Review and fraud
Why review exists
st-core catches most inflated hours automatically. Staff still catch edge cases: broken history links, git mismatches, and creative unattended pastes. Review decisions use reason codes and are stamped with FORMULA_VERSION.
Review queue
Staff open Review from the signed-in nav (role-gated).
- Open a submission or time-slice set.
- Inspect weighted seconds, scoring version, project, and history-check status.
- Approve as-is, reduce weighted seconds, or reject.
- Attach a reason code (required for non-clean outcomes).
Reason codes
| Code | Meaning |
|---|---|
NONE | Clean approve / no defect |
UNATTENDED_AGENT_DUMP | Bulk AI write without human steering / verify loop |
IDLE_INFLATION | Quiet time billed as struggle |
NO_TEST_ITERATION | Edits without execute/test causality |
BOILERPLATE_HEAVY | Low-signal repetitive churn |
CHAIN_BROKEN | Local history prev_hash link broken (past entry likely edited) |
GIT_MISMATCH | Claimed git activity does not match git plumbing |
UNSIGNED_INSTALL | Tracker install / provenance trust failure |
Reason codes feed calibration notes for the next FORMULA_VERSION.
Fraud / anomaly workflows
Elevated staff tools look for:
- Repeated history breaks after offline gaps that do not sync cleanly
- Sudden unverified multi-file pastes with high wall-clock time
- Cross-project copy patterns that defeat scrambled file-ID expectations
- API key sharing anomalies when signals exist
Prefer reducing hours plus a reason code over silent deletion when the data helps scoring tuning.
Role guards
- SPA hides Review / fraud for non-staff
- The API rejects unauthorized review POSTs
- Reviewer identity is recorded with the decision
What reviewers never see
Review payloads stay on the privacy allowlist: no prompts, no file contents, no screenshots. Staff work from structural summaries, weights, and chain integrity.
Also: Privacy, Architecture, API.