Logbook · 1–23 September 2026
Logbook
What this project believed, built, measured and gave up, in the order it happened. Every experiment id links to its entry in the log; every number on this page is transcribed from EXPERIMENTS.md, from the Results page, or — for experiments still running — from the run files named beside it.
The project began as a test of one idea: that a fly's wiring diagram, trained by a local forward-forward rule, could learn to choose which refractory alloy to compute next. Within a day the test had failed, and backpropagation had failed the same way, which put the fault somewhere other than the learning rule. Forward-forward training was given up. What the project became — a generator steered by a readout on the connectome, feeding a five-rung fidelity ladder that asks whether a bcc alloy keeps one phase from 90 to 1000 K — grew out of the failures below. The old Results page is kept unchanged as the record of the three September runs. The September book, its forward-forward chapter included, is archived in the repository under book/theory_2026-09-15/; the site's Book now teaches the current method.
Each era is written the same way: what was believed, what was built, what was measured, what failed or was withdrawn, and what replaced it. The Corrections table at the end lists every claim the log withdrew or corrected.
git log, and the log's eras placed by the dates and times its entries record. Era 0, the idea from about 1 September, has no commits or dated entries and sits left of the break in the day axis. Eras 3 and 4 (E46–E147) share one hatched band: entries E46–E119 carry no date, and the commits either side of them are 13 September 08:16 and 16 September 02:53. Days without a bar had no commits. Each band links to its era.0 · Before the code
The idea came before any code, around the first of September: could a fly's wiring diagram choose which alloy to compute next? The operator places the start there. This machine keeps only later traces. The first MaleCNS render in the book is dated 10 September, 22:57, and so is a crystal-gym demo of an alloy environment. The repository's first commit is from 11 September, 23:04. The log below starts at that commit, so the ten days of thinking before it left no numbers to transcribe.
1 · A brain that did not choose
BelievedA recurrent network on the MaleCNS wiring diagram, scored by the mean "goodness" of all its neurons and trained by a reward-modulated Forward-Forward rule that changes each synapse from local information only, could learn to pick which of 64 fixed refractory alloys to evaluate next.
BuiltThree runs of 40 MACE-MP-0 screens and 10 Quantum ESPRESSO calculations each: the 16,384 best-connected neurons at 0 K, the whole brain (163,972 neurons, 6,143,838 connections) at 0 K, and the whole brain at 1500 K and 5 GPa. They are on the Results page.
MeasuredThe three runs chose with the same probabilities, and the curve was 1/(alloys remaining) — uniform sampling with a fixed seed; the first twelve alloys were the same, in the same order, in all three runs (Results page). Learning moved the weights by ‖Δw‖ = 111 and every choice probability by at most 0.0015 (as E15 restates it). E1 then replaced the local rule with backpropagation through the three settling steps and changed nothing else: the update directions agreed at cosine 0.9916, and the live run reproduced the Forward-Forward screening sequence in 39 of 40 picks. The learning rule was not what failed; the fault was upstream of it. Reading the 97 mushroom-body output neurons instead of the mean over all neurons broke the tie between score and input norm (E8), and taking the encoder and the readout's valence from the anatomy made the ranking stable across seeds, mean pairwise Spearman 0.9995 (E9). A degree-preserving rewire of the 61,210 Kenyon-cell-to-output synapses then moved the ranking (ρ 0.868 ± 0.016, E10), and the measured wiring ranked candidates closer to the objective than rewired copies did, −0.755 against −0.679 (E11) — but ridge regression on the eight concentrations alone reached 0.948 (E12). Plasticity at the one learning site was directionally real and functionally negligible: +0.0097 against a gap of 0.37 (E15).
Failed or withdrawnThe controller over a fixed list, and with it forward-forward training: the rule (LocalFF) appears in no experiment after E8. Along the way E6 found that the connectome importer's five-contact threshold had discarded 76 % of all connections, and E13 replaced the screening potential: MACE-MP-0 ranked the ten DFT alloys at Spearman 0.770 with a bias of −60.4 meV/atom, MACE-MPA-0 at 0.927. E14 counted why the old one failed — its training set holds no quaternary or quinary metallic refractory alloy.
Replaced byThe mushroom body as a learner and a generator, and an environment cheap enough to test it in.
