Repository navigation
FALSIFY-GPU-010 prose: an unreported completion does not null the backend when another ran on the CPU #5045
Description
contracts/aprender/apr-gpu-backend-v1.yaml FALSIFY-GPU-010 (added by #5005, #3719) predicts that backend.ran and backend.fell_back "are null when a completion went unreported or no driver ran".
BackendTally::report() (crates/aprender-orchestrate/src/agent/driver/served_backend.rs) does not do that. One CPU completion decides the run whatever the others said:
completions' used_gpu |
ran |
fell_back (GPU forced) |
|---|---|---|
any Some(false) (e.g. [None, Some(false)]) |
"cpu" |
true |
all Some(true) |
"gpu" |
false |
no Some(false), at least one None |
null |
null |
| no completion | null |
null |
The code is the intended behaviour (a CPU completion is a fact; a GPU verdict needs every completion to say so). The prose is wrong, so a reader of the contract would expect null where the document prints "cpu".
Fix: restate the prediction to match report(). This is a prose-only change: no test, no code, and no change to what any gate accepts.
Found in review of #5005 (follow-up recorded at its C314 sign-off).
Activity
noahgift commentedon Oct 10, 2026
Parked from S1 (#3863): S1 is at 5 open children against a budget of 4. This ticket is the newest, has no pushed branch and no open PR, and had no milestone (set to 0.71.1). It stays open with its owner and re-enters when an S1 child closes (1-in-1-out). Nothing reads sub-issue membership (checked: no code, workflow or contract), so no gate changes.
noahgift commentedon Oct 14, 2026
merged into #5312 (master ticket for epic #3994; triage rule 7b, 2026-10-14). It is a checklist row there; reopen this issue when work on it starts or its PR opens.
noahgift commentedon Oct 14, 2026
Moved: this sub-ticket is now a row in #5325 (apr on GPU: refuse by name before OOM, never fall back to CPU silently), split out of #5312 by theme.