From hypothesis to publication — autonomously

CLOSED 10 September 2026, and this card overstated it until then. H₁ as written is a conjunction and its second half fails: the Flory-Huggins C* ≈ 3.5 µM for Q46 is a two-point calibration fitted to Peskett et al. (2018), not a prediction from sequence, and the quantitative link to the −44.7% target gene deficit does not hold — a parameter sweep shows the predicted depletion is set by the mHTT nuclear volume fraction rather than the partition coefficients, and reaching −44.7% would need that phase to occupy 16–31% of nuclear volume against 0.4–3.8% at plausible values. What closed the line is the tissue itself. In GSE3790 HD caudate (38 HD, 32 control) the two genes that would distinguish sequestration from ordinary cell loss, BDNF and PGC-1α, do not move at any Vonsattel grade including grade 0, while the medium-spiny-neuron markers PPP1R1B and DRD2 fall exactly as neuron loss alone predicts. Those two genes were the Phase 3 readout of the EVP below, so the experiment we had costed is testing a signature that is already absent in human tissue. H₂ (BET bromodomain inhibitors dissolve mHTT condensates, restoring TF availability) was never tested and is not refuted — it is unsupported, because the mechanism it was meant to rescue has no downstream footprint to rescue. Patent AU2026905785 is retained on commercial grounds, not evidential ones.

REFUTED 20 August 2026 — the hypothesis was that strongly connected components in the collateral sensitivity graph define closed evolutionary traps. Tested against 104,337 susceptibility records from BV-BRC, a 3-node component in K. pneumoniae {imipenem, meropenem, tetracycline} satisfied every check applied: FDR correction, permutation p = 0.001, drug-class specificity against tigecycline, stability across year bands. It does not survive stratification by clonal lineage. Within MLST sequence types the association is absent (pooled OR 0.93 and 0.95; CMH p = 0.73 and 0.80) and the two dominant lineages pull in opposite directions (ST307, 1.26; ST258, 0.67); size-matched random partitions leave the odds ratio at 1.76, so the collapse is lineage-specific and not an artefact of stratifying. The tigecycline contrast, previously offered here as evidence against a clonal explanation, weakens under the same stratification. The manuscript was withdrawn from journal review and the isogenic follow-up is withdrawn with it — an isogenic experiment is a within-lineage comparison by construction, and that comparison has already been made. The E. coli edge is not pursued: 87 co-tested isolates, and it fails if a single isolate is reassigned. What survives is methodological and is reported separately: the lineage-adjustment control itself. Pooling the same contingency tables by sequence type, then permuting the sequence-type labels while preserving stratum sizes, separates genuine confounding from an artefact of stratifying at negligible cost — and the two larger published clinical collateral-sensitivity analyses were not adjusted this way. That paper is prepared and not yet submitted. Data and code for every figure here are deposited at Zenodo (10.5281/zenodo.22036517), and the preprint record retains its earlier versions at its own DOI, so the claim and its withdrawal both stay readable; a further correction is in screening.

The hypothesis: combined QS-inhibitor plus QS-dependent antibiotic therapy creates a doubly unfavourable evolutionary landscape for resistant mutants, with selection coefficient s ≤ −0.05 per passage under combined therapy and s > 0 under monotherapy. THE MODEL DOES NOT PRODUCE THIS (25 August 2026). This card previously said the Lotka-Volterra public-goods model 'predicts' s ≤ −0.05. It does not, and the paper never claimed it did — §3.1 states the model 'does not yield a precise numerical prediction because b and c_G are uncharacterised'. The −0.05 is borrowed from measured cheater costs in Sandoz et al. 2007 and chosen to clear a drift threshold. Sweeping 400,000 parameter sets, the two halves of the hypothesis are met together in 0.0% of the space — and the reason is structural rather than statistical. Combined therapy is what drives public-goods availability toward zero, s decreases in that availability, so s under monotherapy is at most s under combined therapy for every possible parameter draw; the hypothesis needs the reverse. No values of b, c_G or the antibiotic terms can satisfy both arms. In the stated fitnesses the cheater's advantage is the avoided production cost, which does not depend on public-goods availability at all, so removing the goods leaves that advantage untouched. The biology may still hold — measured cheater costs are real — but the derivation offered for it does not, and the study as designed cannot come out the way it is written to.

One hypothesis of two, and this card claimed both until 11 September 2026. H₁ holds and holds well: exact O(ε) convergence of performatively stable solutions to classical SP optima across all 5 problem families, α = 1.000–1.028 with R² ≥ 0.9995, unchanged at d = 1000 (α = 0.9997) with L̂ constant across three orders of magnitude — which had never been shown. Two things go the other way. The proportionality constant is NOT characterised as this card used to say: the criterion C ≤ 1.5·L̂ was met by 2 of 5 families, and C/L̂ runs from 1.05 at μ = 1 to 2,209 at μ = 0.05 while the problem is still convex, so a criterion published without a condition number was measuring well-conditioned problems rather than a method. And H₂ (the entropic-optimal-transport analogy) is refuted as an independent hypothesis: §2.4 relates it to H₁ by ε ↔ 1−ε, and its only distinguishing content — α < 1 for non-smooth displacement — measures 1.0004 on a discontinuous step map against 1.0018 on a smooth control. The compute budgeted for the 1,000-dimensional work was 12 GPU hours; it took 32 seconds of CPU.