2 · A learner, an environment, and a benchmark it lost
BelievedA cluster expansion fitted to a validated potential would make the environment instant; the mushroom body, given a sparse odour code and a proper teaching signal, would learn the free-energy landscape and search it better than sampling.
BuiltAn eight-element icet expansion on a fixed bcc lattice at 3.2935 Å, 344 parameters, fitted to MACE-MPA-0 mixing energies, cross-validated at 5.67 meV/atom (E16b). MACE-MPA-0 checked at 6.61 meV/atom on 300 near-equilibrium structures of the RHEA DFT database (E18). Configurational entropy by thermodynamic integration (E21) and a composition-level free energy at 362 compositions per second (E23). The mushroom body with rectified, sparse Kenyon cells and compartment-specific dopamine (E27).
MeasuredWith a sparse code the circuit learned (held-out Spearman −0.23 → +0.15, E27); with dopamine reporting a prediction error it reached 0.853 against 0.561 for ridge on the composition (E28). As a run-and-tumble generator it beat Latin-hypercube sampling at 60 evaluations, −471.6 ± 7.6 against −456.8 ± 6.6 meV/atom (E30). A Gaussian process then beat it on 10 of 10 seeds (E33), and two pre-registered comparisons were losses (E35, E37). E44 found that every one of those comparisons had changed the generator and the selector at once; with only the generator varied, the circuit drew with a space-filling design (−479.9 ± 1.6 against −479.0 ± 1.3). On the physics, E42 reproduced Sobieraj et al.'s Ta–Ti–V–W at 383 ± 75 K against their 500 K, with Ta–W — one of their two most influential pairs — as its strongest, and E45 mapped ordering over 219 compositions.
Failed or withdrawnE29: the fly's error is bias — 99.8 % reproducible across training orders — so three variance-based uncertainty signals all came out anti-correlated with the error. E31 withdrew the free-energy fast path's 0.96 meV/atom accuracy (its reference cell self-interacted; 1.33 stands). E33 found an unexplained 11 meV/atom shift in the random baseline, which left E30's t = 4.6 unsafe. E36 showed E34's turning points were sampling noise. E17 did not confirm E3's "128 atoms is the minimum" in DFT, and E41 showed E39's agreement with a published transition had been read from a single Monte Carlo run.
Replaced byConformal intervals as the circuit's honest uncertainty (E40), and a target the generator could be pointed at: where alloys order, and at what temperature.
3 · The whole brain as a generator, and a survey on the wrong lattice
BelievedWiring in the whole connectome as a rate model would let the brain generate the search; a survey of ordering temperatures over the bcc expansion would name the alloys that stay one phase.
BuiltThe whole brain as a rate model over 163,972 neurons with measured transmitter signs (E46); the survey's qualifiers from the bcc expansion; a kinetic rung from vacancy formation and migration barriers (E57, E58); a free off-lattice screen at 4 ms per composition (E59); six generator arms behind one interface (E60).
MeasuredAt its operating point the whole brain was a 7 × 8 matrix: an affine map reproduced its descending-neuron heading at cosine 0.980 (E47). At a corrected operating point it searched no better than that matrix (+1.26 meV/atom, t = 0.34) and 36 meV/atom worse than a random heading; two pre-registered failure conditions fired (E49). The survey's qualifiers sat on a lattice that does not form: C15 Laves HfV₂ and ZrV₂ lie 123 and 119 meV/atom below the bcc solution (E54), and against a 136-phase off-lattice hull all seven qualifiers sat 98 to 185 meV/atom above it at 90 K — while MoNbTaW, which the survey had rejected, sat below it across the window (E56). With kinetics added, the requirement had an answer, and it was that rejected alloy (E58). On the real target, elitist and cross-entropy search beat the fly and learning was not shown (fly against its frozen twin p = 0.40, E60) — until E61 found the Kenyon-cell code starved to 0.07 % active on the fly's own proposals, three cells of 4,064. With the threshold set on the code that actually fires, the fly beat its frozen twin on 16 of 16 seeds (+4.85 ± 0.81, p = 1.2 × 10⁻⁵) and matched, without beating, the elitist hill-climber (p = 0.138).
