[GOLD-RING] compiler: compound assignment and hoisted nested fn - #2217
Merged
gHashTag merged 640 commits intoAug 19, 2026
Conversation
…E the run (Refs #1959) Nine waves improved one answer on one dataset and hit its ceiling. T354 asserted the limit is "the capacity of six-input truth tables on UNSW-NB15" -- but that sentence contains a dataset, and nothing has ever varied it. THE TEST: run the same architecture family on three tasks and measure the GAP between a dense reference and the sparse truth-table network on each. A constant gap is a property of the architecture; a moving gap is a property of the task. Absolute ceilings differ by task and tell us nothing alone; the gap is the invariant, and it has never been computed once. FORECAST REGISTERED BEFORE THE RUN: (1) the gap is LARGER on MNIST and Fashion than on UNSW, because UNSW 593 binary features are mostly irrelevant (T306) so a 6-of-593 draw loses little, while MNIST 784 pixels are informative almost everywhere. Predicted gaps: UNSW 5-7 pp, MNIST 10-18, Fashion 8-15. (2) Therefore T354 phrasing is too strong and needs narrowing to "on tasks whose features are individually weak". If the gap is CONSTANT within 2 pp, T354 stands as written. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
…he architecture's (Refs #1959) T354 IS REFUTED. It asserted the limit is "the capacity of six-input truth tables on UNSW-NB15" -- a sentence containing a dataset that had never been varied. Same architecture, trainer, quantiser, normalisation and balancing; only the task changes: UNSW-NB15 base 55.06 dense 89.67 sparse 82.97 penalty +6.70 MNIST-bin base 51.39 dense 96.01 sparse 81.16 penalty +14.85 Fashion-bin base 50.00 dense 91.52 sparse 88.03 penalty +3.48 Spread 11.37 pp. The cost of collapsing a layer into six-input truth tables is NOT a constant of the architecture; it depends on how much a task's decision needs many features at once. "Six-input truth tables have a ceiling" is true; "the ceiling is 84%" was a fact about UNSW wearing an architecture's clothes (T355). Two of three forecast bands landed and the miss is the interesting one: UNSW 5-7 (6.70 correct), MNIST 10-18 (14.85 correct), Fashion 8-15 (3.48 WRONG). The reasoning -- informative pixels punish a 6-of-784 draw -- holds for MNIST and INVERTS for Fashion, whose binarised silhouettes are decided by a few coarse pixels. Feature count and modality do not predict the penalty; evidence concentration does (T355a). WE HAVE BEEN BENCHMARKING ON OUR SECOND-WORST TASK FOR NINE WAVES. Sparse accuracy is 88.03% on Fashion against 82.97% on UNSW, at half the penalty. We followed the field's benchmark triplet without asking whether it suits this datapath. It does not (T355b). THE FIRST LAYER'S NARROW VIEW IS NOT THE BOTTLENECK. Registered: fan-in 12 buys 1.5-3.0 pp, and under 1 pp refutes the "six-of-593" framing. Measured +0.83 pp for 832 extra LUT -- a thousand LUT per point against width's 393. Both bad, fan-in 2.5x worse. The pre-registered consequence applies: three waves of reasoning that started from that image were about the wrong thing (T356). Note also fan-in 10 scoring below fan-in 8 at three seeds -- an intervention below our own noise floor is not a lever whatever its cost (T356a). Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
838: if a conclusion names a dataset, the dataset is an untested variable -- T354 said "on UNSW-NB15" and the second half was load-bearing; varying it moved the penalty from +3.48 to +14.85 pp. 839: measure the GAP, not the ceiling -- dense-minus-sparse is the invariant that answers "architecture or task?" and had never been computed once. 840: nine waves of benchmarking on an unexamined choice; a benchmark inherited from a neighbouring literature is a hypothesis about your system, not a given. 841: an effect below your noise floor is not a lever whatever it costs. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
… the run (Refs #1959) W758 explained an 11-point spread in the sparse penalty with "how concentrated a task evidence is". That is a WORD, and three points would fit any word one cares to propose. This file uses ELEVEN tasks: the three originals plus eight MNIST digit-pairs sharing input dimension, trainer, seeds and subsample, so nothing in the setup is confounded with the quantity under test. THE MEASURE, defined BEFORE any correlation: C6 = (sum of the 6 largest per-feature mutual informations) / (sum of all) -- what fraction of the total single-feature evidence a SIX-input neuron could see if it chose perfectly. Six because that is the fan-in the six-bit rule fixes. No free parameters, nothing to tune after the fact. FORECAST REGISTERED BEFORE THE RUN: (1) C6 correlates NEGATIVELY with the penalty, Pearson r <= -0.7 over 11 tasks; (2) the three originals order Fashion > UNSW > MNIST by C6, mirroring penalties 3.48 < 6.70 < 14.85; (3) the relation is monotone but not linear. If r is weaker than -0.5, concentration is not the mechanism and W758 explanation must be withdrawn to a bare observation. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
…eleven tasks (Refs #1959) W758 explained an 11-point spread in the sparse penalty with "how concentrated a task's evidence is". W759 turned that word into a measure defined BEFORE any correlation -- C6 = (sum of the 6 largest per-feature mutual informations) / (sum of all) -- and tested it on ELEVEN tasks: the three originals plus eight MNIST digit-pairs sharing dimension, trainer, seeds and subsample. r(C6, penalty) = +0.128, t = +0.39 at n = 11. Not merely weaker than the registered -0.7 threshold: the WRONG SIGN and indistinguishable from zero. The pre-registered consequence applies verbatim -- W758's explanation is withdrawn to a bare observation that the penalty varies between +0.28 and +14.85 pp and we do not know why (T358). Forecast (2) was refuted before the correlation was computed: I predicted Fashion > UNSW > MNIST by C6 and measured UNSW 0.146 > MNIST 0.081 > Fashion 0.042 -- Fashion has the LOWEST concentration and the LOWEST penalty, the exact inverse of the story. The mechanism was contradicted by its own three founding points (T358a). Nothing else measured predicts it either: dense accuracy r=-0.398, headroom r=+0.398, neither significant (T358b). THE OUTLIER THE DIGIT PAIRS EXPOSE: every MNIST digit pair costs +0.28 to +5.66, while MNIST-bin -- the same pixels, a different labelling -- costs +14.85. The architecture is not bad at "MNIST"; it is bad at that particular binary partition of it. A task's identity is its labelling, not its input distribution (T359). The penalty ranges over a factor of FIFTY across eleven tasks and the project has no predictor for it. Any claim that this datapath suits task X must be MEASURED on X, not argued from X's properties -- a weaker statement than W758 made, and the one the data supports (T359a). Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
…1959) THE SAME SILICON, FIVE AND A HALF POINTS HIGHER. W757s Pareto front rebuilt on Fashion-bin with identical architectures and therefore identical area: H16 L2 at 123 LUT gives 86.91% against UNSWs 81.37%, and the whole curve sits 5.1-5.9 points higher. The front is flat on Fashion too -- 6.5x the area for 1.38 pp -- so T352a shape conclusion survives the change of task while the level does not. Best single-die artefact this programme has: 123 LUT at 86.91%, obtained by changing the benchmark rather than anything in the datapath (T360). It does NOT license claiming the datapath "suits" Fashion -- T359a withdrew that kind of inference. The honest statement is "measured higher on Fashion, cause unknown" (T360a). Lessons 842-845: turn explanatory words into measures and test on more points than inspired them; define the measure and its refutation threshold before computing the correlation; a tasks identity is its LABELLING not its inputs; changing the benchmark moved more than nine waves of tuning. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
