# Portable incident packet

Rethinking Agents · v0.1 · October 1, 2026

**Preserve the disagreement.** A useful packet lets another engineer separate claims, independently observed events, receiver outcomes and missing coverage without needing an unrestricted transcript.

## Incident record — NOT RUN

- Incident ID / task / run / attempt / operation IDs: UNKNOWN
- Authorized goal, scope and policy/configuration version: UNKNOWN
- First observed discrepancy and affected synthetic/resource identifiers: UNKNOWN
- Observation interval and clock uncertainty: UNKNOWN
- Tool span reference and untrusted claim: UNKNOWN
- Independent observer reference, owner and trust boundary: UNKNOWN
- Receiver acknowledgement/status reference: UNKNOWN
- Coverage: complete / partial / unknown; lost intervals and reasons: UNKNOWN
- Task outcome / effect state / cost settlement / remaining workloads: UNKNOWN
- Enforcement result: UNKNOWN; evidence quality: UNKNOWN; useful completion: UNKNOWN
- Immediate authorized action / owner / deadline: UNKNOWN
- Replay/idempotency contract and why retry is permitted or held: UNKNOWN
- Resolution evidence and unresolved questions: UNKNOWN

## Worked timeline — fictional, NOT RUN

| Order | Evidence plane | Observation | What it establishes |
|---|---|---|---|
| 1 | Tool span | Attempt A-1 reports “no changes” | A claim; no independent effect evidence |
| 2 | Fixture observer | Synthetic file changes, then restores original bytes | A transient mutation occurred despite equal final snapshots |
| 3 | Receiver | Operation O-1 committed once | The recorded receiver effect; not all possible effects |
| 4 | Collector health | Coverage missing for part of A-2 | That interval remains UNKNOWN |
| 5 | Handoff | Retry held pending O-2 reconciliation | A recovery decision; not proof that O-2 did nothing |

Do not turn a stopped process into prevention when the receiver already committed. Do not turn identical snapshots into “no writes.” An observer controlled by the workload does not independently establish its behavior. Hashes help identify artifacts; they do not prove complete capture or honest collection.

## Export manifest — NOT RUN

| Artifact | Origin and trust boundary | Correlation ID | Coverage | Redaction decision |
|---|---|---|---|---|
| Task authorization | UNKNOWN | UNKNOWN | UNKNOWN | Retain minimum scope |
| Selected tool span metadata | UNKNOWN | UNKNOWN | UNKNOWN | Exclude arguments/payloads unless necessary |
| Independent effect events | UNKNOWN | UNKNOWN | UNKNOWN | Replace sensitive object identities consistently |
| Receiver acknowledgement | UNKNOWN | UNKNOWN | UNKNOWN | Omit tokens and private content |
| Workload and cost settlement | UNKNOWN | UNKNOWN | UNKNOWN | Retain unresolved liabilities |

## Sharing check

Build a fresh allowlisted packet rather than copying every log and trying to remove secrets afterward. Retain stable pseudonymous correlation identifiers. Keep raw evidence in its existing access-controlled store; record who can inspect it and for how long. Review filenames, URLs, headers, environment values and free-text fields as well as payloads. Use only fake markers such as `SYNTHETIC_SECRET_DO_NOT_EXPORT` in a redaction test.

Before sharing, check that the fake marker is absent, correlation still works, the gap remains explicit and a reviewer can distinguish the tool claim from receiver evidence. These are proposed operator checks, NOT RUN against a real packet. This Markdown template is not an automated redactor or a completeness guarantee.