Failed or withdrawnThe whole brain as a generator (E47–E49). All seven survey qualifiers — Laves artefacts of a bcc-only model (E54, E56). Every mixing energy reported to that point, too negative by 226 ± 43 meV/atom because the elements were referenced on a strained shared lattice (E53); the fix reached the script and not the stored expansion (E55). E51: the plasticity rule's sign was inverted through the output neurons, and results before the fix rested on weights driven to about 150 times their anatomical value. E62's "tungsten freezes the alloy through its migration barrier" (E64: a controlled series gave corr(W, E_m) = −0.03; tungsten raises the vacancy formation energy instead). E65, which accused the expansion of failing outside its box, was itself a convention error. E68: ordering-temperature error bars were 1.7–2.1 times too tight.
Replaced byA verdict with three terms — ordering on the lattice, stability off it, reachability — asked by rungs of rising cost.
AddedSetting the Kenyon-cell threshold on the code that actually fires, not on its raw input (E61). See What we added.
4 · A ladder, audited against itself
BelievedA ladder of rungs — screen, expansion Monte Carlo, MACE lattice, kinetics, DFT — would make the generator's finds trustworthy; widening to twelve elements (Cr, Co, Cu, Ni) would open the search.
BuiltThe DFT rung (E75–E78); the eight elemental corners pinned into the expansion (E80); an off-lattice hull taken from the materials databases, with bcc orderings removed by coordination number (E84–E86); a twelve-element expansion (E85); MAP-Elites and an archive arm (E114–E125); a written specification of every rung, with costs and blind spots (E119).
MeasuredFirst principles agreed with MACE on the gaps the ranking uses, to 1 and 4 meV/atom (E78). The expansion was wrong by 58 meV/atom at pure Mo, ten times its cross-validation error, so the generator had been paid for finding the model's corners (E72). The screen was rewarding the ordering hazard: three of four generator candidates ordered inside the window (E96). The ladder reached its top rung for the first time only after its early-return rule was reversed (E105; MoNbTaW, p = 0.9404 in 286 s). Against 75 published ordering temperatures, rung 1 read 0.54 of the published scale (E108); E110 and E111 traced most of it to a flat baseline under a 1/T² tail (×1.35) and a mean-field reference (×0.77). Measured against RHEA DFT at matched lattice and relaxation, the bottom rung's error was 8.1 meV/atom (E130). Searched continuously rather than on a hand-built grid, 11 compositions passed every correction where the grid had found 2 (E133). The connectome generator had not been run in any of these comparisons; run, its best find was 4 meV/atom off cross-entropy's at a much lower rate (AUC_Q 1.00 against 13.57, E136), on 5,311 neurons — 2.5 % of the annotated brain (E137).
Failed or withdrawnE83's nickel expansion (two reference conventions in one fit). E84's twelve-element hull (a two-atom B2 Mo–Ta cell turned every refractory answer positive, E85). E91's claim that kinetics made high-entropy alloys competitive (E92: the melting-point proxy for the activation energy over-predicted by up to 1.58 eV, always toward stability). "Mo–Ta is the best composition in the space" (E96). The rung-0 ordering gate as a clearance — off stoichiometry (E99), then at equiatomic too, where it read 0.09–0.77 of the Monte Carlo value (E142). The Cr–Ni basin (E100; E103, E104: a 168 meV/atom model error). E114's verdict against MAP-Elites (E115: implementation, not method). E117's 55 feasible compositions (12 in E131, 2 in E132). E128's 44.6 meV/atom bottom-rung error (E129: the relaxation convention). E134's "the high-entropy class fails on physics" (E135: MoNbTaTiW orders above 1500 K where the gate said 121 K). E138, E141 and E143 disagreed about whether rung 1 depends on where its sweep stops; it does when the transition sits near or above the ceiling, and no rung had ever consulted the melting point.
Replaced byA rule written in E116 and broken repeatedly afterwards: measure before changing the code. And a decision to stop fitting the bottom rung to a potential — RHEA's DFT would be the training set.
AddedA ladder written down rung by rung — question, cost, measured error, blind spots (E119) — with a prediction and a bar registered before each result. See What we added.
5 · RHEA's DFT as the training set; the head gets a sensory surface
BelievedAn embedded cluster expansion (pyeCE) trained on RHEA's DFT would give a better bottom rung; the whole connectome, not the 5,311-neuron mushroom body, would be the substrate.