…run (Refs #1959) W759 tested one pre-registered predictor on eleven tasks and it failed at r=+0.128. The obvious next move -- try more predictors -- is exactly how a programme fools itself: ten candidates on eleven points finds one by chance. THE DESIGN THAT PREVENTS THAT: sixty tasks, all LABELLINGS OF ONE DATASET (45 digit pairs, 10 one-vs-rest, 5 groupings), so the input distribution is identical across all sixty and only the decision changes -- T359s observation turned into the experiments design. The sixty split into DISCOVERY (30) and CONFIRMATION (30) by a fixed seed before anything is measured. Eight candidates are ranked on DISCOVERY only; the single best is tested ONCE on CONFIRMATION, and that number is the result. FORECAST REGISTERED BEFORE THE RUN: at least one candidate reaches |r| >= 0.5 on DISCOVERY -- with eight candidates on thirty points that is nearly certain by chance. Predicted: the winner does NOT replicate, landing |r| < 0.3 on CONFIRMATION, and the honest conclusion stays "no predictor known". I expect `head` to win discovery and shrink on confirmation. If a candidate holds |r| >= 0.5 on BOTH halves, the project has its first real rule for choosing tasks. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
…asurements (Refs #1959) The file kept the right posture -- only reproducible numbers, omit doubtful rows -- and was six days stale, predating everything measured since W746. Now records: the single-die artefact (123 LUT at 81.37% UNSW / 86.91% Fashion, 99.46 MHz, 1 cycle, output stage included); the three-die trained network (232 LUT, zero DSP, 100/100 agreement on real rows); the measured cost rules (six bits per neuron, ternary inputs take fan-in 3, area is not a lever, the golden resolve costs 8 DSP or 2750 LUT); and both openXC7 primitive defects with their flags. Section 4.5 states what is NOT claimed: no LUT*ns comparison (the fields Fmax is not in our record and fetching failed), no suitability claim for any task (no predictor survived confirmation), and no accuracy from before W749 (uncontrolled pre-activation scale). Also records the correction that every LUT figure before W752 omitted the decision neuron. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
…umes (Refs #1959) The first run died at 51 of 60 on the alarm timeout and lost ALL fifty-one results, because json.dump was at the end. A long sweep that only persists on success persists nothing. Now writes after every task and skips tasks already on disk, so a restart resumes. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
…ims (Refs #1959) Section 4 said the single-cell figure could not be compared to network-level published figures. That gap is now closed on one benchmark and the answer is not flattering: 123 LUT at 81.37% against TreeLUT 89 LUT at 92.0% -- 1.38x the area at 10.6 points less accuracy. Our area is competitive; our accuracy is not. THE GOLDEN ALPHABET MEASURED AGAINST ITS OWN CLAIM. The line is named for phi. Alphabet size is worth +0.735 pp, shape +0.149 pp significant on 1 of 3 tasks, against inter-layer normalisation at +29.15 pp -- and the multiplier phi removes from weight application returns in the pair resolve, costing 8 DSP48E1 or ~2750 LUT. A negative result about our own headline idea, stated as one. The algebra stands; the practical advantage does not. FOUR NEW NON-CLAIMS: no claim the golden alphabet is preferable in hardware; no LUT*ns comparison with any published system; no claim this datapath "suits" any task (no predictor survived a confirmation split); no accuracy from before W749 is comparable. Also records both openXC7 primitive defects, and the three-die trained network (232 LUT, zero DSP, 100/100 agreement on real rows). Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Added: a trained ternary network across three dice (232 LUT, zero DSP, 100/100 agreement on real rows); the trainer-to-silicon export path that made it possible; a known-bad-primitive guard in t27c yostat; a 33-bit data register lifting the 31-bit payload limit; the openXC7 SRL16E defect report. Fixed: openXC7 emits a wrong bitstream for SRL16E while the netlist is correct (0/6 rows against 24/24 with -nosrl, source byte-identical); and every LUT figure before W752 omitted the decision neuron. Changed -- claims narrowed: the golden alphabet measures +0.735 pp against normalisation +29.15 pp and its pair resolve reintroduces a multiplier; no accuracy before W749 is comparable; no suitability claim for any task; no LUT*ns comparison. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
) Adds a compact, measured FPGA section high in the README: the three-die trained network (232 LUT, 0 DSP, 100/100 agreement), the best single-die artefact (123 LUT at 81.37% UNSW / 86.91% Fashion) beside the field reference (TreeLUT 89 LUT at 92.0%), and the measured six-bit cost rule. States plainly that our area is competitive and our accuracy is not, and that the golden alphabet -- the line is named for phi -- measures +0.735 pp against normalisation at +29.15 pp, with the multiplier returning in the pair resolve. The algebra stands; the practical advantage does not. Also surfaces both openXC7 primitive defects on the front page, since they produce a wrong bitstream from a correct netlist and pass the acceptance criterion. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
#1959, Refs #2173) TOTAL MUTUAL INFORMATION PREDICTS THE SPARSE PENALTY. W759 tested one pre-registered predictor on eleven tasks and it failed. W760/W761 ran labellings of ONE dataset -- MNIST digit pairs, one-vs-rest and groupings, so the input distribution is identical and only the decision changes -- split into discovery and confirmation halves by a fixed seed before anything was measured. mi_tot scored -0.761 on discovery and -0.881 on confirmation. The registered forecast predicted the discovery winner would NOT replicate (|r| < 0.3); it is refuted, and for once in our favour. The more total single-feature evidence a labelling carries, the LESS a six-of-N draw costs (T361). Why W759 saw nothing: c6 measures CONCENTRATION, mi_tot measures AMOUNT. The quantity that matters is how much evidence exists, not how unevenly it is distributed -- the opposite of the story W758 told (T361a). Stated narrowly: this licenses prediction ACROSS LABELLINGS OF ONE DATASET, which is what was measured. It does not yet license prediction across datasets (T361b). THE SWEEP THAT LOST FIFTY-ONE RESULTS: the first census died at 51 of 60 on its timeout and lost all of them, because json.dump ran only after the loop. A long sweep that persists only on success persists nothing. Now writes after every task and resumes (T362). Also: filed the openXC7 SRL16E defect as t27#2173. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
