Verification artefact package for "After the Euclidean Highway: Hyperbolic Expert AI as the Next Innovation"

This deposit contains the corrected, self-contained verification artefact package for the HySAT (Hyperbolic Structure-Aware Training) placement principle developed across six expert small-language-model projects (20,951,552 cumulative training-exposure samples, re-derived from the raw training logs and corpus manifests, with a 6,363,188-sample verified-unique subtotal plus the ORAA access-tier corpus; more than 770,000 cumulative optimizer updates, 772,484 with MetaTeach entered at its verified lower bound; six completed HySAT training runs; zero NaN events in those completed runs).The evidence is disclosed at three explicit tiers. First, automated gates in verify_claims.py reproduce the H2 manifold-drift result (3,474/3,474 recorded steps within 1e-3; 81.95% within the stricter 1e-4 diagnostic threshold) and the H5 batch-size pair-formation gate (five evaluable b >= 8 rows; zero falsifiers). Second, released traces, evaluation aggregates, interval records, trainer states, hyperparameter configurations, matched ablations, and controlled toy/mechanism tests support direct inspection of the corresponding claims. Third, manifest and research-access records identify claims whose production artefacts or historical per-run terminal records were not retained publicly.The historical negative-result record is stated as 17 documented HyperLoRA-family training collapses, comprising NaN termination or unrecoverable divergence. The package does not claim that all 17 terminal events are independently observable as NaN in deposited per-run traces. One preserved early AdmitBrain E2 trajectory directly records an unrecoverable divergence plateau through step 13,000 (total loss approximately 4,288-4,339) and contains no textual NaN/Inf in the retained portion. A later E2 trainer-state record is a separate stabilized run. The included crash_reproduce.py is a controlled mechanism test, not a reconstruction of every historical collapse.Version 1.7 completes the documentation-consistency pass begun in v1.6: the package README now carries the correct version identity (five stale v1.5 strings in the published v1.6 README are fixed); the failure-condition map names the micro-batch pair-formation floor, cites only deposited files, adds the MetaTeach v15 final-run row with its evidence tier (final launch command not retained), and keys the AdmitBrain success row to the final deposited run; the toy README states the instability criterion at family-ledger tier (per-run v-norm traces of the historical collapses were not retained); and the changelog corrects the earlier scheduler sentence (constant learning rates in the four loss-layer fine-tuning projects; BS Sovereign records a 3 percent cosine warmup; MetaTeach shows per-round scheduling). No data, trace, or executable-arithmetic changes.Version 1.6 corrects mathematical documentation only (no data or executable-arithmetic changes): a false inequality in four documentation/comment locations is replaced by the valid conservative bound sinh(r)/r >= exp(r)/(4r) for r >= ln 2; categorical sufficiency language for the three stabilisation conditions is replaced by observed-condition language matching the manuscript (the conditions are the observed stabilisation conditions of the deposited record, not proven universal sufficient conditions); and the claim map states the conditional convergence separation with its sufficient-direction and bounded-iterate assumptions. Full per-item provenance in CHANGELOG.md.Version 1.5 re-derives every per-project training denominator from the raw training logs and corpus manifests; adds condition-blinded three-judge re-scoring aggregates for the CEO Finance coach-configuration benchmarks and the S3 Creative seed-efficacy benchmark, with the H6 orchestration direction recorded as reversed under the blinded protocol and the S3 result confirmed; generalises the shared hyperbolic engine to arbitrary curvature with a deposited invariant test; adds the CEO Finance v2 production training log, the MetaTeach v15 round-level metrics, and the nine matched-run traces cited in the Supplementary Information; renames files whose earlier names misattributed their runs; and moves internal evaluation material to the research-access tier. Full details are recorded in CHANGELOG.md. These corrections strengthen the reported scope; the completed-run zero-NaN result, the matched-ablation results, and the H2/H5 automated outcomes are unchanged.