BuiltA label pipeline that removes each RHEA frame's rattle and strain with one MACE difference (E150), and a strain cut at 0.10 Å chosen by the noise among random decorations: a label noise floor of 12.4 meV/atom against a 29 meV/atom ordering spread (E151). The first eCE on those labels (E158). For the head: the composition broadcast onto all 17,479 afferent neurons (E157, E163), a learned readout (E159), plastic sites outside the mushroom body (E166), per-walker gains read from the head (E167), and a GPU port of the propagation, 47 times faster (E184). The flywheel: DFT cells at the search's own finds (E172–E186).
MeasuredOn compositions it never saw, the eCE ranked configurations at ρ 0.803 against 0.651 for the old expansion (E158). With the anatomy's readout the whole brain searched about ten times worse than the mushroom body alone, AUC_Q 0.15 against 1.48 (E152); its olfactory input reached 2.7 % of the brain, and a broadcast onto the afferent class reached 25.1 % (E157). The broadcast lifted AUC_Q from 2.61 to 11.57 (E163); per-walker gains lifted it to 25.93 (E167). The first DFT cells seemed to show the eCE compressing formation energies two- to four-fold (E172, E176, E177). E180 traced that to regression references and E183 to fitted Hf, Ti and Zr references; with all nine references from corrected pure cells, v5 agreed with six DFT cells to 3–10 meV/atom (E186).
Failed or withdrawnThe first eCE had seen no ordered cell (E149), and both training sets carried displacement and volume energy the model cannot represent (E150). "The expansion is 45 times too flat" (E150, E151: correction error on random binaries). E152's first substrate control could not have been one — the mushroom-body file had been overwritten by the whole brain. E159's "reading LAL and DNa is the mechanism" (E161: 99.6 % of the learned weight sat on the output neurons). E167's "the head sets each walker's gain" (E170: shuffled gains did as well). E169's steering from the descending neurons collapsed the search (AUC_Q 0.63). E181's "the model's own extrapolation failure" (E183: the labels' references, not the model).
Replaced byv5 as the bottom rung (E186), and a head to be judged next on a reward that can tell good from better.
AddedLabels a lattice model can learn from: each RHEA frame's rattle and strain removed by one MACE difference, with the strain cut chosen by the noise among random decorations (E150, E151). See What we added.
6 · The head, on a reward that ranks
BelievedThe fly's advantage over model-free search would survive a reward that ranks its finds; plastic sites in the fan-shaped body and under serotonin would add learning outside the mushroom body.
BuiltThe search paid from v5 (E193) — the old reward had scored every find at −10.5 to −11.7 meV/atom (E187). Five-arm comparisons (E195, E195b). Controls for the head's gain field (E198). A fly branch with local search, imagination and a Bennett prediction-error rule (E202–E205).
MeasuredThe reward was coupled to process history: an ordering estimate inside it shared one random generator across the process, so every multi-arm comparison run in one process had been scored on a history-dependent reward (fixed in commit 247ed9d; E199 is void). On the pure reward the readout learns the reward — score–reward ρ +0.90 live against −0.03 frozen (E201) — and over three seeds the fly finds more distinct qualifying alloys than the elitist search at the same rate: 299 ± 31 against 238 ± 14, AUC_Q 147.1 against 144.8, a tie (E195c, E201b). v5 predicted the DFT formation energy of the generator's best find, Mo₃₃Ti₁₉Nb₂, to 8 meV/atom (E194), and of Mo₃₅Ta₁₂Ti₇ to 3 (−99.1 meV/atom); it overstated Mo₃₆Nb₁₈ by 28 (E196).
Failed or withdrawnThe fan-shaped-body site: runs with and without it were bit-identical, because the site sits one propagation step beyond the recurrence and moved the scores by exactly 0.0 (N6); the gap between E171b and E178 was code drift. The head's own gain field was the worst of three on both rewards (E198; E198b: live 136.9 against fixed random 153.2 AUC_Q) — the modulation channel carries nothing. Local search halved the fly's finds (E202), imagination made them shallower (E203), and E205 combined two falsified changes. E195b's "elitist ≥ fly on every column" belonged to the gated reward (E201b).
Replaced byA narrower claim for the generator — its advantage is width, not depth — and a decisive test of whether the wiring matters at all (E224, below).