…s (Refs #1959) T361b stated that mi_tot success across labellings of ONE dataset did not license prediction across datasets. W759s eleven-task set spans three datasets and its penalties were already measured; only the predictor needed computing. Registered forecast: (1) raw mi_tot fails across datasets, |r| < 0.5; (2) the normalised version does better; (3) dense travels better than either. Measured: mi_tot raw r = -0.684, t = -2.82, SIGNIFICANT -- (1) refuted. mi_tot/n_features r = -0.676, slightly WORSE -- (2) refuted, so the quantity that matters is TOTAL evidence, not evidence per feature. dense r = -0.398, the weakest and not significant -- (3) refuted hardest. One of three predictions survived in direction and none in detail (T363). MECHANISM, both scopes now measured: the more total single-feature evidence a task carries, the less a narrow random view costs -- redundancy is what makes sparse sampling survivable. Within one dataset r = -0.78/-0.88 over 45 labellings; across three datasets r = -0.684 over 11 tasks. T361b is relaxed (T363a). This is the programmes most useful positive result and it is not about arithmetic: a task-selection rule computable in one pass, telling you before you build anything what a six-input truth-table datapath will cost in accuracy. MNIST-bin has the least evidence (3.14) and the worst penalty (+14.85); 0v1 has the most (44.75) and the least (+0.28) (T363b). Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
…1959) Dmitrii has asked "which format for ternary on FPGA" four times, which means the answer is not where it is looked for. The format skill now opens with it: Weights: a TRUTH TABLE, not a format, at <=6 input bits per neuron. Where a table will not fit: {0,+-1,+-2,+-4,+-8}. Accumulator: int12-int16. phi does not belong in the datapath. With the measurements behind each clause: 2.00 LUT/neuron at <=6 bits and no arithmetic at all (dense 54,914 LUT -> 128 as tables, 429x, 3.6x Fmax); the SIX-BIT RULE with the ternary-is-two-bits corollary that nearly shipped a 10,250-LUT design as an 800-LUT one; PoT with its real citations and the warning that `pot9` is an internal tag; the phi table showing +0.735 pp against a resolve costing 8 DSP or 2750 LUT; the intervention ranking that puts the alphabet seventh; mi_tot as the task-selection rule; and the mandatory -nodsp -nosrl flags. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
… own headline (Refs #1959) The full 60-labelling census put ntrain at r = +0.745/+0.730, almost matching the winner -- a warning, not a finding, because the census mixes digit pairs (~10,800 rows, mean penalty +2.36) with one-vs-rest tasks (~54,000 rows, +3.42). Any predictor correlated with task TYPE would replicate beautifully while measuring nothing. Restricted to the 45 digit pairs, where ntrain varies by 1.15x: mi_tot -0.810 (t=-9.07), mi_max -0.751, dense -0.734, ntrain -0.382. NTRAIN SIGN FLIPS -- that inversion IS the confound, made visible -- and mi_tot is STRONGER inside the clean group than across the mixed one, which is what a real predictor does (T364). The replication discipline was necessary but not sufficient: a discovery/ confirmation split protects against fitting noise and not at all against a confound present in both halves (T364a). Final rule with both controls applied: mi_tot predicts the sparse penalty at r = -0.81 within a homogeneous group of 45, -0.72 across a mixed 60, -0.68 across three datasets, over penalties spanning +0.24 to +14.71 pp. The only predictor this programme has that survived both a confirmation split and a confound check (T364b). Lessons 846-849. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
…ts (Refs #1959) Dmitrii asked for a measured top of numbers for ternary weights from our catalogues. Distinction stated first: the 83-format catalogue enumerates FLOAT ENCODINGS (bit widths, exponent/mantissa), not weight VALUES, so the candidate list is built from the projects own theorems and the quantisation literature. Alphabet shape held fixed at NINE levels for every base -- A(b) = {0} u {+-b^k, 0<=k<=3} -- because T286 showed size is worth +0.844 pp and shape +0.085, and confounding them is how a base wins for the wrong reason. Candidates: 1.0 (degenerate control), 2.0 (dyadic/PoT, the incumbent), phi (golden, our line, the only multiplier-free degree-2 scale), plastic (r^3=r+1), psi4 (r^4=r+1), supergolden, tribonacci, sqrt2 (algebraic but NOT in the r^d=r+1 family), silver (1+sqrt2), 3.0 (TERNARY BASE -- never measured, an omission worth naming), e (transcendental control), and linear spacing. FORECAST REGISTERED BEFORE THE RUN: (1) all bases within 1.0 pp except 1.0 which loses by 2-6; (2) area ordering 2.0 < 3.0 ~ lin << every irrational; (3) the top is decided by AREA not accuracy and 2.0 wins. If any base beats 2.0 by more than 1.0 pp in accuracy, the shape effect is larger than three waves say and T286/T317 need revisiting. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
…Refs #1959) Dmitrii asked for a measured top of numbers for ternary weights from our catalogues. The question exposed a gap: the 83-format catalogue enumerates FLOAT ENCODINGS -- bit widths, exponent/mantissa splits -- and contains no answer about what VALUES a weight level should take. That catalogue cannot answer this question and nobody had noticed. Eleven bases, NINE LEVELS EACH (shape fixed so only the number varies), accuracy on UNSW-NB15 with the corrected bench and area as a placed 64->8 layer: dyadic 2.0 89.62% 752 LUT linear -- 812 ternary 3.0 89.76% 1417 sqrt2 89.64% 1991 plastic 89.66% 2709 golden phi 90.01% 2726 e -- 2856 silver 89.80% 2971 b=1.0 (={0,+-1}) 89.52% -- THE WHOLE ACCURACY SPREAD ACROSS EVERY BASE IS 0.49 pp; the area spread is 3.95x. phi is nominally first (+0.39 pp) and costs 1,974 extra LUT for it -- 5,062 LUT per point. The top is decided by area and dyadic wins it, as forecast (T365). The forecast's second half is refuted and is the more interesting half: b=1.0, which collapses to {0,+-1}, was predicted to lose by 2-6 pp and loses by 0.49. A three-level alphabet is within half a point of a nine-level golden one on this bench (T365a). BASE 3 HAD NEVER BEEN MEASURED, IN A TERNARY PROJECT: 89.76% at 1,417 LUT, second-cheapest geometric base and better than dyadic on accuracy. The obvious candidate was absent from every prior sweep, which is what happens when the catalogue consulted is about float encodings (T365b). Ordered by what a designer pays: dyadic -> linear -> base 3 -> everything irrational. Accuracy does not order them. The number is an area decision and the literature's answer -- powers of two -- is right (T365c). Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
…(Refs #1959) W763 measured the base cost in a DENSE adder tree: 3.95x spread (dyadic 752, base 3 1417, phi 2726, silver 2971), because that layer performs real arithmetic. The TRUTH-TABLE layer performs none -- the neuron is enumerated on the host and emitted as a case statement. Registered forecast: every base costs 2.00 LUT/neuron within +-0.15, spread collapses to 1.0x. REFUTED in the interesting direction: the spread collapses to 1.9x, and the ORDERING INVERTS. Base 3 costs 1.05 LUT/neuron, 46% LESS than dyadic, and for a project whose premise is ternary that is the first measurement in which the ternary base wins anything (T366). Why it inverts: a table cost is set by how many input patterns disagree with the most common output, since yosys folds the rest into the default arm. Fast-growing bases (3, e, silver) push the sum past the threshold on most patterns so one symbol dominates and the table compresses; slow-growing bases (phi, sqrt2, plastic) leave the sum near zero, the symbols mix evenly, and nothing folds. Cheapness in a table is table SKEW (T366a). T365c scope corrected: "the number is an area decision" holds for adder-tree datapaths (3.95x) and points the OTHER WAY in table datapaths (1.9x). Neither statement is about accuracy, which varies 0.49 pp across every base (T366b). Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
