rqgm-editable-parent-selection / reports /operational-rules.md
BurnyCoder's picture
Publish the verified paper-review protocol before model inference
387a4cf verified
|
Raw History Blame Contribute Delete
1.8 kB

Operational details fixed before inference

The week starts at campaign launch. At each new confirmatory group, reserve the slowest pilot per-item time for the root and every earned or proposed endpoint, including token-matched endpoints. Add the projected search time for that group. Attempt the first primary pair even if the estimate is pessimistic. Before further groups, if their search and earned scoring no longer fit, write the time-only stopping inputs and jointly freeze the completed configurations. Keep incomplete coverage descriptive. These rules clarify the timing-only choice in the original registration.

One scheduled validation count allows one expansion window. An invalid edit consumes that window and its model tokens. Fallback attempts consume time and a common per-run allowance of one additional selector execution per validation-budget unit. Following exhaustion or loss of the worker, the harness can execute the known uniform root draw on the same random seed. Every attempt is retained. The final validation budget continues to count observed scientific outcomes. Model service errors retry the interrupted logical event, preserving the original error in the local ledger.

Search runs share an identical validated model-call cache within the relevant split compartment. The fixed reader can reuse an identical passage across compartments. Role-specific inputs, decoding settings, backend version, model digest and compartment contribute to the key. Report actual and logical compute separately. Publications contain curated public run records. Private traffic and original manuscripts remain local. The public GitHub Pages explorer updates at published milestones; local search observations refresh on each scheduled validation and expansion.