AddedControls that can falsify the generator: a frozen twin, rewired and random heads, and model-free swarms given the fly's own diversity devices, all on one reward (E201, E224). See What we added.
7 · Rung 1: a sampler at twice the temperature
Believedv5's ordering temperatures were about half the published ones — Mo–Ta 500 ± 50 K (E188), MoNbTaVW 367 ± 65 K against published 750 / 742 K (E190b) — so the model under-ordered and needed active learning on its own sampled states (E191, E192) and refits.
BuiltTwo refits, v6a and v6b; a reversible two-leg sweep at 10–30 times the sampling (E207); ground states by annealing (E211); a from-scratch Metropolis sampler on v5 and an exact nearest-neighbour Ising control (E215, E216). Alongside: the magnetic blind spot measured as four numbers — spin minus non-spin, Cr −12.4, Ni −61.1, Co −170.2, Fe −467.4 meV/atom (E216) — and a first fcc bottom rung, where a model on Li et al.'s set reached 17.9 meV/atom on held-out random (SQS) cells after a model on Waters' set failed every bar.
MeasuredBoth refits moved the transition down, not up (N4). The sweeps were reversible on all three systems (E207), and cell size, sampling depth, direction and ceiling each failed to move the Mo–Ta number (E200, E207, E214). Then E215 put a known 1410 K Ising transition through pyeCE's sampler and got 612 K. The cause was in pyeCE's source: its models output a per-atom energy for each two-site cell, and the acceptance sums those outputs, so the sampler runs at twice the temperature it reports. Measured on the sampler's own swap, the ratio is 0.500 for v5; a second sampler gave 1.9–2.0; E215c on a 20 K grid gave 2.000 ± 0.03. Two code referees found a second encoding: pyeCE's short-range-order column is (α + σ)/2, not α, and is decoded exactly for ordering pairs. On the physical axis v5 put Mo–Ta at 910–1037 K and MoNbTaVW at 697–702 K. The nine-system scorecard, run on the second sampler, placed six of six comparable systems within 22 % of the published CE + Monte Carlo values (E210b); the seventh, Cr–Ta–Ti–W at 2.92 times, was one pair over-ordered — DFT confirmed that chromium orders against tantalum, at 44.9 meV/atom against v5's 107.9 (E222).
Failed or withdrawn"Rung 1 reads half the published value" as a statement about the model. The supercooling hypothesis (E207). "v5's ordering energy is nearest-neighbour" (E213: a collinearity artefact; v5 is nearest-neighbour plus a real second-shell coupling, E220) and the 1409 K yardstick built on it. The thermostat-ceiling reading (E214). Every short-range-order onset read in pyeCE's units. E190's validation at 128 sites (a resolution failure). "Published 600–1000 K for Mo–Ta", which came from prose; Kim & Widom's number is 2020 K. The log keeps a ledger of the withdrawals of 21 September and what replaced each.
Replaced byRung 1 as a Metropolis sampler on the physical temperature axis, wired into the ladder on 21 September, and a scorecard whose every row names its observable, lattice convention and spin frame.
AddedThe sampler's factor of two, found on a transition known exactly and confirmed by a second sampler written from scratch, and the decoding of its short-range-order column (E215, E215c). See What we added.
8 · Choosing the model that ships
Believedv5 under-binds Mo–Ta by about 30 meV/atom (E191, E211), a defect a refit on this project's own DFT cells would repair (v7).
Builtv7 and v7b; a scorer on the eleven DFT cells no model chose; an audit of the labeller (E228); three-seed ensembles; embedding and pair-range arms (E229, E233, E237); ExactECE, an exact numpy re-evaluation of pyeCE's forward; Lean 4 proofs of six conventions the ladder's numbers rest on, each tied to a test (formal/).
MeasuredThe refit metric was blind: across three seeds v5's held-out error moved 7.2 / 7.4 / 7.5 meV/atom while its error on the eleven unbiased cells moved 22.9 / 16.6 / 23.8 — a seed spread of 8.5 meV/atom per cell. MACE is not soft (force slope 1.014); the labeller's volume fit missed its minimum by +17.2 meV/atom (E228). Capacity was the Mo–Ta lever: embedding 9 took B2 Mo–Ta from +29.5 to +0.2 (E229) but over-bound the search's own random cells with a narrow seed spread (E223, rung 4); pairs to 10 Å helped everything except Mo–Ta (E233). Embedding 6 was the one that held both: eleven-cell error 7.5, B2 Mo–Ta −5.2, the search's random cells 6.1 (E237). RHEA's deepest ordered labels agreed with QE within 13 meV/atom (E232), so the ordered-state error had been the model's. ExactECE matched pyeCE to 0.023 meV per swap on seven models and ran 54 times faster at embedding 9, after E236 had been stopped because the embedding-9 Hamiltonian was 55 times slower to sample.
