# Suite35 PACE 1.1: native AQ revision

The original1.0 package provided a portable normalized-input implementation but left the actual AQ source integration to the user. Version1.1 completes the known AQ integration using the experiment sources supplied in this conversation.

## Behavior changes

- `./run_suite35.sh` now runs actual AQ data automatically with all stages and resumable verification. An explicit demo JSON still runs synthetic data.
- Canonical schemas, finite-CF-covered sidecar positions and17 numeric course fields are handled directly. Missing sidecars reconstruct course metadata; missing canonical cohorts fall back to existing expanding split CSVs.
- Exact numeric training grades are already in `target_grade10`. Optional full `DDTBHP`/known raw-index histories are discovered automatically, reconciled by grade identity or checked tuples, and cut by each enrollment's semester. Target identities can join prior histories without inserting target grades into canonical events.
- Native chronology inherits the earlier experiments' previous-semester availability assumption. No externally prepared calendar or publication timestamp file is requested. Internal comparison carriers are identified explicitly.
- Full labeled training support is default; optional covered support reproduces that older population choice. Source/support differences are logged and all current arms share rows.
- Native matrices load one cohort at a time. Histories use compact padding and repeated-prefix preparation. Original source files are immutable and byte-fingerprinted.
- Archival metric-only baseline cells supplied by Suite32 are bundled with provenance and automatically compared on matched faculty/fold cells.
- Reports keep detailed results, support audits, subgroup findings and continuation rules. Epoch records and checkpoint replay remain mandatory. Phase reports are saved separately; combined tables print once. Repeated identical SSL reuse messages are summarized. Log capture reopens the live file per line and finalizes it atomically from an independent spool, so stale snapshots cannot remove earlier lines or strand later results on an old descriptor.

Three encoders, three SSL objectives, all eight primary controls, chronological selection/fresh refits and F1–F5/F6–F7 freeze semantics remain part of the full study. No gain on AQ is claimed without running the real study. See `verification/VERIFICATION.md` for the precise local evidence and limits.
