Volume 12 - The Crucible: What Survived the Fire - Five Forms of Error and the Rebuilding of the Instrument

Volume 12 is not an expansion of the KishLattice survey. It is an audit of the instrument that produced it, and the audit found five distinct ways in which our own pipeline could manufacture a result that looked like physics. The count of domains reported as STRONG fell from twenty-eight to nine. One headline result — stellar transverse velocity locking at 15/π–16/π, carried in this programme since Volume 5 — is withdrawn. Two of four falsification controls turned out to be demonstrating scale annihilation rather than projection wrongness. A designated null was found to reproduce the very domain it was meant to null. The register assignment of every domain in the survey was shown to depend on a choice of physical unit that had never been pre-registered. A pre-registered falsification condition from our own prior work appears to be met on our own published record, and is reported as such. None of it was found by a critic. All of it was found by running measurements designed to test hypotheses that turned out to be wrong. What this volume establishes Five artifact forms are named, each with a mechanism, a demonstration, and a screen that now runs on every domain on every run: Form A (transform artifact), Form B (mundane concentration, demonstrated here by showing that pure Rydberg 1/n² and Z² scaling produces a register peak with no lattice present), Form C (preprocessing-induced concentration), Form D (null-support mismatch), and Form E (quantised support). Two structural limits are also identified, and they are not defects of the instrument but properties of the data that had never been measured. The first is a resolution criterion: to distinguish register N from its neighbour, a domain must span log-range greater than N² × 0.021590. This is the standard Rayleigh criterion, it has been true of every volume this programme has published, and it was never checked. Four of the nine surviving domains cannot resolve their own reported register. The second is a degeneracy: the register index and the scalarization modulus enter the statistic only as a product, so the data can identify a preferred product but cannot split it into a register and a coordinate. The volume also introduces and pre-registers a second one-dimensional engine that tests the period of log x rather than its phase. Four of the five artifact forms trace to the fact that the existing statistic measures phase — a quantity fixed by our choice of unit and offset, which no physics constrains. Under the period statistic, unit invariance is an algebraic identity rather than an empirical robustness claim, and the promoted data product becomes framework-agnostic: the lake stores raw values, and the modulus enters only at test time. Direction of effect Every artifact form identified here biases the measurement against the framework. None is capable of manufacturing a positive peak from nothing. Every one of the twenty-eight STRONG domains in Volume 11 was scored into a headwind. Over five weeks of engine work the ledger went 28 → 21 → 9, and one number moved up: the surviving anchor strengthened at every tightening of the instrument. What survives One domain — 189,330 NIST atomic transitions — has passed every screen the programme could invent: continuous support, top-level field resolution, zero below-floor mass, immunity to the log-2 commensurability family, register stability across independent unseeded runs, and the highest resolution margin in the survey. It is also not yet safe: a Rydberg null that could account for it has been pre-registered and not yet run, and the volume says so. Errata and pre-registrations Six errata are filed, dated, with the referee's rulings attached — including two cases where the referee overruled the architect, and three where the architect corrected himself. All pre-registrations, including the period-engine trial set and the six-lake 2D calibration protocol, were filed before any code was written. The volume adopts the look-elsewhere correction as an entry gate for every domain, requiring a local z of 6.22 across a trial space of 23 registers by approximately 50 domains. Seven standalone audit scripts are released with the volume, each reproducing a corresponding figure or table from the published lakes. On what is claimed The evidence today is one domain. That is a lead, not a framework. The cross-domain register agreement reported in Volumes 5 through 11 does not currently survive the resolution criterion, and the bridge programme is in question until its domains are rebuilt with fuller dynamic range. All of that is stated in the volume rather than left to be inferred. This work lays the foundation for the two-dimensional engine, where the scramble null — inert in one dimension, as Erratum 1 established — finally becomes load-bearing. Twenty-eight went into the fire. What comes out will have earned it.