Failed or withdrawnv7 and v7b were not deployed: they bought accuracy on the new cells with accuracy on RHEA's (held-out 9.3 and 9.9 against 7.2), and v7 had counted one B2 cell three times. The per-group VASP-against-QE offset (no such offset where the frames overlap). Eighteen training rows had been written as the wrong decoration in every fit, and E234's fit crashed on a row skipped without warning. No new-label arm shipped (E238): median Ti, Cr and Hf references over-bound the random cells; E239's DFT anchor cells are measuring the right ones.
Replaced byRung 0 and rung 1 on E237's embedding-6 three-seed ensemble on the older labels — E238's pre-registered fallback — with rung 1 sampling the ensemble mean through ExactECE.
AddedAn exact evaluator for the embedded cluster expansion, 54 times faster at embedding 9; seed ensembles as the model's uncertainty; Lean 4 proofs tied to tests of the production code; DFT anchor cells for the label frame's references (E239). See What we added.
9 · The first full walk
BelievedThe first search walked up the corrected ladder would name compositions with no phase change anywhere from 90 to 1000 K.
BuiltE223's search and a walk of its 32 finds; a rung-1 sweep that continues to 80 and 60 K; the re-walk on the shipped model (E240); a survivor chosen by agreement of three independent models (E241); a DFT test of that survivor (E242).
MeasuredE223 left 24 of its 32 finds at "no transition found above 100 K", which the window's 90 K floor made unpassable by construction; with the sweep extended, all 24 carry a verdict (E240). Reading the survivor's curve before spending DFT time exposed two defects in how rung 1 handled the floor: it consulted only the energy-slope channel, although five finds had a fluctuation-channel peak inside the window; and the lattice rung scored every floor pass as censored, so rung 2 never saw them. Both were fixed and E240's saved results re-scored without re-sampling: 18 of 32 p values changed, and 11 finds pass rungs 0–2 (p 0.50–0.77). E241's rule, fixed before E240 finished, picked Mo₅₁Ti₃₈W₃Ta₃ (p 0.771; MACE driving force −110.0 meV/atom; all three votes). The known-answer check for the shipped model's rung 1 passed: equiatomic Mo–Ta orders at 1149 ± 114 K, against a published CE + Monte Carlo 2020 ± 545 K (E242).
PendingTwo experiments decide what the project has found.
E242 · pending
Does Mo₅₁Ti₃₈W₃Ta₃ order? The shipped model's lowest-energy 54-atom arrangement of the survivor and a random arrangement of the same composition are being computed with DFT at the ladder's standard (60/720 Ry, k 4 × 4 × 4). Written before the run: an ordering energy of at most 5 meV/atom confirms the find as the ladder's first full pass; 5–15 leaves it not established; above 15 the find fails. The model's lowest state is only the lowest the model can find, so a small DFT ordering energy cannot rule out a different order that DFT prefers; the entry says so.
Already passed: the known-answer check on the shipped model's ordering rung — equiatomic Mo–Ta orders at 1149 ± 114 K against a published CE + Monte Carlo 2020 ± 545 K.
E224 · in progress
Does the connectome's wiring matter? Eight arms, five seeds each, paired by seed: the real wiring (A), a degree-preserving rewire (B), a random head of matched sparsity (C), a frozen twin (D), and four baselines — elitist search, elitist with the fly's diversity devices, a ridge readout driving the same swarm, and MAP-Elites (E–H). Seeds recorded so far: A 5, B 5, C 5, D 4, E–H 0 of 20.
| Seed | Real wiring (A) | Rewired (B) | A − B |
|---|---|---|---|
| 0 | 333 | 301 | +32 |
| 1 | 278 | 332 | −54 |
| 2 | 333 | 315 | +18 |
| 3 | 15 | 240 | −225 |
| 4 | 203 | 334 | −131 |
Distinct qualifying finds over 800 verifications. The bar written before the run: all five A − B differences of one sign, with a median of at least 31. 2 of 5 differences are positive and the median is −54. On the five pairs recorded that bar is not met. Seed 3 was a cold-start stall for the real and the random head — qualifying finds per 50-round window 0 / 0 / 1 / 34 and 0 / 0 / 0 / 1 — which a replay is examining. No verdict is drawn until every arm is in.
