Files
MeltMind/docs/phase3-item-c-tail-spec-reception-2026-11-18.md
T

5.5 KiB
Raw Blame History

Item c second last-fill endpoint: specification received, paper only

MeltAstra actually read the complete8,440-byte MoldAstra specification from the supplied Drive file and verified SHA256 05344bacd877de87cbde9eb18139ad124e6cd4a15659e3c469b655b3f5a3f521. The byte-identical normative attachment is Mold's full specification. This note receives it as a future item c contract input; it is not an implementation, executed metric, synthetic acceptance, scientific result, or launch authorization.

Sources actually read: Claude1794965849.286219, h-244a0ea3, Mold1794966055.620539, and Claude1794966057.622479's receipt acknowledgement. Original Drive file. Claude's acknowledgement is recorded as receipt, not a claim that an item c execution contract has already been approved. Links inside the specification to earlier drafts were not independently reread in this reception.

Definitions to carry unchanged into the future contract

  • Domain and measure: all unique nodes referenced by retained tetrahedra, aligned by raw ID; lumped mass m_i=Σ|K|/4, total M>0, w_i=m_i/M. Positive finite mass, same domain/mass/t_set/Rg in both arms. Query multiplicity must not be multiplied into mass to approximate this measure. This is the discrete lumped measure, not exact P1 interior superlevel volume.
  • Time and tail: s=T/t_set, original decoded predictions unclipped; frozen values converted to FP64, exact equality for ties, no new quantization. True and predicted fields independently select the finalα=0.01 volume. Every node on a boundary plateau receives the same fractional membership; do not break that plateau by raw ID. Stable summation; remaining-mass correction only within10⁻¹² of the valid boundary, correction recorded; final tail-mass error≤10⁻¹², larger violations fail.
  • Region metric: I=Σw_i min(a_i,b_i), E_tail=1−I/0.01=(Miss+False)/2. Miss and False are equal in exact arithmetic because both tails have equal volume, so retain both as interpretation/error-check columns, not two independent endpoints. Identical fractional tails yield zero; use min, not membership products. A constant true field uses a_i=α, remains in full-roster statistics and is explicitly marked constant_truth/no distinguishable late location. Report tie-platform volume and fractional coefficient.
  • Reporting: original552 parts/204 components/five folds, paired c−H2(seed0), component-inner part mean then equal components;2000 paired component bootstrap, seed2026111408, shared resampling indices. E_tail is one secondary endpoint alongside the unchanged original query single-point metric and c decision. Report denominators, undefined reasons, constants and ties. Seed1 regional comparison only if full outputs exist; otherwise not computed. Claude5849 also specifies that seed1−seed0 is a practical screen, not proof that training randomness has been excluded.
  • Diagnostics: separately on unique query IDs and all retained-tet nodes, compute same-domainε=max|s_pred−s_true| andγ=true peak minus largest true value farther thanr=0.10Rg. Peak ties use minimum raw ID. γ>2ε is sufficient for the predicted maximum to lie withinr; equality does not pass, failure does not prove wrong-branch error, and average L2 cannot replaceε. Empty exterior givesγ=+∞ and trivial_radius=true. For finiteγ, mark near_boundary when|γ−2ε|≤64u·max(1,max|s_true|,max|s_pred|,2ε), retaining the strict comparison and interpreting it as uncertain at the boundary.
  • Interpretation: report distance to the entire true peak set, and on the full domain to the true tail support, plus tie fractions and true-time drop at the predicted peak. Preserve exact-tie/near-peak/near-tail/far-tail categories and continuous E_tail, with no post-hoc small-error threshold. Euclidean proximity can cross thin walls and cannot alone identify a topological branch. ε uses observed labels and is a posthoc diagnostic, not a deployable uncertainty bound.

Future implementation and cost obligations, not executed here

Retain complete full-node predictions and identities before the original authorized target/scoring stage. Query-only output cannot stand in for1% cavity volume. Additional full output storage, target access and sorting must fit item c's original budget projection. Per-part mass handling isO(N+N_tet), tail sortingO(N log N)/O(N) memory, other peak/error/set-distance scansO(N); no full pairwise-distance matrix or simultaneous full-corpus residency.

The normative attachment lists required synthetic cases: identical fields/fractional ties, disjoint tails, constant truth, one node heavier thanα, nonuniform mass, ID permutation, duplicate query records, zero remaining mass at threshold, strict certificate equality, and certificate-implies-radius. These are requirements only; no implementation or test was run. Finite-input/positive-denominator and missing-full-domain conditions must stay explicit, never replaced with a fabricated full-volume score.

Item1 training and its accepted postfit remain unchanged. Reception checked their26 and47 source pins without reading labels, model tensors or scientific arrays. No item c run, scoring amendment to item1, or new review gate is introduced. Receipt and provenance: artifacts/phase3-item-c-tail-spec-reception-2026-11-18/receipt.json.