… two trits (Refs #1959) Dmitrii corrected the framing: every metric and top must be measured for the TERNARY world. Everything measured so far is counted in LUT -- a BINARY primitive of the substrate we happen to own, not of the target. Recounted in trits: the alphabet sizes that pack with ZERO WASTE in ternary storage are exactly the powers of three -- 3, 9, 27. The Nine-Rung Law puts the ceiling at NINE, and nine is 3^2, EXACTLY TWO TRITS. The two sizes this programme singled out by measurement -- the minimum useful alphabet (3) and the ceiling (9) -- are both perfect in trits and both wasteful in bits, 9 levels burning 7 of 16 binary codes (T367, arithmetic not measurement). EVERY AREA FIGURE IN THIS FILE IS IN THE WRONG UNITS FOR THE TARGET. LUT counts measure a binary FPGA: correct for this substrate, wrong for a ternary one where the natural unit is the trit-cell and a 6-input binary LUT has no counterpart. T340, T352 and T366 are all substrate measurements and none has been restated in ternary units, because no ternary substrate exists to measure on. Named rather than papered over (T367a). What the 0.49 pp result actually says: a THREE-LEVEL alphabet -- {0,+-1}, ONE TRIT, the founding premise -- scores within 0.49 pp of a nine-level golden one and costs 114 LUT against 123-128 in tables. Half the weight storage for half a point; pooled over three tasks the size effect is +0.735 pp, so the honest range is half to three-quarters of a point for one trit versus two. A strong position for a ternary line, under-reported in W763 as "the interesting half of a refutation" (T367b). Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
) T367 recounted alphabet packing in trits. Pushed further the reframing reaches three more places, two of which restate measurements the programme already has. RADIX ECONOMY (classical, Knuth TAOCP v2): E = b*log_b(N). At N=10^6, base 3 scores 37.73 against binarys 39.86 -- base 3 is the nearest integer to e and the most economical integer radix, beating binary by 5.7%. Textbook, not ours, but it is the substrate this line targets and no measurement in the file had been stated against it (T368a). THE SIX-BIT RULE IS THE THREE-TRIT RULE. T331 measured 2.00 LUT/neuron at <=6 input bits. A ternary input is 2 bits, so fan-in 3 is 6 bits is EXACTLY 3 TRITS, and a natural three-trit table is 3^3 = 27 entries held in a binary LUT6 of 64 -- 42% used. 58% of every LUT in the datapath is wasted by the SUBSTRATE, not the design (T368b). WEIGHT STORAGE: for the two sizes that matter, 3 and 9 levels, binary wastes 21% and ternary wastes nothing. On the trained network that is 39,040 bits binary against 19,520 trits = 30,938 bit-equivalents, 21% less unconditionally (T368c). AN ARITHMETIC SLIP CAUGHT IN OUR OWN OUTPUT: the first printing carried a hand-written summary reading "37%, 21%, 37%" while the computed column read 21%, 21%, 5%. The summary was a hardcoded string. A number typed beside a computed number is not a check on it -- same class as reading a line count for a quantity, sixth disguise (T368d). Establishes that the founding choice is optimal IN ITS OWN UNITS. Establishes nothing about accuracy or the 10.6-point gap. The ternary case is an efficiency case and a real one; it is not an accuracy case and must never be presented as one (T368e). Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Saves Dmitriis standing correction to the format skill: measure in TRITS, not LUT. Radix economy (base 3 optimal among integers, 5.7% over binary), alphabet packing (powers of three waste nothing; the Nine-Rung ceiling is 9 = 3^2 = two trits), the six-bit rule restated as the THREE-TRIT rule (3^3 = 27 entries in a 64-entry binary LUT, 42% used), the 21% storage tax, the inverted base ordering in tables, and the plain answer that three levels IS a strong position -- one trit, within 0.49 pp of nine golden levels, 114 LUT against 123-128. With the boundary stated: all of it is an EFFICIENCY case and none of it is an accuracy case; the 10.6-point gap to TreeLUT is untouched. Lessons 850-853: measure in the targets units; say when a catalogue cannot answer the question; the same variable can order two architectures oppositely; a number typed beside a computed number is not a check on it. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
…hdrawn (Refs #1959) THE THIRD TRIT BUYS NOTHING. T367 established 3, 9 and 27 as the zero-waste ternary alphabet sizes; T288 put the ceiling at nine, but nine is where OUR ladder stopped and 27 = 3^3 is the next perfect packing. Measured: UNSW 9->27 +0.15 pp (t=0.65, ns); Fashion 9->27 -0.13 pp (t=-2.03, ns, NEGATIVE). "Two trits and no more" now holds as a measured statement in the targets own units (T369). AND THE 3->9 STEP IS SMALLER THAN THIS FILE HAS BEEN CLAIMING. T317 put the size effect at +0.735 pp pooled; here it is +0.10 (UNSW, ns) and +0.25 (Fashion, significant). The honest range for one trit versus two is 0.1-0.75 pp and the low end is not significant. The case for two trits is weaker than the file implied (T369a). "THE TERNARY TABLE WE DO NOT HAVE" WAS THE WRONG FRAMING. Dmitrii pushed back and asked what t27 was built for. SOUL.md: "T27 is a spec-first architecture where mathematical truth, not implementation, is the source of truth." The 27-entry ternary table is not a thing this project lacks -- it is exactly the thing the language exists to state. A .t27 spec names the object; the compiler emits what the substrate can hold, and today that emission costs 58% waste. THE WASTE IS A PROPERTY OF THE TARGET, NOT THE SPECIFICATION (T370). That 58% is therefore not a defect to apologise for but the measured cost of running a ternary specification on binary silicon -- the number that quantifies what a ternary substrate would return. It belongs in the argument for the Tiny Tapeout line, not in a list of shortcomings (T370a). Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
…1959) A tri blog post on the ternary recount, bilingual as the existing post is. Carries the picture (27 reachable entries in a 64-entry LUT6), the alphabet packing table (powers of three waste nothing), the six-bit rule restated as the three-trit rule, and the 27-level measurement that closes the ceiling at two trits. It also carries two things against us, at the same volume: the 3->9 step is +0.10 (ns) / +0.25 pp on this bench against the +0.735 recorded elsewhere, so we had been quoting the top of the range; and none of the ternary accounting touches the 10.6-point accuracy gap to TreeLUT. And the framing correction: the 27-entry table is not something we lack, it is what a spec-first language exists to state -- SOUL.md, "mathematical truth, not implementation, is the source of truth". The 58% is the measured cost of emitting a ternary spec onto binary silicon, which is the number that says what a ternary substrate would return. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
…ORE the run (Refs #1959) The theorem file contains TWO numbers for one quantity: T317 records the alphabet-size effect (3 -> 9 levels) at +0.735 pp pooled over three tasks, and W765 measured the same step on the same architecture at +0.10 (UNSW, ns) and +0.25 (Fashion). That is a defect regardless of which is right. The only difference between the benches is the training budget: alphabet_fixed.py runs 8 epochs, bases.py runs 30. Everything else matches. HYPOTHESIS: a larger alphabet is a CRUTCH FOR AN UNDER-TRAINED NETWORK. Given enough gradient steps the network compensates for a coarse alphabet, so the benefit of extra levels decays with budget. If true, the programmes ranking is not a list of independent effects -- the alphabets +0.735 pp was measured in a regime the training-budget fix has since removed. FORECAST REGISTERED BEFORE THE RUN: (1) the effect decays monotonically with epochs -- about +0.7 at 8, +0.15-0.3 at 30, under +0.1 and not significant at 60; (2) both recorded numbers are therefore correct measurements of DIFFERENT regimes, and T317 needs its budget stated rather than its value corrected; (3) the decay appears on both tasks, since the mechanism is optimisation not data. If the effect is FLAT in epochs, the benches differ in something I have not identified and the contradiction is unexplained -- worse than either number being wrong. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