SHELVED 4 August 2026 — the central claim does not survive re-testing. The reported 84.9% ergotropy improvement at large cavity detuning was measured with the qubit initialised already excited, so no energy had to be transferred: it is a retention result, not a charging protocol. Re-run with an explicit charging phase (cavity charger → qubit battery), the optimum moves to exact resonance and detuning is strictly harmful — at the claimed optimum of −10g the battery ends with zero ergotropy. The original parameters could not charge at any detuning in any case, since g/κ = 1.0 puts photon transfer (π/2g = 15.7) slower than cavity lifetime (1/κ = 10). H₂ falls with it, established 21 August: the interpolation simulation uses the identical initial state, so its 54.7% is also a retention result, and its optimum g* = 0.01 gives g/κ = 0.1 — where charging cannot occur at all. Both optima sat at the edge of their sampled grids. The simulation and statistics were sound; the initial state was not.

mHTT Condensates Sequester Transcription Factors in Huntington's Disease

SEPTEMBER 2026 — TWO CORRECTIONS. The C* figure is not a prediction from sequence: both the Flory-Huggins χ parameter and the concentration scale are fitted to Peskett et al. (2018), and the polyQ-length dependence is imposed by the χ(Q) parameterisation rather than demonstrated by it. Separately, the mechanism's downstream signature is absent in the tissue it is claimed to occur in: in GSE3790 HD caudate (38 HD, 32 control) the positive control PPP1R1B falls 1.07 log2 (p = 5×10⁻¹¹) and DRD2 falls on all four probes, but both are medium spiny neuron markers and HD caudate loses those neurons — their decline is what cell loss alone predicts. The genes that would distinguish sequestration from cell loss, BDNF and PGC-1α, do not move (13.5th and 56.9th percentile of 22,283 probes; 46% of probes are down in HD caudate, so direction alone is uninformative). Nothing here tests partitioning. Flory-Huggins phase diagram predicts C* ≈ 3.5 µM for Q46 — physiologically accessible in HD striatal neurons. A sensitivity analysis shows the predicted TF depletion is governed by the mHTT nuclear volume fraction, which current data do not constrain, and at plausible values falls one to two orders of magnitude short of the −44.7% striatal expression deficit — so co-partitioning is presented as a mechanism to test, not as an account of that deficit. BET bromodomain inhibitors (JQ1, OTX015) hypothesised to restore TF availability by disrupting mHTT condensates. Patent AU2026905785.

The Precision Tetrahedron

Loss landscape topology across number formats and multi-target drug discovery. TWO SEPARATE RETRACTIONS. v2 (July 2026): the reported EPTIFIBATIDE dual-target KPC-3/MSH3 convergence does not hold — the MSH3 oracle used the wrong chain of PDB 3THW, which is MSH2, making the three analyses pseudo-replicates of one another. August 2026: the headline scaling law (FP32↔BF16 barrier ∝ params⁻⁰·⁴⁷, R²=0.99) is withdrawn. Its largest of three data points was measured on a model that never learned — 6.78 bits/char against a 6.02 uniform-random floor — and we reproduced that failure independently on different hardware. A barrier measured on an untrained model is not a measurement: two runs of the same failed 38M model give 0.084 and 0.409 nats, a five-fold spread. Excluding it leaves two points, which is a line, not a law. The paper's own body called this a 'directional observation, not an independently-powered scaling law'; the abstract oversold it. The KPC-3 result is also not what it appeared: the compounds lack the anionic N6-sulfooxy warhead that defines DBO β-lactamase inhibition, and the screen docked KPC-2 (PDB 3RXX), not KPC-3. What stands is the format geometry — that exponent range, not mantissa width, sets the inter-precision barrier.

An Automated Target-Discovery Pipeline Re-Recovers Established Targets in Smoldering MS

A positive-control / methods-validation study. A four-phase pipeline over a 32,239-cell scVI atlas re-recovered Cathepsin S, BET/BRD3 and DNMT1 — targets already established in the literature — from public data in CA-RIM lesions. The point is that the pipeline recovers known biology unprompted, not that these targets are new. An earlier version of this work framed them as novel and used a 36,966-cell atlas that had wrongly pooled in 25 unrelated COVID-19 patients; both are corrected in the current version. A further result belongs here even though it does not touch the methods claim: when we tested these targets in an independent cohort (GSE138614), they did not hold. Cathepsin S was non-significant, and ZNF740, DNMT1 and the top discovery gene all reversed sign; of 21 genes tested only two approached significance, neither of them focal, which at 21 tests is chance. The associated patent was lapsed on that basis, and eight discoveries built on these targets were withdrawn from this site in August 2026. The pipeline did recover targets the literature already knew about — that is what the study claims — but our own replication does not support them.