one @@ -146,10 +254,11 @@ So a flip is exactly a change in the predicate `state > 0`. `perInputConfidence(m, spec, samples, threshold)`, `rankedInputConfidence(conf)`. - **one-line health summary** — `healthLine(ad)` gives - `settledness / diversity / churn trend / disagreement`, each with its own - verdict word (`settling`, `fidgeting`, `frozen`, `coherent`, ...), and - `automataVerdict(ad)` reduces the trajectory to one word. `formatAutomataReport(ad)` - prints everything. + `settledness / diversity / churn trend / disagreement / shape`, each with its + own verdict word (`settling`, `fidgeting`, `frozen`, `coherent`, `healthy`, + `too long`, ...), and `automataVerdict(ad)` reduces the trajectory to one word. + `formatAutomataReport(ad)` prints everything, including the clause-shape report + when `ad.shape` was computed. ### API @@ -161,7 +270,7 @@ let ad = automataDiagnostics(tmpl, samples, spec, settleThreshold = 0.5, nHistBins = 9, window = 100, measureChurn = true) # ad.machine, ad.settledness, ad.diversity, ad.churn, ad.disagreement, -# ad.histogram, ad.inputConfidence, ad.summary +# ad.histogram, ad.inputConfidence, ad.shape, ad.summary echo ad.summary # one line echo automataVerdict(ad) # "settling" | "fidgeting" | "collapsed" | ... echo formatAutomataReport(ad) # the full readout @@ -226,6 +335,37 @@ on a `tm_core` temporal one-pass proxy, live-order): - constant inputs flagged: bits 38/39 (the known never-written ones) plus 19/36/37 in this fixture - context: pooled warm accuracy 35.72% vs the 34.24% majority = **+1.48pp** +**Clause-shape reading (group 8, same run):** `mean=19.17 median=16.00 p10=4.00 +p90=44.80 std=13.89 min=1 max=57 nonEmpty=173 empty=27/200` -> verdict +**`too long`**. Per polarity `posMean=19.80 / negMean=18.70`; per class +`c0=11.4 c1=13.4 c2=31.2 c3=22.6 c4=21.6`. Per-block share is diffuse (no block +dominates): wall-near 10.8%, distance-band 8.9%, flight-band 8.7%, radial-frac +7.6%, speed-band 7.0%, closing 6.8%, **UNUSED 6.4%** (the always-true negations of +the never-written bits 38/39 are free padding), the rest 2-6%. Coverage: +`firing/sample=48.53 (24.3%) effectiveClauses=97.85/200 top3Share=5.3%` — the +voting is NOT concentrated on a few clauses. One-line health: + +``` +settledness=0.484 (settling) | diversity=0.267 (moderate) | churn=0.094/100 falling (settling) | disagreement=0.145 (coherent) | shape=19.17 (too long) +``` + +**Does `s` rescue this gun? MEASURED, no.** Recompiling the offline driver with +`-d:TM_S_DEF=` (source untouched) retrains the gun end to end: + +| `s` | mean clause len | shape verdict | pooled warm acc | margin vs majority | +|---|---|---|---|---| +| 1.5 | 15.82 | too long | 32.22% | **-2.03pp** | +| 2.0 | 14.99 | too long | 34.03% | **-0.22pp** | +| 3.0 (shipped) | 19.17 | too long | 35.72% | **+1.48pp** | +| 5.0 | 20.47 | too long | 34.72% | **+0.47pp** | + +Lowering `s` shrinks the clauses (15.8 at `s=1.5`) but makes accuracy WORSE; +raising it pads them and also loses. The shipped `s=3.0` is the best of the four, +and NO value gets the mean anywhere near the healthy 3-8 band. Combined with the +settledness finding, the shape is consistent with **"no consistent short rule +exists in this representation/target"** — the clauses are long, diffuse and +padded, and the knob cannot fix a signal problem. + The gun settles onto the within-battle labels but its settled rules barely beat the majority class, which points at the TARGET / representation rather than the inertia `N`. See the Task 5 report for the inertia discussion. @@ -254,10 +394,12 @@ let m = machineFromTeams(TM_NBITS, TM_CLASSES, TM_NCLAUSES, TM_NSTATES, TM_S, ```sh nim c -r -d:release --path:common_libs common_libs/tests/test_tm_diag.nim # 48 pure unit checks nim c -r --path:common_libs common_libs/tests/test_tm_automata_diag.nim # 55 automata-metric unit checks +nim c -r -d:release --path:common_libs common_libs/tests/test_tm_clause_shape.nim # 66 clause-shape unit + synthetic checks nim c -r -d:release --path:common_libs common_libs/tests/diag_synthetic.nim # Task 