Tabnetics Diakrino Campaign Results
Scope
Tabnetics Diakrino is a trained feature-ranking sidecar integrated into the Tabnetics feature-selection pipeline. It proposes bounded feature additions to a protected classical core; the downstream classifier, split policy, and evaluation protocol remain paired across arms. This campaign evaluates that integration, not an official TabPFN classifier.
Final Campaign
The final closeout covers 64 HDLSS datasets, three paired seeds (29, 47, and 83), and eight arms, for 1,536 reconciled result rows. The primary P4 profile combines multi-view Tabnetics Diakrino proposals, native-null admission, and support-only JMI arbitration. Its protected-core baseline is B0.
P4 versus B0 has a mean balanced-accuracy delta of +0.0088, with 33 / 14 / 17 dataset-level wins, ties, and losses. Its Holm-adjusted p-value is 0.0828; it passes the later, outcome-informed strict p < 0.10 product-profile gate and the tier-harm gate. This was not the original confirmatory decision rule.
Equal-addition diagnostics remain descriptive rather than promotion-blocking: P4 versus classical-next-best is +0.0076 BA (Holm 0.3562), versus seeded random extras is +0.0055 (Holm 0.8521), and versus permuted Tabnetics Diakrino ranks is -0.0032 (Holm 0.9383). Consequently, the result supports the protected augmentation profile but does not establish that the ranker itself is the source of the observed improvement.
Decision-rule provenance
The original closeout verdict, recorded on 12 July 2026 at 09:49:45 UTC, was no_evidence_pfn_helps under the then-current Holm p < 0.05 and required equal-budget-control gates. The historical identifier is retained here to distinguish that report from later renamed reports; it does not denote an official TabPFN classifier.
T-PFN-FS-GATE-02 / #274 subsequently revised the decision policy on the same frozen recovery-v2 matrix. Its final record, at 10:43:41 UTC, explicitly calls the change outcome-informed: Holm p < 0.10, with C1–C3 retained as required specificity diagnostics but no longer promotion blockers. The initial #274 body still required those controls; the final record and signed commit 863bbebb record the completed policy. Thus the later product-profile promotion is not a new independent confirmation or evidence that neural rankings beat matched-budget alternatives. Neither nonsignificant controls nor a policy revision prove equivalence.
Original and revised reports have distinct recorded byte hashes: see #267's corrected digest record and #274's final record above. These are historical recorded results, not a fresh recomputation. Exact reproduction requires recovery of the original outcomes, bindings and consumed analyzer identity; the current renamed analyzer carries the revised rule and must not silently replace the historical one.
Publication Contract
The result matrix was reconciled against source, input, split, feature-order, checkpoint, bundle, budget-authority, and host-telemetry identities. The renamed implementation does not consume pre-rename artifacts; a future Tabnetics Diakrino claim requires fresh source-bound emissions under the renamed contract.
This page is included in the public Pages and release-snapshot generators. It records campaign evidence only; publication remains an explicit later release action.
Documentation and webpages on this site are generated from authoritative internal sources using a combination of deterministic rules and generative AI. Errors are possible. Please report issues via GitHub Discussions or email [email protected].