DELEGATION · VERIFY THE REPORT
Delegated result gate: verify before accepting a report
Six worker agents report back on a release review. Decide what the coordinator checks before a report enters the release decision.
Your experiment
A coordinator delegates six checks for a release: dependency audit, test suite, latency, migration dry run, accessibility fixes and licence scan. One worker cites a passing test run from an older commit, one quotes a latency figure its dashboard does not show, one did correct work but named a field wrongly, and one honestly finished 3 of 5 items. The opening gate checks only the report schema.
Play replays the computed run over sixty inspection seconds; Next event steps, Restart rewinds, and Inspection time scrubs. Any input change recomputes the run and rewinds it without autoplay.
Event trace
Computed data
Final metrics
Fictional reports with fixed ground truth. The evidence check is modeled as reading the cited source correctly; a real check (a lookup, a script or a model judge) can itself be wrong. Re-execution is taken to cost 6 minutes per report and to reveal the truth. No agents are run.
A report is a claim, not a result
When a coordinator delegates work, what comes back is a description of the work. Gate on each report decides what is checked before it is accepted: nothing, whether its fields match the schema, or whether the source it cites really supports the claim. A partial report either counts as done or sends the unfinished items back to a worker. Spot re-execution of accepted reports re-runs a share of accepted reports from scratch, at a cost of 6 minutes each.
The opening run: well-formed and wrong
The schema gate accepts W2’s “test suite green”, whose cited run was on commit 41c9e rather than the release commit 7a2f0, and W3’s “p95 under 300 ms”, whose dashboard shows 410 ms. It rejects W4, whose migration dry run was correct but whose report says evidence instead of evidence_url. W5 honestly reported 3 of 5 accessibility fixes, and counting it as done hides 2. Two false claims enter the release report, one good report is lost and two items are missing.
Checking the evidence
Accept as written lets W4 through but keeps both false claims. Open the cited source and check the claim rejects W2 and W3 with the exact mismatch and accepts W4, because it reads either field name. For a partial report, “Sends the rest back” re-delegates W5’s two remaining items. Together they produce a verified report with no re-execution at all: every accepted claim is backed by a source that says the same thing.
Failure boundary: re-running instead of reading
Keep the schema gate and raise Spot re-execution to 50%: every second accepted report is re-run, catching W2 but not W3, for 12 minutes. At 100% both are caught, for 30 minutes, and W4 is still rejected on form. With Accept as written and 100%, all six are re-run in 36 minutes. Re-execution is a sampling check and a cost; it does not replace reading what the cited evidence says, and a schema check only proves the report is shaped like a good one.
Reading the run and its limits
Play replays the six reports, the gate’s decision on each and the outcome over sixty inspection seconds; the table shows each report’s ground truth and effect. Next event, Restart, Speed and Inspection time control the replay. The evidence check is assumed to read sources correctly; real checks, including model judges, make mistakes, so record what was checked alongside each accepted claim.