3 proof (17 checks) nim c -r -d:release --path:common_libs common_libs/tests/diag_automata_validation.nim # Task 5 A/B/C proof -nim c -r -d:release --path:common_libs common_libs/tests/diag_tm_pattern_offline.nim # Task 4 real reading +nim c -r -d:release --path:common_libs common_libs/tests/diag_tm_pattern_offline.nim # Task 4 real reading + shape ``` -See `common_libs/tests/diag_synthetic.nim` for the ground-truth validation and +See `common_libs/tests/diag_synthetic.nim` for the ground-truth validation, +`common_libs/tests/test_tm_clause_shape.nim` for the clause-shape validation and `common_libs/tests/diag_tm_pattern_offline.nim` for the real reading. diff --git a/common_libs/tm_diag/diagnostics.nim b/common_libs/tm_diag/diagnostics.nim index 9d9fe52..2516b69 100644 --- a/common_libs/tm_diag/diagnostics.nim +++ b/common_libs/tm_diag/diagnostics.nim @@ -1,7 +1,7 @@ ## tm_diag/diagnostics.nim — THE DIAGNOSTICS KIT (Tasks 2 + 5). ## ## Everything runs OFFLINE: a trained `TmMachine` plus a labelled sample set. -## No battles, no harness, no Java. The seven groups are: +## No battles, no harness, no Java. The eight groups are: ## 1. pre-flight DATA checks (dataChecks, shuffledLabelControl) ## 2. clause introspection (clauseInfo, clauseSummary, topClauses) ## 3. per-feature contribution (featureContributions, deadInputs, ...) @@ -12,6 +12,10 @@ ## voteDisagreement, stateHistogram, ## perInputConfidence, healthLine, ## automataDiagnostics) +## 8. CLAUSE SHAPE (Task 2/4) (clauseShapeVerdict, lengthStats, +## blockLengthContributions, clauseCoverage, +## clauseShapeDiagnostics, shapeLine, +## formatClauseShapeReport) ## ## A `DiagSample` carries the LITERAL vector (positive literals [0..nBits), ## negations [nBits..2*nBits)) exactly as the shipped TM stores it, plus the @@ -184,6 +188,68 @@ type lengthHist*: seq[int] ## index = clause length, value = count (non-empty) meanLength*: float maxLength*: int + ## ── clause-SHAPE distribution (Task 2, group 8) ── + minLength*: int ## shortest non-empty clause + medianLength*: float ## p50 of the non-empty length distribution + p10Length*: float + p90Length*: float + stdLength*: float + posMeanLength*: float ## positive-polarity clauses only + negMeanLength*: float ## negative-polarity clauses only + perClassMeanLength*: seq[float] + perClassMedianLength*: seq[float] + perClassHist*: seq[seq[int]] + emptyFraction*: float ## emptyClauses / totalClauses + +proc percentile*(sorted: openArray[int], p: float): float = + ## Linear-interpolation percentile (numpy default) over an ASCENDING array. + ## Empty -> 0.0; p is clamped to [0,1]. + if sorted.len == 0: return 0.0 + let pp = clamp(p, 0.0, 1.0) + let rank = pp * float(sorted.len - 1) + let lo = int(floor(rank)) + let hi = int(ceil(rank)) + if lo == hi: return sorted[lo].float + sorted[lo].float + (rank - float(lo)) * (sorted[hi] - sorted[lo]).float + +type + LengthStats* = object + n*: int + minL*: int + maxL*: int + mean*: float + median*: float + p10*: float + p90*: float + stdDev*: float + hist*: seq[int] + +proc lengthStats*(lengths: openArray[int]): LengthStats = + ## Full distribution of clause lengths: n, min/max, mean, median, p10/p90, + ## standard deviation and a histogram (index = length). + result.n = lengths.len + if result.n == 0: + result.minL = 0 + result.maxL = 0 + return + var sorted = newSeq[int](lengths.len) + for i, l in lengths: sorted[i] = l + sorted.sort() + result.minL = sorted[0] + result.maxL = sorted[^1] + var sum = 0 + for l in sorted: sum += l + result.mean = sum.float / result.n.float + result.median = percentile(sorted, 0.5) + result.p10 = percentile(sorted, 0.10) + result.p90 = percentile(sorted, 0.90) + var varSum = 0.0 + for l in sorted: + let d = l.float - result.mean + varSum += d * d + result.stdDev = sqrt(varSum / result.n.float) + result.hist = newSeq[int](result.maxL + 1) + for l in sorted: inc result.hist[l] proc clauseInfo*(m: TmMachine, samples: openArray[DiagSample], spec: FeatureSpec): seq[ClauseInfo] = @@ -200,14 +266,31 @@ proc clauseInfo*(m: TmMachine, samples: