86 protected KIOS frames, six reconstructed conditions, 516 views.
Follow the evidence.
The original checkpoint replay matched all 24 published metric cells exactly and four prediction tables byte-for-byte. The 1.0.0rc1 candidate is retrospective evidence—not a final release, independent test or flight validation.
See the record ↗136,359 baseline and 71,420 Phase 23 predictions scored. 24 / 24 metric cells matched.
516 cross-model views and 860 within-model pairs reuse KIOS. Phase 26 is a separate prospective gate.
Hashes you can check.
These identify the retained model, input and release-candidate bytes. Matching a recorded hash checks byte identity; it does not by itself authenticate provenance, episode pairing or scientific validity. The linked receipts and protocols provide that context.
| Artifact | Status | SHA-256 |
|---|---|---|
| Official KIOS archive | Verified | 9d27a3e9616c188fb72a727633705735da5df168ce47efeed389dcb3455fdbdb |
| Phase 22 baseline weights | Verified | 3a1801b192d624f8dcdda4bc5d9a9157309000df67a4c30d62368c37901feddd |
| Original source split manifest | Separate from the protected-frame selection | 32623634069c4f3082b79a6a42bca408c636f4b626e5f7691276257999544905 |
| Protected Phase 25 frame manifest | Verified for the accepted replay | 8605cf1cbf9e8769c0324bf066078f9964f232ac3128131c8bd0b689bb7b64b7 |
| Phase 23 exact weights | Recovered; all six condition aggregates matched exactly | 43240be969708c32b3e11340c9846b3a588baf361378398677c794d16a455310 |
| 1.0.0rc1 frozen manifest | 184 files; scoped candidate validated | 157cf7b37f1793e1b85c158d89245a1b751fe441481b8083b905e8a5e0efcc6f |
| Phase 26 independent test | Method locked; test inference not authorized | NO_DATASET_ADMITTED |
Check the current research candidate.
From a clean, full-history checkout, this read-only gate verifies 184 frozen files, the original replay receipt, exact recovered source, 860 within-model pairs, saved controlled-occlusion matches and generated Atlas data. It does not retrain a detector or admit an independent dataset.
pip install -e ".[dev]"
git fetch --no-tags origin \
f090da03d20b2425c5addccb4d19117c8991bcc1 \
6a8623cbeb54951f43b8ca65293598e7c41a474e
python scripts/validate_research_release.pyThe narrower python scripts/reproduce_release.py still checks Phase 25A only. The candidate report explicitly states inputs_replayed: false: this gate checks retained artifacts rather than repeating image/model inference.
The accepted inference replay used Python 3.13.2, Ultralytics 8.4.160, Torch 2.7.0+cu118, NumPy 2.2.3 and Pillow 11.0.0. Repeating it requires the exact external archive, reconstructions, original checkpoints and CUDA settings. Disable optional-package auto-install with YOLO_AUTOINSTALL=False; never substitute a checkpoint or silently change the runtime.
Preserve each experimental chain.
The ten recovered source files are exact commit f090da03…; no historical UI branch was merged. Original controlled sweep: INCONCLUSIVE, raw clean gate FAILED, no treatments. The separate Q95 occlusion v2 freeze remains separate: its 145,169 saved detections and 516 frame/target cases were replay-verified, not newly inferred in this revalidation.
python scripts/verify_phase22_occlusion_v2.pyPhase 26 needs its own test.
The candidate screen checks every KIOS reference image against the verified archive before comparing new images. A separate development-only rule lock rejects test outcome rows. Neither step admits IMAV as an independent test or produces a Phase 26 result. None of the seven audited sources was admitted; that is not a claim that no qualifying public dataset can exist. The next-candidate execution lock requires complete authenticated screening against all 422 KIOS real frames, independent session lineage, negatives, ontology/annotation and rights review. Pass --near-distance 9 to require accepted dHash separation ≥ 10. A clear screen is not admission or permission for sealed-test inference.
python scripts/verify_phase26_admission_protocol.py