reproduction-bundle/STATUS.md

# STATUS — LOGDIFF (`OAM1jJsMGp`)

**Session:** full-score campaign. **Last updated:** 2026-07-20. **State:** in progress; official score 6/10 and exact molecular runs active. The first four full 1,000-step AND shards are validated with 4/4 chemically reconstructable, distinct molecules and 8/8 successful generated-ligand dockings; the scoped checkpoint is public while shards 4--7 continue.

Public Space: [DineshAI/OAM1jJsMGp](https://huggingface.co/spaces/DineshAI/OAM1jJsMGp) at current revision `38daa9517ef7af65d409ffd0325cba289e969014`; the official verdict below was issued against this exact revision.

Fresh official verdict: **6/10**, high quality, at judged Space SHA `38daa9517ef7af65d409ffd0325cba289e969014` (`2026-07-20T01:09:46+00:00`); C1/C5 are verified, C2 is falsified, and C3/C4 remain inconclusive. The new C4 feedback explicitly cites the 4/32 scope, the below-target `64.53` product, and the unstarted AND-NOT campaign, confirming that both complete campaigns remain required.

GitHub remains `MachineLearning-Nerd/icml26-repro-OAM1jJsMGp-logdiff-exact-boolean-guidance` at `6449752`; the current campaign changes are intentionally uncommitted.

## Anchored evidence

- C1 Proposition 3.1: verified by exhaustive finite-state oracles and real trained-UNet algebra checks.
- C2 CMNIST/Shapes3D Table 2 range statement: falsified as written by the paper table itself; only 2/8 LOGDIFF and 5/8 constant-baseline cells fall inside the stated ranges.
- C3 CelebA Table 3: table transcription matches, but the pinned author release lacks all nine declared checkpoints, has a broken config reference, and uses a different FID implementation/protocol.
- C4 GRM5/RRM1 Tables 5-6: exact targets, all three checkpoint hashes, the real TargetDiff sampler, and AutoDock Vina are recovered; full generated-ligand composition remains active.
- C5 Proposition C.2: verified over 20,650 independent-product events and 6,350 taxonomy events, with invalid overlap rejected.

## Source-scale runs

- Exact upstream CMNIST composition classifier completed 50 epochs on the full 60,000-example split.
- Fixed-timestep evaluation completed over the full 10,000-example test split at seven timesteps.
- Independent exact upstream clean-judge training and full 10,000-example evaluation completed. The checkpoint SHA-256 is `ff8dbb5253278ad46e0033959045f9558927b0e814cd04f98fc747d6629adfb6`; digit/colour accuracies are `0.9814/1.0` and digit/colour NLLs are `0.0867544/0.000025857`.
- Molecular recovery includes verified DualDiff data, GRM5/RRM1 target pair, pretrained weights, a real sampler smoke run, successful reference-complex Vina gates, and an executed generated-ligand docking/pocket-alignment smoke path for both targets. The smoke is explicitly not claim evidence because it uses 1/1000 denoising steps and has 1/32 validity. The exact Appendix-D.3 transition-alpha bug is repaired, the beta-2 restartable runner passed four-worker 80-step and eight-worker 80-step gates, and the 32 x 1000-step AND run is active with four workers and ten retries per shard. Samples 0--3 completed on their first attempts after approximately 3.5 hours each; all four have finite 1,000-step posterior trajectories, valid reconstructions, and distinct SMILES. A hash-validated partial merge (`b0cf5223...`) permits honest interim docking without altering the continuing 32-sample campaign.
- The four-ligand checkpoint completed paired `vina_full` docking at exhaustiveness 32: 8/8 generated-ligand target evaluations succeeded; mean GRM5/RRM1 affinities are `-8.00375/-8.06675` kcal/mol, mean AND product is `64.53267625`, diversity is `0.887054`, valid-unique fraction is `1.0`, and paper-defined quality fraction is `0.5`. Same-run reference affinities are `-5.866/-7.640`. The product is below the paper's LOGDIFF `73.20 +/- 3.18` and DualDiff `71.87 +/- 3.33`, so it is published as a partial execution checkpoint rather than headline verification. AND-NOT separation is not inferred from AND samples.
- Official CelebA is complete and verified at 202,599 images; the released clean-filter splits contain 100,247 train, 12,459 validation, and 12,186 test images. Public DDPM `shalpin87/diffusion_celeba@91d0ff0` is pinned locally and clean-fid 0.1.35 is installed for the independent C3 route.

## Publication and verification

- Static Space is public and RUNNING; Hub and app return HTTP 200.
- Exactly the intended 11-page hierarchy is present, with Conclusion last.
- Required tags, pinned executive summary/poster, artifact cell, and public evidence bucket are present.
- Local structural validation passes except the pre-existing legacy Space-name rule requiring a new `repro-*` identifier; no duplicate Space was created.
- Remote secret and absolute-path leakage scans pass.
- The independent CMNIST judge page was synced and read back at current Space SHA `93d90f36`; its remote SHA-256 exactly matches the local page (`840c7bef...`).
- The Claim 4 checkpoint page was synced and byte-for-byte read back at public Space SHA `38daa951`; local and remote SHA-256 are both `7ac3c77c...`. The public checkpoint artifact also read back with matching SHA-256 `c44cd570...`; its unauthenticated download endpoint returns a public Xet redirect. All 29 tests pass.

## Next

Continue the full 32-ligand AND campaign, then merge/dock it and run the separate AND-NOT generation and docking campaign. In parallel when resources permit, run the independent real-CelebA negation route. Update the evidence bundle/logbook with complete metrics and request re-judging. Do not stop on failures; materially different recovery approaches continue until the anchored claims reach their maximum defensible score.