…ibrate (Refs #1959) T363 relaxed T361b on r = -0.684 across three datasets. Fitting the relation on the homogeneous 45 digit pairs and testing on the three datasets NOT in the fit: penalty_pp = -2.2843 * ln(mi_tot) + 8.9025 (in-sample r=-0.839, RMSE 0.64 pp) Fashion mi 30.40 predicted 1.10 measured 3.48 error -2.38 UNSW mi 11.04 predicted 3.42 measured 6.70 error -3.28 MNIST mi 3.14 predicted 6.29 measured 14.85 error -8.56 MAE 4.74 pp out of sample against RMSE 0.64 in it, and ALL THREE ERRORS HAVE THE SAME SIGN -- the rule systematically under-predicts outside its fitting group. mi_tot ranks tasks correctly and calibrates only within a dataset (T371). THIS KILLS THE TOOL I WAS ABOUT TO SHIP. t27c taskfit was to print a predicted penalty from one pass over the data; on the three datasets we care about it would have printed 1.1, 3.4 and 6.3 where the truth is 3.5, 6.7 and 14.9 -- confident, wrong, and all optimistic. A predictor validated only by CORRELATION will happily produce a calibrated number nowhere near right, and correlation was the only validation T363 ran (T371a). What survives: mi_tot is a RANKING signal, r=-0.68 across datasets and -0.81 within one. Not a calibrated estimator and never to be printed as one; the honest tool prints an ordering and a bracket (T371b). Why the sign is always the same, as a hypothesis not a finding: the fit came from pairs spanning 0.24-5.55 pp and all held-out tasks lie above that range, so a logarithm fitted inside a narrow band under-shoots outside it by construction -- a property of extrapolation, not evidence about ternary datapaths (T371c). Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
…ine (Refs #1959) W765 declared the file self-contradictory: T317 put the size effect at +0.735 pp and W765 measured +0.10 / +0.25. W766 registered a mechanism -- decay with training budget -- and measured four budgets. THE FORECAST IS REFUTED: the effect is FLAT in budget (UNSW +0.512/+0.169/+0.328/+0.351, never significant at 8 seeds; Fashion +0.324/+0.403/+0.295, always significant). The registered consequence was to look elsewhere, and elsewhere was T317s own table, which reports PER TASK: UNSW +0.381, MNIST +1.478, Fashion +0.346, mean +0.735. W766 measures UNSW at ~+0.34 and Fashion at ~+0.34 -- agreeing with T317 on both. The +0.735 is a three-task mean lifted by an MNIST outlier and W765 compared it against two individual tasks. There was no contradiction in the file; there was a contradiction in how I read it (T372). The wasted experiment still established two new facts: the size effect is FLAT in training budget from 8 to 60 epochs -- not a crutch for under-training as the mechanism supposed -- and on UNSW it never reaches significance at 8 seeds, so T317s UNSW significance came from 30 seeds and not a larger effect (T372a). Lessons 854-857: check whether two numbers are the same quantity before chasing the discrepancy; correlation licenses a ranking, only out-of-sample error licenses a number; extrapolation outside a fitting range fails in a predictable direction; "the thing we do not have" is usually a substrate statement. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
…rate has a number (Refs #1959) THE EMPTY CELL IS FILLED. Fashion at 60 epochs: +0.200 pp, t=4.82, significant. Full grid: UNSW +0.512/+0.169/+0.328/+0.351, never significant at 8 seeds; Fashion +0.324/+0.403/+0.295/+0.200, always significant. Flat confirmed on both. The alphabet-size effect is not a crutch for under-training -- W766s mechanism is dead on both tasks -- and it is worth about a third of a point per task, which is what T317 said all along (T373). THE RANKER REFUSES TO PRINT A POINT ESTIMATE. taskfit.py orders tasks by mi_tot and reports a bracket from the two nearest measured anchors: Fashion 0.3-3.5 (true 3.48), UNSW 0.9-6.7 (true 6.70), MNIST 5.7-14.8 (true 14.85). Ordering right, every true value at the TOP edge -- the systematic direction T371c predicted. The tool states in its own output that a fitted curve under-predicted three held-out datasets by 2.4, 3.3 and 8.6 pp (T374). CIRCULARITY NAMED: those three datasets are themselves anchors, so this is a consistency check and not a validation (T374a). WHAT A TERNARY SUBSTRATE WOULD RETURN [СМОДЕЛИРОВАНО]. Unit: one configuration cell. One neuron: 128 bit-cells (only 54 reachable) against 27 trit-cells -- 4.74x fewer cells, 2.99x less configuration entropy. The trained networks hidden layers: 4,096 bit-cells against 864 trit-cells (T375). Three things it does NOT say, listed because the number is quotable and the caveats are not: nothing about speed, nothing about silicon area (a trit-cell costs more transistors and how many depends on circuit design we do not have), nothing about accuracy. It is an estimate of configuration memory and only that, and the adder-tree output layer is excluded (T375a). Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
858: a tool that cannot be calibrated should print an order and a range, not a number -- refusing a point estimate is a feature when the estimate would be wrong, and the alternative was one commit from shipping. 859: say "consistency check" when the test cases are the training cases, and name the circularity in the same breath as the result. 860: attach the non-claims to a quotable number in the same paragraph -- 4.74x will be repeated and its caveats will not unless they travel with it. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
…o golden sieve exists (Refs #1959) T374a named the circularity: the tools three demo datasets were themselves anchors. Twelve NON-ANCHOR labellings from the census, penalties measured in W760/W761 with no reference to the tool. Registered forecast: (1) ranking holds, rho <= -0.6; (2) bracket coverage under 60%, misses BELOW the lower edge; (3) order usable, range not yet. (1) CONFIRMED AND BEATEN: rho = -0.902, stronger on held-out tasks than the -0.81 measured inside the census. (2) half right: coverage is 67%, better than predicted, but all four misses are below the lower edge, exactly the predicted direction. The circularity is discharged -- mi_tot ranks tasks it has never seen (T376). THE BRACKETS DEFECT IS T364s CONFOUND AGAIN: every miss is a digit pair, and the anchor set mixes two populations -- whole datasets (3.5-14.9 pp) and digit pairs (0.2-5.6) -- occupying the same mi_tot range at different penalty levels. A bracket across both is too high for pairs by construction. The fix is separate anchor sets per task family, not a better curve (T376a). NOW LICENSED: mi_tot may ORDER candidate tasks (out-of-sample rho = -0.902) and bracket them WITHIN one task family. Not across families, and still no point estimate (T376b). NO GOLDEN SIEVE EXISTS HERE. git log --all --grep=sieve returns two commits, both PRIME sieves (Eratosthenes W498, Miller-Rabin W370); neither touches phi. The nearest relative is Zeckendorf representation in specs/numeric/lucas_accumulator .t27 -- every integer uniquely a sum of NON-CONSECUTIVE Fibonacci numbers, which is exactly base-phis forbidden 11. It is a representation, not a sieve, and has never been used as a weight encoding here (T377). Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Owner
Author
|
Ladder update — 0004a and 0005 are on the branch. The headline: the `and` clause never worked, and now it does.