openArray[DiagSample], proc clauseSummary*(infos: openArray[ClauseInfo]): ClauseSummary = var maxLen = 0 var lenSum = 0 + var posSum = 0 + var posN = 0 + var negSum = 0 + var negN = 0 + var lengths: seq[int] + var posLengths, negLengths: seq[int] + var nClasses = 0 + for inf in infos: + if inf.cls + 1 > nClasses: nClasses = inf.cls + 1 + var classLengths = newSeq[seq[int]](nClasses) for inf in infos: inc result.totalClauses + if inf.cls >= 0 and inf.cls < nClasses: + if inf.length > 0: classLengths[inf.cls].add inf.length if inf.length == 0: inc result.emptyClauses else: inc result.nonEmpty lenSum += inf.length + lengths.add inf.length if inf.length > maxLen: maxLen = inf.length + if inf.polarity > 0: + posSum += inf.length; inc posN; posLengths.add inf.length + else: + negSum += inf.length; inc negN; negLengths.add inf.length if inf.votes > 0: inc result.firedAtLeastOnce if inf.polarity > 0: inc result.posFired @@ -220,6 +303,26 @@ proc clauseSummary*(infos: openArray[ClauseInfo]): ClauseSummary = if inf.length > 0: inc result.lengthHist[inf.length] result.meanLength = if result.nonEmpty > 0: lenSum.float / result.nonEmpty.float else: 0.0 + let stats = lengthStats(lengths) + result.minLength = stats.minL + result.medianLength = stats.median + result.p10Length = stats.p10 + result.p90Length = stats.p90 + result.stdLength = stats.stdDev + result.posMeanLength = if posN > 0: posSum.float / posN.float else: 0.0 + result.negMeanLength = if negN > 0: negSum.float / negN.float else: 0.0 + result.emptyFraction = + if result.totalClauses > 0: + result.emptyClauses.float / result.totalClauses.float + else: 0.0 + result.perClassMeanLength = newSeq[float](nClasses) + result.perClassMedianLength = newSeq[float](nClasses) + result.perClassHist = newSeq[seq[int]](nClasses) + for c in 0.. healthyHi -> too long + ## mean < collapsedMax -> collapsed + ## collapsedMax <= mean < healthyLo -> short (below band, not yet collapsed) + ## healthyLo <= mean <= healthyHi -> healthy + if summ.nonEmpty == 0: return "n/a" + let mean = summ.meanLength + if mean > healthyHi: "too long" + elif mean < collapsedMax: "collapsed" + elif mean < healthyLo: "short" + else: "healthy" + +proc lengthHistText*(hist: openArray[int]): string = + ## One line per length, e.g. `len 3: 12 ####`. + var mx = 0 + for c in hist: + if c > mx: mx = c + for l in 0..3}: {hist[l]:>6} {bar}\n" + +type + BlockLengthContribution* = object + name*: string + first*: int + count*: int + totalLits*: int ## literals from this block across the clauses + meanPerClause*: float ## totalLits / number of counted clauses + share*: float ## totalLits / all literals across counted clauses + clausesUsing*: int ## clauses with >= 1 literal from this block + perClauseMax*: int ## most literals this block contributes to one clause + +proc blockLengthContributions*(m: TmMachine, spec: FeatureSpec, + skipEmpty = true): seq[BlockLengthContribution] = + ## Per input BLOCK (walls / us / motion / bullets from the `FeatureSpec`): how + ## many literals does it contribute on average across clauses? + ## + ## A block that contributes many literals to EVERY clause is dominating the + ## model; a block contributing none is dead (the block-level view of the dead + ## INPUT list). `skipEmpty = true` (default) averages over non-empty clauses + ## only, so empty clauses do not dilute the shares. + result = newSeq[BlockLengthContribution](spec.blocks.len) + for i, b in spec.blocks: + result[i].name = b.name + result[i].first = b.first + result[i].count = b.count + var nClauses = 0 + var total = 0 + for c in 0..= 0: + inc perBlock[bi] + inc total + for i in 0.. 0: inc result[i].clausesUsing + if perBlock[i] > result[i].perClauseMax: + result[i].perClauseMax = perBlock[i] + for i in 0.. 