# 091 validation record

2026-10-04. Real target model was used throughout; no synthetic network replacement.

1. Read existing environment and reused exact model/SST2 cache. Initial sandbox HTTP request failed to reach configured proxy (curl exit7); authorized anonymous public download then succeeded. AG News API resolved eb185aade064a813bc0b7f42de02595523103ca4; train/test/card hashes locked,20 total resources.
2. Initial preprocessing smoke exit1 on news row112808 (422tokens>256). No model inference/training before failure. Original split/log/failure preserved in initial_preprocess_failed; only trace root anonymized as <workspace> in distributed stderr. Protocol amended before inference to news512 with same rows. Target stays256, evaluation512, no truncation.
3. Successful smoke exit0,14steps/40generations,19.1148035seconds/RSS5302763520B. preflight.json recorded conservative extrapolation×20=382.29607seconds<1200s, RSS<6442450944B before formal run.
4. Formal exit0,224steps/800generations,223.4107753seconds/RSS5634686976B. Four real538MB checkpoints private, nonzero q_proj changes; stages include download/preprocessing/training/evaluation. All preset arms complete. No confirmation inference, no repeats chosen by score.
5. Full raw audit:800outputs,64training rows reencoded,224step schedules and budgets, class mapping, prompt/token/labels, per-task statistics and paired transitions, code identity; max manual CE error2.384185791015625e-7. All raw records privately retained before export. Long news echoes omitted from public copy along with reversible output IDs; distribution.json provides exact row keys/hash/count. Post-export audit reports the exact number of fully reparsed vs redacted records and rebuilds aggregates using retained flags for the latter; it cannot independently see their private texts.
6. Four Transformers files and PyTorch SGD re-downloaded from official fixed commits, byte-identical to installed packages; licenses byte-identical. Upstream version/SHA/functions in SOURCE_MAP.md. Self-written scripts complete, original copied helpers from090 included in package. Execution-source hashes match recorded identity; audit/export/docs added after training, not mistaken for the exact inference-source set.
7. This study encountered a substantive floor: AG News baseline8/32,30outputsA. No task/prompt replacement after observing it. replay25 does not improve either primary count; replay50 harms both. No claim of established general retention or catastrophic forgetting. Selection rule chooses target_only on development only. Later cross-template/human/multi-seed/confirmation work is unexecuted.
8. Archive and public-source validation records live outside code to avoid self-referential hashes. Fresh extracted/public copies run syntax, independent audit, resource hash checks and real four-arm smoke; no need to rerun full800generations for each packaging check. Checkpoint byte hashes may vary by serialization filename, compare saved step values/prediction tokens for smoke equivalence. These validations do not establish cross-platform bitwise portability.