Found by ddmin: an 80-line "contextual parser-state" repro reduced to four lines, and the context was one Recorded revision: 0003's per-clause skip was withdrawn by its own regressions (clause junk handed to module level errors hard where the old fallback skipped safely; four files regressed, all four recover). The and-fix keeps the win by making most blocks lower completely. Zero regressions at every shipped rung; cumulative patches in 🤖 Generated with Claude Code |
`parse-complete --show` printed "nothing discarded" for a file the corpus mode charged 2,438 tokens -- same binary, same file. The counter increments in three places; the span recorder lived in only one (`skip_to_next_top_level`). Skipped brace bodies (forall and friends) and statement-level recovery counted tokens and recorded nothing, so the W634 read-what-vanished tool was blind to two whole channels -- the mechanism behind W892's finding that the first inventory missed 311+ BDD lines in one file. Both remaining channels now push (line, lexeme) under the same 20,000-span cap. Verified: per-file span totals equal the corpus counter exactly across all 126 discarding files, 42,926 = 42,926, zero mismatches. Refs #1959 Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
…a "type"
The residual map blamed struct literals in clause values; exact-line probes
acquitted them (they parse). ddmin found the real killer three clauses later:
after a bracketed group, parse_array_literal consumed a following Ident
UNCONDITIONALLY as the Zig-style element type (`[3]u8`), across newlines --
so `and params = [1.0]` followed by `when result = ...` ate `when` as a
"type" and the whole block fell back. `and` clauses survived all along only
because KwAnd is not an Ident, which is exactly why 0005's fix made this look
like a struct-literal problem one clause earlier. Presence is not causality.
The rule now: an Ident after `]` is an element type only if it sits on the
same line as the `]` (the corpus's legacy same-line shapes, incl.
`&[_]provider-schema::ToolCall{}` in specs/provider/stream.t27, keep their
parse) or is immediately followed by an initialiser brace -- and in
clause-value mode a clause keyword (given/when/then/assert/and) is NEVER a
type, on either line.
Adversarially verified: a three-lens break panel ran 72 probe attempts against
the first version of this guard and found (1) one-line `given xs = [1, 2]
then ...` pairs still eaten by the same-line arm, and (2) a SILENT
false-green: `then {1} == xs` forged the brace test -- `then` became the
"type", `{1}` its initialiser, and the assertion vanished under "nothing
discarded". Both closed by the clause-keyword exclusion; the forged-brace
probe ships as a negative test (its honest outcome today is fallback, because
`{expr}` in expression position is the pre-existing W578 limitation, not this
rung). Two cross-line bare-type shapes (`[2]` newline `f64`) trade away their
parse; the corpus holds zero instances (0 new parse-fails).
Ladder, measured (624 non-scratch specs):
0005 42,926 discarded tokens
0006 37,786 (-12% this rung; -44% from the 67,760 base)
parse-fails 171 (zero new vs baseline; the SSOT pair stays fixed)
lost-tests inventory: 21,444 -> 22,329 of 23,033 BDD lines READ (96.9%)
Refs #1959
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Owner
Author
|
Rungs continue: instrument honesty + 0006. Three updates since the last note:
Ladder: 67,760 → 37,786 discarded tokens (−44 %), parse-fails 173 → 171, zero regressions at every rung. One new one-pager awaits an Architect word: 🤖 Generated with Claude Code |
…em twice
The rung: given/when/and clauses whose value is not an `ident =` binding lower
as expression STATEMENTS (`given uart_tx_send(0x55)`, `given f() == 1`), and
the bench keyword pair `measure f(x)` / `target latency_us < 5.0` lowers to
StmtExpr nodes named by their clause -- never to asserts: a bench target is a
goal, not an invariant, and inventing check semantics is what
FORALL-DECISION.md declines to do. The colon-prose forms
(`measure: nanoseconds to ...`) stay honest fallbacks: prose is not an
expression.
v1 survived the corpus sweep and fell to the adversarial panel (84 attempts,
3 lenses) three ways:
1. CONJUNCTION ABSORPTION -- `given f(0x55)` + `and g(0x66)` folded into ONE
StmtExpr holding `f(0x55) and g(0x66)`: the second side-effect call became
conditional on the first's truthiness, under "nothing discarded". The
0005 clause-break only recognised `and ident =`. Rule now: in clause-value
mode a LINE-LEADING `and` is a clause, full stop -- in the BDD layout a
genuine operator lives on the line it operates on.
2. ROLE LOSS -- `and` after `then` lowered as a bare expression, silently
shedding its assert. Gherkin semantics now: `and` inherits the role of
the clause before it, but CONTENT WINS (`and ident = ...` binds anywhere;
an and-binding does not break the assertion chain, so `then a == 1` /
`and y = 2` / `and y == 2` yields two asserts and one binding).
3. PAREN ASYMMETRY -- `given (x + 1) == 2` died in the tuple gate while the
SAME value parsed under measure. A failed tuple shape now restores and
lowers as an expression clause.
Known sharp edges, measured and accepted: a mid-expression failure
(`given f(x) +` trailing operator) still costs its whole block -- the honest
whole-block fallback, same as before this rung; and junk PROSE inside a block
that now fully lowers surfaces at module level as a hard error (corpus
instances: zero; the baseline ratchet stands guard).
Ladder (624 non-scratch specs):
0006 37,786 discarded tokens
0007 34,175 (-50% from the 67,760 base)
parse-fails 171, zero new vs baseline; consume-all 327 -> 328
Refs #1959
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
…ding T56 counted FOUR discard channels; W899's instrument fix reached three. `skip_to_semicolon` was the fourth -- eating tokens with no count and no span. Its biggest victim: a semicolon-less top-level `const` followed by another const/var (both outside is_top_level_start ON PURPOSE -- they appear inside keyword blocks) sent the tail-skip through the next declaration, the next fn, and everything to the following stray `;` or EOF. specs/ar/coa_planning.t27 -- the corpus's single largest "discarder" at 2,438 counted tokens -- actually retained ONE declaration in its AST; the rest of the file was eaten mostly UNCOUNTED. Two generations of readers bucketed it "Rust-form bodies"; the intervention map (delete the suspect, re-measure) convicted the two semicolon -less consts in minutes. The rung: (1) skip_to_semicolon records every token it eats -- all four channels now feed both accounts; (2) at depth 0 a LINE-LEADING const/var ends the skip -- the next declaration is not expression tail. Same-line `[]const u8` shapes keep their W568 parse (the line test spares them). DISCLOSED STATUS CHANGE, deliberately: coa_planning and restraint now FAIL to parse honestly (their fn bodies hold `for .. in` / `match ::` constructs that were never parsed in the files' lives -- with the eaten regions restored, the hard errors surface). Parse-fail count stays 173: two zombie parses out, the two SSOT files in. Every other regression check is green; the corpus loses 30,023 tokens (-56% from base) at this rung. Refs #1959 Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
… commit for the code) Shipped in the same compiler.rs revision as 0008; recorded separately for the ladder's cause-per-rung ledger: - 0009a `bench name: expr` -- the one-line colon body parsed under `invariant` and dropped under `bench`; the intervention map convicted the FORM (trivial expressions dropped too). Lowered exactly as invariant-colon: the predicate becomes an assertion (0004a's own precedent for bench then-clauses). - 0009b `measure: <prose>` / `target: < 50ns` -- convicted as FORMS; the content is frequently prose no expression grammar can honestly read. The rest of the line is captured VERBATIM into the node's value: tokens read and preserved, no semantics invented (FORALL-DECISION.md's own rule). - 0010 `invariant name : EXPR;` -- the trailing SEMICOLON was the convicted cause (colon, spacing, `||` all exonerated); it is now consumed after the lowered assertion. Corpus effect of the batch: 30,023 -> 27,452 discarded tokens (-59.5% from the 67,760 base); consume-all 326 -> 385; discarding files 125 -> 66; parse-fails unchanged at 173 (the 0008 membership). Lost-tests inventory: 22,601 of 23,033 BDD lines READ (98.1%). Refs #1959 Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
…s that hardened 0008-0010
0011: `value in [-1, 0, 1]` / `x in {0, 1, 2}` -- KwIn as a same-line
membership operator at comparison precedence, with a brace-set RHS parsed in
place. For-in loop headers are untouched (a for-in never reaches expression
position; the same-line rule keeps it that way).