0: result[i].totalLits.float / nClauses.float else: 0.0 + result[i].share = if total > 0: result[i].totalLits.float / total.float else: 0.0 + +type + ClauseCoverage* = object + samples*: int + totalClauses*: int + meanFiringClauses*: float ## mean non-empty firing clauses per sample + meanFiringFraction*: float ## ... / totalClauses + effectiveClauses*: float ## participation ratio of the vote counts + top3Share*: float ## share of all fires from the 3 busiest clauses + firingClauseCount*: int ## clauses that fire at least once + +proc clauseCoverage*(m: TmMachine, samples: openArray[DiagSample]): + ClauseCoverage = + ## How many clauses are actually needed to cover the cases? + ## + ## `meanFiringFraction` is the fraction of clauses that fire on any given + ## sample. `effectiveClauses` is the participation ratio + ## `(sum v)^2 / sum v^2` of the per-clause fire counts: if only 3 clauses cast + ## most of the votes, it reads ~3 regardless of the configured count. + result.totalClauses = m.nClasses * m.nClauses + result.samples = samples.len + if samples.len == 0 or result.totalClauses == 0: return + var counts = newSeq[int](result.totalClauses) + var fireSum = 0.0 + for s in samples: + var firing = 0 + for c in 0.. 0: inc result.firingClauseCount + totalVotes += v + sqVotes += v.float * v.float + if sqVotes > 0.0: + result.effectiveClauses = totalVotes.float * totalVotes.float / sqVotes + var sorted = counts + sorted.sort(SortOrder.Descending) + let k = min(3, sorted.len) + var top = 0 + for i in 0.. 0: top.float / totalVotes.float else: 0.0 + +type + ClauseShapeDiag* = object + summary*: ClauseSummary + verdict*: string + blocks*: seq[BlockLengthContribution] + coverage*: ClauseCoverage + healthyLo*: float + healthyHi*: float + collapsedMax*: float + +proc clauseShapeDiagnostics*(m: TmMachine, samples: openArray[DiagSample], + spec: FeatureSpec, healthyLo = 3.0, + healthyHi = 8.0, collapsedMax = 2.0, + skipEmpty = true): ClauseShapeDiag = + ## The one-call clause-shape readout: length distribution + healthy-band + ## verdict + per-block length contribution + coverage/concentration. + result.summary = clauseSummary(clauseInfo(m, samples, spec)) + result.healthyLo = healthyLo + result.healthyHi = healthyHi + result.collapsedMax = collapsedMax + result.verdict = clauseShapeVerdict(result.summary, healthyLo, healthyHi, + collapsedMax) + result.blocks = blockLengthContributions(m, spec, skipEmpty) + result.coverage = clauseCoverage(m, samples) + +proc shapeLine*(d: ClauseShapeDiag): string = + ## One-line clause-shape summary, in the same style as `healthLine`. + let s = d.summary + &"shape mean={s.meanLength:.2f} med={s.medianLength:.1f} " & + &"p10={s.p10Length:.1f} p90={s.p90Length:.1f} max={s.maxLength} " & + &"({d.verdict}) | coverage={d.coverage.meanFiringFraction*100:.1f}%/sample " & + &"eff={d.coverage.effectiveClauses:.1f}/{d.coverage.totalClauses} " & + &"top3={d.coverage.top3Share*100:.0f}%" + +proc formatClauseShapeReport*(d: ClauseShapeDiag): string = + ## Full clause-shape readout: distribution, per-polarity / per-class means, + ## per-block contribution and coverage. + let s = d.summary + result.add &"shape: mean={s.meanLength:.2f} median={s.medianLength:.2f} " & + &"p10={s.p10Length:.2f} p90={s.p90Length:.2f} std={s.stdLength:.2f} " & + &"min={s.minLength} max={s.maxLength} nonEmpty={s.nonEmpty} " & + &"empty={s.emptyClauses}/{s.totalClauses} " & + &"({d.verdict}, healthy band {d.healthyLo:.0f}-{d.healthyHi:.0f})\n" + result.add &" polarity: posMean={s.posMeanLength:.2f} negMean={s.negMeanLength:.2f}\n" + result.add " per-class mean/median:" + for c in 0.. 0: + result.add &" | shape={d.shape.summary.meanLength:.2f} ({d.shape.verdict})" proc automataDiagnostics*(tmpl: TmMachine, samples: openArray[DiagSample], spec: FeatureSpec, epochs = 1, seed = 777'u64, @@ -1092,6 +1395,7 @@ proc automataDiagnostics*(tmpl: TmMachine, samples: openArray[DiagSample], result.histogram = stateHistogram(result.machine, nHistBins) result.inputConfidence = perInputConfidence(result.machine, spec, samples, settleThreshold) + result.shape = clauseShapeDiagnostics(result.machine, samples, spec) result.summary = healthLine(result) proc automataVerdict*(d: AutomataDiag): string = @@ -1139,3 +1443,5 @@ proc formatAutomataReport*(d: AutomataDiag): string = for c in 0.. 0: + result.add formatClauseShapeReport(d.shape)