input/v4/DRIFT.md

Post-freeze drift log

The experiment spec was frozen as version 1.0.0 before reportable collection. The following implementation-level changes were necessary. None changes a primary metric, start grid, seed set, selection rule, or success tolerance.

D-001 — transitive dependency pin

• Detected: local clean-environment smoke test, before full collection.
• Change: pin tensordict==0.5.0 alongside the authors' torchrl==0.5.0.
• Reason: the current resolver otherwise installs tensordict==0.13.0, which is API-incompatible with torchrl==0.5.0 (interaction_mode import error).
• Effect on interpretation: none; this restores the dependency pairing from the 0.5 release line and is recorded in every environment manifest.

D-002 — explicit official-baseline wrapper types

• Detected: first full local Claims 1–3 attempt.
• Change: validate and handle the released ARS VecNormalize asset separately from the released PPO/SAC DummyVecEnv assets.
• Reason: the three official checkpoints do not share the same wrapper type.
• Effect on interpretation: none; reward normalization is rejected, and observation de-normalization is used only where the official wrapper declares it.
• Evidence policy: the failed attempt remains append-only raw evidence; the corrected attempt uses a new batch id.

D-003 — Pendulum reset sensitivity panel

• Detected: source audit before full Claim 5 collection.
• Change: add a secondary evaluation of the released Pendulum ARS checkpoint that resets through the complete VecEnv chain.
• Reason: the author's primary evaluation resets the innermost environment directly, so its first observation bypasses VecNormalize.
• Effect on interpretation: the author-exact buggy protocol remains the preregistered primary comparison. The corrected protocol is labeled as a sensitivity analysis and cannot replace the headline result.

D-004 — canonical table filename

• Detected: deterministic table-builder implementation.
• Change: use summaries.csv rather than the singular summary.csv listed in the planned-output section of SPEC.md.
• Reason: the table contains multiple summary grains and the plural form aligns with the other table names.
• Effect on interpretation: naming only. Row grain and aggregation rules are declared in the canonical manifest.

D-005 — runner provenance

• Detected: pre-HF-Job release review.
• Change: add the runner's own SHA-256 and resolved path to every raw manifest.
• Reason: the HF Jobs CLI uploads the local PEP 723 script separately from the mounted input bundle.
• Effect on interpretation: none; this strengthens traceability.

D-006 — ephemeral TensorBoard scratch path

• Detected: first Claims 1–3 HF Job attempt.
• Change: write TensorBoard event files and RL-Zoo monitor scratch files to a disjoint local /tmp tree instead of the mounted evidence Bucket.
• Reason: one concurrent TensorBoard writer received FileNotFoundError from the Bucket's FUSE/object-storage layer even though the directory had been created. The structured training metrics, captured worker stdout, and policy coefficients were already written separately as canonical evidence.
• Effect on interpretation: none. TensorBoard files are transient transport artifacts; all analysis inputs remain append-only in the Bucket. The failed Job and its error manifest are retained under their original batch id.

D-007 — local evidence staging before Bucket publication

• Detected: integrity audit of the completed v3 Claims 1–3 HF Job.
• Change: write and validate the entire append-only batch on local /tmp storage, then copy the closed files to the mounted Bucket and require a byte-for-byte recursive comparison before the Job can succeed.
• Reason: repeated open(..., "a") writes to records.jsonl on the Bucket FUSE mount lost complete append ranges while still returning success. The v3 manifest's in-memory counters declared 2,400 episodes, but the hashed file contained only 2,200; 200 baseline episodes and related records were absent. The scheduler's COMPLETED state was therefore not sufficient evidence of batch integrity.
• Effect on interpretation: none. The experiment configuration, seeds, metrics, and runner are unchanged. The invalid completed v3 batch and the canceled v3 Claim 4/5 Jobs remain disclosed; only a subsequently validated batch may enter canonical tables.