The 64-attempt adversarial panel over 0008-0010 found eight breaks; every one
is closed and four ship as tests:
- const-stop v3: the skip now also ends at a line-leading TOP-LEVEL opener
(it previously ran THROUGH `test t` into the body and minted module consts
from test-local bindings), and a const/var stop requires a DECLARATION HEAD
(`Ident :` / `Ident =`) -- the wrapped pointer type `const P = *` newline
`const u8;` no longer mints a module constant named u8.
- var gains the same tail-skip const has: junk after a recovered var
(`then v == 2` at module level) is swallowed-and-counted instead of
hard-erroring the whole module.
- prose capture is BENCH-ONLY (a stray `measure:` inside a test captured the
block's only then-assert into inert prose -- a vacuous test under a green
light) and stops at same-line clause words (`measure: 5 target: 6` is two
nodes again; `target: < 50ns then x == 1` keeps its assert).
- bench-colon parses its head under the clause flag (a next-line `and y == 3`
was absorbed under `or` precedence into a strictly WEAKER assertion) and
CONTINUES into the shared clause parser with the assertion role preset --
`bench b: x >= 0` followed by measure/then clauses lowers completely, and
Gherkin `and` after the head asserts.
- invariant-semicolon: var/enum/struct/using now end the block as cleanly as
const/fn (the boundary list omitted them and the invariant dropped its OWN
assert when any followed). The GLOBAL boundary set is untouched.
Known pre-existing edge, documented not caused: a keyword-test body opening
with `const x = 1` ends the clause block at the boundary and the binding
parses at module scope -- frozen does the same for unpoisoned files; fixing it
is a lowering decision (statement clauses), not a recovery decision.
Ladder (624 non-scratch specs):
0010 27,452 discarded
0011+v3 26,749 (-60.5% from the 67,760 base); consume-all 385;
discarding files 66; parse-fails 173 (the disclosed membership:
two zombies out, two SSOT files in); zero other changes
Refs #1959
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Owner
Author
|
Ladder complete through 0011 — the probe-fixable frontier is nearly exhausted. Summary of the last four rungs:
Every rung carried a 3-lens adversarial break panel after its corpus sweep; the panels found (and the revisions closed) a forged-brace silent false-green, conjunction absorption of side-effect clauses, module globals minted from test bodies, prose swallowing sibling asserts, and a strictly-weaker assertion from One disclosed membership change: parse-fails stay 173, but Inventory: 98.1 % of the 23,033 BDD lines are READ (55 % were dropped at W890). What remains is decisions, not probes: the forall word ( 🤖 Generated with Claude Code |
…ing a global
`const h = 5` between clauses ended the block at the KwConst boundary and the
binding parsed at MODULE scope: a test-local name became a global for every
backend while the block's remaining clauses dropped. Now const/var (and `let`,
which lexes as KwConst) and bare `ident = expr` assignments between clauses
lower as statement clauses -- StmtLocal / fn-body-shaped StmtAssign -- under
three guards, each earned by a measured failure during this rung's own
verification:
1. only AFTER at least one accepted clause, at that clause's indent or
deeper (a shallower const/var is a real module declaration and still
ends the block; a body OPENING with const keeps the old edge, now an
honest per-line drop rather than a minted global);
2. CLEAN TERMINATION -- semicolon, new line, boundary or EOF. parse_expr
returns Ok on a partial brace-if read (W578), and accepting that
"success" left the loop on mid-line junk and cost the whole block
(jones_topology_filter went 35 -> 75 dropped before this guard);
3. an arm failure restores and ENDS the block instead of whole-block
fallback -- module-level statement parsing then reads what it can.
Measured (624 non-scratch specs):
0011 26,749 discarded tokens
0012 26,713 (-60.6% from the 67,760 base); jones_topology_filter
35 -> 16 (the whole invariant body incl. the bare assignment now
lowers; only W578 brace-if innards remain); zero regressions,
parse-fails 173 (the disclosed membership), all panel probes hold
Refs #1959
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
… guards that came back
A 75-attempt three-lens panel against 0012 found twelve breaks; the worst
class was SCOPE THEFT with corrupted dataflow: at a column-1 tie an unindented
block absorbed the module const the NEXT test read (`rc=0`, "nothing
discarded", and a global silently became a local). Every break is closed:
- column-1 tie DISABLES the statement arms (module declarations and body
statements are indistinguishable by column there);
- the anchor is the MINIMUM accepted clause column, not the first (an
over-indented first given made a col-5 body statement look "shallower" and
hoisted it while its own siblings were accepted at col 5);
- ADJACENCY: a statement clause sits on the line immediately after the
previous token -- a blank-line gap means module scope, as every reader
assumed;
- clean landing is measured against the LAST CONSUMED token's line, not the
statement's ("line > start" blessed a mid-line stop on a later line and
split one physical line into two minted statements);
- the bench-colon continuation gate admits KwConst/KwVar and presets the
column anchor (a `let` after the head was minted as a module const holding
a conjunction).
Documented, not fixed here: tab-vs-space byte columns invert the visual
indent relation (the corpus is tab-free; a tab rule needs a lexer decision);
a dirty statement that breaks to module level meets const's own truncating
fast path (pre-existing); a later clause failure still takes the whole block
(the standing whole-block honesty rule).
Corpus: unchanged to the token -- 26,713 discarded, 66 files, consume-all
385, parse-fails 173 (the disclosed membership) -- while six panel probes
flip from theft or silent line-splitting to correct scope or honest drops.
Three probes ship as tests.
Refs #1959
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Each cause below was CONVICTED BY INTERVENTION in the W905 causal map (delete
the suspect in a scratch copy, drops vanish) before any code moved:
1. LVALUE STEPS -- `and state.gamma[0] = 2.0`, `when buffers.scores[0..4] =
scores`: the expression path parsed the lvalue, stopped on `=`, and the
over-consumption guard felled the block. A dotted/indexed lvalue landing on
a plain `=` now lowers as StmtAssign, under the same clean-landing rule as
every statement clause.
2. KEYWORD AS FIELD LABEL -- `Contract { ..., invariant: "..." }`: the struct
literal field loop broke on the keyword and the block fell. Any word-shaped
token followed by `=`/`:` is a field label in that position.
3. `var` AS A NAME -- `given var = 5` / `then var == 5`: the binding arm
accepts var/const keywords directly before `=` as names, and in
clause-value mode ONLY, a var/const keyword not opening a declaration reads
as an identifier in operand position.
4. UNIT PHRASES -- `target throughput > 1000 steps/sec`: the same-line residue
after the bound is captured verbatim into the node's value, exactly like
the colon form's prose. Read, not interpreted.
Remaining from the six: inline lambdas in then-clauses and given-comprehensions
(~235 tokens) -- left priced, not smuggled.