Read at build time from runs/odt_map/stage_b_e224_<arm>_s<seed>.json; no runs/e224_card.txt yet.
Failed or withdrawnThe first survivor pick, Mo₅₂W₂₅Ta₂₂, was one of the five floor-rule failures and is void. E240's prediction that the Mo–Ta-rich finds would order more than 150 K higher on the new model was falsified. The five finds the fluctuation channel failed are "not established", not proven to order: that channel has its own cold-end artefact, a spurious peak at 80 K on Mo–Ta.
AddedReading a transition from both heat-capacity channels, a sweep that continues below the window's floor, and a survivor chosen by agreement of independent models rather than by one learned model (E240, E241). See What we added.
Corrections
Every claim the log withdrew or corrected, with the experiment that did it. Dates are the log's; entries E46–E119 carry none and are marked by their era.
| When | Claim | Why it was withdrawn or corrected | Where |
|---|---|---|---|
| 11 Sep | The controller failed for want of credit assignment | Backpropagation reproduced the local rule's picks, 39 of 40 | E1 |
| 12 Sep | The encoder is the binding constraint | Changing it moved nothing; the readout did | E8, E9 |
| 12 Sep | Measured wiring beating a rewire shows value | Ridge on eight concentrations beat both | E11, E12 |
| 12 Sep | 128 atoms is the minimum cell | Measured in MACE only; the DFT shift from 16 to 54 atoms is within noise | E3, E17 |
| 12 Sep | 23 % chance one occupancy flips a ranking | About 7 % with DFT's own scatter | E2, E17 |
| 12 Sep | Free-energy fast path accurate to 0.96 meV/atom | Reference cell self-interacted; 1.33 stands | E24, E31 |
| 12 Sep | The fly beats random sampling at t = 4.6 | Random baseline shifted 11 meV/atom, unexplained | E30, E33 |
| 12 Sep | Smearing paths have one to six turning points | Sampling noise; the exact curves are monotone | E34, E36 |
| 12 Sep | Mo–Ta strongest pair and 700 K for MoNbTaVW | Read from one Monte Carlo run; not robust | E39, E41 |
| 13 Sep | ± 173 K scatter on every transition | The estimator (argmax), not the physics | E45, E52 |
| 13 Sep | The fly lost to a Gaussian process | Two variables changed at once; with one, a draw | E33, E37, E44 |
| era 3 | Whole-brain complexity is needed to generate the search | A 7 × 8 affine map searches the same | E47, E49 |
| era 3 | Seven survey qualifiers | Laves phases a bcc model cannot see | E54, E56 |
| era 3 | Every mixing energy reported | Strained references, 226 ± 43 meV/atom too negative | E53, E55 |
| era 3 | Results before the plasticity fix | The rule's sign was inverted | E51 |
| era 3 | Tungsten freezes the alloy through its migration barrier | Controlled series: corr(W, E_m) = −0.03 | E62, E64 |
| era 3 | The expansion fails outside its fitting box | The comparison mixed conventions | E65 |
| era 3 | Ordering-temperature error bars | Model spread dominates; 1.7–2.1× too tight | E68 |
| era 4 | Ordering temperatures before the corner refit, as numbers | The refit moved them by up to 265 K | E80, E82 |
| era 4 | E75's DFT comparison | It subtracted cells holding different atoms | E77 |
| era 4 | Nine-element nickel expansion results | Two reference conventions in one fit | E83 |
| era 4 | E84's twelve-element hull | A bcc ordering in an off-lattice hull | E84, E85 |
| era 4 | Ta–W as the twelve-element answer | Sparse training data, not chemistry | E88, E89 |
| era 4 | Kinetics makes high-entropy alloys competitive | Activation-energy proxy over-predicts up to 1.58 eV | E91, E92 |
| era 4 | Mo–Ta is the best composition in the space | The screen rewarded the ordering hazard | E96 |
| era 4 | The ordering gate as a clearance | Unsafe off stoichiometry, and at equiatomic (0.09–0.77) | E99, E142 |
