Research
Findings
Notes from the proving ground — integrity failures that transcript-only evals miss, and what changes when agents reconcile against the system of record.
-
01
Claim
Before the write runs, lock a business-operation identity — not a random retry id.
-
02
Execute once
The tool runs. If the response is lost, the outcome is marked unknown — never assumed.
-
03
Observe
Before any retry, check the system of record. Evidence, not a guess.
-
04
Receipt
applied / not_applied / ambiguous, plus the next step. Every study below grades against this.
False failures on τ³-bench: when the cancel worked and the agent said it didn’t
Fault-injected grading on Sierra τ³-bench write tools. Scale study: 2,016 traces — afterstate 100% safe and complete; timeout slice still shows naive at 12% verified completion.
Read the study →Where Afterstate breaks
780-cell adversarial matrix. Transport faults: 100% safe. Identity collision (deliberate misconfig): 0% safe. Observer lag: safe but incomplete. Blended afterstate safety: 92% — and we published why.
Read the boundary →QuickServe × Afterstate — integrate in four steps
Simulated design partner. Before: 10 duplicate writes, $85.60
double-processed. After: 0 duplicates. Real
init --apply, verify scorecard, and receipt log.
Transport retries are solved. Agent retries are not.
Primary-source evidence that agent-level re-invocation sits above transport idempotency — and why observe-before-retry belongs in the write path.
Read the note →