Measured (624 non-scratch specs):
0012 v2 26,713 discarded
0013 25,905 (-61.8% from the 67,760 base); consume-all 385 -> 387;
discarding files 66 -> 64; parse-fails 173 (the disclosed
membership), zero new; every prior family and panel probe holds
Refs #1959
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Owner
Author
|
The ladder closes at thirteen rungs: −61.8 %, and the residue is priced. Final account for this branch (
Every rung carried probes, an adversarial break panel, and a full-corpus certificate; the map that priced the residue used delete-and-remeasure convictions, and four of its six small causes shipped as rung 0013 in one wave (lvalue steps, keyword field labels, What remains needs two words from a human, not more probes:
A number (1–4) for the first and a word for the second on this thread finishes the arc. 🤖 Generated with Claude Code |
The last two convicted causes from the priced frontier:
1. INLINE LAMBDA -- `then all(result.v, fn(x) x >= 0.0)`: represented without
a new node kind as an ExprCall named "fn" (parameters joined in
extra_field, single-expression body as the only child), clause-value mode
only. Convicted at W905 by swapping the lambda for a named predicate.
2. COMPREHENSION SUFFIX -- `given encoded = encode(x) for x in {...}`: the
same-line `for ...` tail is captured verbatim into the binding's
extra_field. Read and preserved; no semantics invented.
Measured (624 non-scratch specs):
0013 25,905 discarded
0014 25,670 (-62.1% from the 67,760 base); consume-all 387 -> 389;
discarding files 64 -> 62; parse-fails 173, zero new; every
family and panel probe of all fourteen rungs holds
With this rung the PROBE-FIXABLE TAIL IS ZERO: everything that remains is the
forall decision (~74% of the residue), the specs/ar dialect decision, and the
long-standing W578 brace-if expression question. The ladder rests until a
human speaks.
Refs #1959
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Owner
Author
|
Option 2 is now REHEARSED — say «2» and it lands in one wave. A 58-line parse-only capture of forall bodies was built, measured on the full corpus, and reverted pending your word:
Patch and numbers: 🤖 Generated with Claude Code |
Owner
Author
|
Both decisions are now rehearsed — two words take the corpus to −93 %. The dialect question got the same treatment as forall: built, measured on the full corpus, reverted.
Zero new parse-fails in every configuration; patches in One number (forall: 1–4) and one word (dialect: 1–4) on this thread close the arc. 🤖 Generated with Claude Code |
Owner
Author
|
Entry point for whoever wakes up: the one-page status board. Both arcs are closed, both remaining decisions are rehearsed, mutation-checked, and priced; five human actions finish everything:
The autonomous loop now holds quietly: ratchets green (parse-baseline 173 = 173 at every check), working branch clean, prototype identical to ring HEAD, scratch manifested. Answering here wakes the work. 🤖 Generated with Claude Code |
…zero regressions master and this branch independently built several of the same things; the merge takes the UNION with each side's stronger half: - compound assignment: the full operator set (+= -= *= /= %= |= &= ^=) -- ours brought /= and %=, master brought |= &= ^= and the compound_binop emitter helper, which now serves all eight in all three backends; - struct-literal suppression in paren-less conditions: ours' reentrant COUNTER (nested conditions) over master's bool, same rule; - Zig casts: master's W566 narrowing @truncate for provably-unsigned integers COMPOSED with ours' W592 float builtins (@floatCast/@floatFromInt/ @intFromFloat) -- floats first, then the narrowing choice; - type annotations: ours' superset (& borrow + ?optional + quoted machine types); SystemVerilog keyword table: ours' superset (-g2012 full list); - seal refusal: master's per-backend failures check (stronger than ours' vacuous-only check) with our CLI family (battery/known/provenance/verdict/ silicon/...) intact; - specs: master's content taken wholesale (the ladder's thesis -- teach the grammar, do not rewrite the tests -- applies to merges too). Certificate, measured on the merged tree (650 non-scratch specs): the merged compiler parses a STRICT SUPERSET of what rung 0014 parses -- 0 regressions, 5 files newly parse (types, formal, linker, vcd_trace, ir); corpus discard 6,782 tokens across 38 files, consume-all 421; every ladder probe and panel test green; parse_baseline.txt refreshed with the merged binary on the merged tree. Refs #1959 Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
gHashTag
enabled auto-merge
August 19, 2026 09:54
Contributor
Contributor
|
📓 NotebookLM Notebook linked to this PR
This notebook contains session context, decisions, and artifacts for this work. |
…instrument Eight iterations against master's own corpus ratchet (33 -> 11 -> 9 -> 8 -> 7 -> 6 -> 4 -> CLEAN), each failure diagnosed to its construct: - the fourth-channel stop ran THROUGH `test` headers into keyword bodies and a semicolon-less const swallowed files; the skip now also ends at any line-leading top-level opener; - `measure:` prose no longer breaks on English `when`/`and`/`given` (stop-set is measure/target/then/assert); - a stray `}` in a flat module -- master's own #2186 balancing braces -- is a RECORDED skip, not a hard error (W577's no-silent-truncation promise kept via accounting); - keyword-identifier family closed: `module` in expressions (transmuted into the Ident arm so calls/fields still attach), `let mut name` (Rust modifier), `match` blocks captured verbatim with zero silent loss, `as float`/`as int` cast aliases, for-range heads no longer eat the loop body as a struct literal; - one spec fix by master's own precedent (#2186): the unterminated `rag_retrieve_architecture` in igla/coder/arch.t27 gets its closing brace; - the expectations ledger re-blessed: 109 entries REMOVED as fixed by the ladder, 4 re-priced under the four-channel instrument that counts what the frozen one could not see (the ledger's old zeros were blind, not clean). Also: instrument artefact #14 found in the loop's own tooling -- `echo "$(basename $f) rc=$?"` prints basename's exit code, not the parser's; every rc in this branch's verification scripts is now captured before echo. Certificate: suite --ratchet --corpus-only exits 0, RATCHET CLEAN; every ladder probe and panel test green. Refs #1959 Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Contributor
|
📓 NotebookLM Notebook linked to this PR
This notebook contains session context, decisions, and artifacts for this work. |
Contributor
PR DashboardGenerated at: 2026-08-19 11:42:25 UTC
Summary
Seal Status
|
…honored and flagged; NOW updated (Refs #1959) Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Contributor
|
📓 NotebookLM Notebook linked to this PR
This notebook contains session context, decisions, and artifacts for this work. |
Contributor
PR DashboardGenerated at: 2026-08-19 12:24:24 UTC
Summary
Seal Status
|
This was referenced Aug 19, 2026
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Two grammar changes that make the L6 numeric SSOT (
gf16.t27) parseable and sealable by the repository's own compiler for the first time —*=was never in the bootstrap grammar (git log -Sis empty) and the SSOT's Taylor loop uses it.Per FROZEN.md this touches
bootstrap/src/compiler.rsunder FROZEN_HASH, so it ships as a branch for Architect approval — verified in a detached copy, never applied outside this branch.Change 1 — compound assignment
-= *= /= %=: four lexer cases beside the existing+=, a five-way parser match, three codegen sites writing the operator through (Verilog desugars tolhs <= lhs <op> rhsexactly as+=already did).Change 2 — nested
fn, hoisted fromparse_fn_body, with the capture check at the only scope hoisting crosses: the nested fn must not reach the enclosing fn's params/prior locals; any other free name resolves identically before and after the move. Negative repro fails withcaptures enclosing locals ["scale"].Evidence (repros in
bootstrap/tests/goldring/): gf16 parses completely, all four backends emit real hashes; tf3 fixed by change 1 alone; unpatched control still fails at gf16:824; corpus regression sweep old-vs-new with byte-identical hashes required — 0 regressions in the first 679/1,079, final number follows as a comment. Blast radius is deliberately exactly the SSOT family; the 34 remaining parse failures are three dialects (docs/reports/gold-ring/DIALECTS.md) whose canonicalisation this patch does not pre-empt.🤖 Generated with Claude Code
Refs #1959