| era 4 | Cr–Ni is competitive and was never reported | It was reported; a 168 meV/atom model error | E100, E103, E104 |
| era 4 | MAP-Elites costs 99 meV/atom | Implementation defects | E114, E115 |
| era 4 | A rank-deficient covariance broke cross-entropy search | The sweep it called for falsified it | E116 |
| 16 Sep | 55 feasible compositions | 12, then 2, once withdrawn instruments were honoured | E117, E131, E132 |
| 16 Sep | The hull change dominates eight- vs twelve-element numbers | Median 0.0 meV/atom on refractory compositions | E120, E126 |
| 16 Sep | Bottom rung 44.6 meV/atom wrong against DFT | Relaxation convention; 8.1 when matched | E128, E129, E130 |
| 16 Sep | The feasible set is 2 | An artefact of a hand-built grid | E133 |
| 16 Sep | A Mo–Ta–Cu family is new | Mo–Ta with 4 % copper | E133, E136 |
| 16 Sep | High-entropy alloys fail the requirement | Sampling artefact; the gate was wrong about 14× | E134, E135 |
| 16 Sep | Rung 1 does / does not depend on its ceiling | It does when the transition is near or above it | E138, E141, E143 |
| 16 Sep | The novelty pressure result | A bug zeroed the qualifiers; confounded | E140, E144 |
| 16–18 Sep | The first eCE training sets | No ordered cells; displacement and volume in the labels | E149, E150 |
| 16–18 Sep | The expansion is 45× too flat | Correction error on random binaries | E150, E151 |
| 16–18 Sep | Reading LAL and DNa explains the gain | 99.6 % of the readout weight on output neurons | E159, E161 |
| 16–18 Sep | The head sets each walker's gain | Shuffled gains did as well | E167, E170 |
| 18 Sep | The eCE compresses formation energies 2–4× | Regression references, then fitted Hf/Ti/Zr references | E176, E180, E183 |
| 18 Sep | The fan-shaped body learns | Site inert: scores moved by exactly 0.0 | E171b, E178, N6 |
| 19 Sep | Multi-arm comparisons run in one process | Reward coupled to process history (247ed9d) | E199 |
| 19 Sep | Elitist ≥ fly on every column | True only on the gated reward | E195b, E201b |
| 20 Sep | Local search and imagination fix the fly | Both falsified | E202, E203, E205 |
| 20 Sep | The head's live gain field helps | Worst of three on both rewards | E185, E198, E198b |
| 20 Sep | Refits would raise rung 1's temperatures | Both moved them down | E190b, N4 |
| 21 Sep | Rung 1 reads half the published value (the model) | pyeCE's sampler runs at twice the temperature | E215, E215c |
| 21 Sep | The protocol supercools | Reversible on all three systems | E207 |
| 21 Sep | v5's ordering is nearest-neighbour; the 1409 K yardstick | Collinearity artefact; a real second-shell coupling | E213, E220 |
| 21 Sep | SRO columns read as α | The column is (α + σ)/2 | E214, E220 |
| 21 Sep | Published Mo–Ta 600–1000 K | From prose; Kim & Widom report 2020 K | E207 |
| 21 Sep | The thermostat ceiling censors T_c | Falsified | E214 |
| 22 Sep | Sobieraj et al. order Cr–V in Cr–Ta–Ti–W | Their table names Cr–Ti | E222 |
| 22 Sep | A per-group VASP-vs-QE offset | No offset where the frames overlap | E228 |
| 22 Sep | Deploy v7 | Duplicated rows; cost on RHEA's frame | v7, v7b |
| 23 Sep | 18 training rows as recorded | Written as the wrong decoration | E234 |
| 23 Sep | E236's re-walk on embedding 9 | 55× slower to sample; stopped | E236 |
| 23 Sep | E240's floor passes | Two defects in rung 1's floor handling | E240 |
| 23 Sep | Survivor Mo₅₂W₂₅Ta₂₂ | Fluctuation peak inside the window | E240, E241 |
The operator's rule for these entries (21 September): anything withdrawn has to be fixed. The withdrawn entries stay in the log as they were written, with the correction beside them; each is on the Experiments pages.
