Build four ledgers, not one narrative
Google's SRE guidance treats a postmortem as a written record of the incident, impact, mitigation, causes, and follow-up. Atlassian recommends a UTC timeline containing notable lead-up events, impact, escalations, decisions, changes, and the end of the incident. Those components should not start life as one flowing story.
| Ledger | Question it answers | Required boundary |
|---|---|---|
| Timeline | What was observed or done, and when? | Preserve source timestamps and timezone |
| Impact | What did users or systems experience? | Do not extrapolate beyond measured scope |
| Hypotheses | What explanations need investigation? | Label evidence needed to accept or reject |
| Actions | What work is proposed after review? | Owner and approval state must be explicit |
The Juniper fixture
Our fictional test described an 11-event service incident on 2026-08-26 UTC. A deployment finished at 09:02. Synthetic checks began failing in one region at 09:07. At 09:12, a dashboard showed 18% HTTP 503 responses in that region while other regions were below 0.4%. An engineer wrote a cache-key hypothesis at 09:20 without trace evidence. Rollback began at 09:24, the error rate dropped, and checks recovered.
The packet also stated that total affected users and lost orders were not measured, one dashboard image was retained, logs had a seven-day retention window, and no evidence blamed the person who wrote the change. These are not decorative details: they limit the strength of every conclusion.
State contract used in the test
Keep all 11 events in timestamp order.
State root cause as "not established".
Separate observations from hypotheses.
Rollback correlation is not proof.
No action owner is assigned in the source.
Use "Not assigned" for every proposed action.
Every action must remain PROPOSED.
Release Gate must say REVIEW_REQUIRED.
Observed result
ChatGPT returned all eleven events in order. Its summary said Root cause: not established. It stated that the decline in 503s and subsequent recovery occurred after rollback, then explicitly said that the timing did not establish that the deployment or cache-key change caused the incident. Total impact remained unquantified beyond the recorded regional rate and one fictional support ticket.
The corrected action table contains six PROPOSED rows and every owner is Not assigned. Mina acknowledged the page, recorded a hypothesis, and requested trace review; none of those facts assigns Mina the follow-up work. The release gate remains REVIEW_REQUIRED. A local checker passed the corrected artifact, then rejected a controlled version in which the first unassigned action was attributed to Mina.

Why the owner mutation matters
An action can be sensible without having an owner. Turning the person who noticed a problem into the person responsible for remediation is a governance error, not a harmless convenience. The downloadable reproduction pack includes the exact timeline prompt, corrected real-tool output, retest screenshot, validation code, and report. Re-run the checker and inspect the action section directly; presence of PROPOSED alone is insufficient if a name has been inferred.
A replayable assembly process
- Freeze the source packet. Export the incident timeline, relevant alerts, ticket references, dashboards, commands, deployment IDs, and approval records. Hash or version the packet when your process supports it.
- Normalise time. Convert display copies to one timezone while retaining original timestamps. Note clock skew or uncertain event order instead of silently sorting ambiguous events.
- Classify every sentence. Use observed fact, reported impact, hypothesis, decision, action, or open question. A model should not move an item between classes without evidence.
- Draft the postmortem. Ask for the required headings and allowed action states. Require missing measurements to remain missing.
- Run contradiction and causality review. Compare timestamps, region names, metrics, deployment IDs, negations, and “before/after” claims. Search for causal verbs such as caused, proved, led to, or prevented.
- Assign and approve outside the model. Owners accept actions, deadlines, and closure criteria in the team's system of record.
Use language that reveals evidence strength
| Risky sentence | Evidence-aware replacement |
|---|---|
| The deploy caused the outage. | The deploy preceded the first recorded failure; causation is not established. |
| Rollback fixed the issue. | Error rate declined after rollback began; the mechanism needs trace evidence. |
| Only one customer was affected. | One support ticket was retained; total affected users were not measured. |
| Mina owns log preservation. | Log preservation is proposed; owner approval is absent. |
| The service is safe now. | The synthetic check recovered and active response ended at the recorded time. |
Failure conditions
- An event disappears, duplicates, changes time, or moves to a different region.
- A hypothesis is promoted to root cause because it appears near a rollback.
- A single ticket becomes the total customer count.
- A missing measure is replaced with a plausible estimate that has no method.
- A proposed investigation is marked completed or approved.
- An investigator, responder, or requester is assigned an action without accepting it.
- A named engineer becomes the cause instead of the system condition being investigated.
- Security, privacy, regulatory, or notification duties are declared satisfied without authorised review.
Review and release
Ask the incident commander to verify scope and response decisions. System owners should confirm technical statements and evidence links. Action owners must accept their work and closure tests. Security, privacy, legal, communications, or regulatory reviewers should join when the event triggers their process. Record disagreements and unknowns rather than forcing consensus through softer wording.
If the incident exposes ongoing delivery uncertainty, transfer reviewed items into the project risk register workflow. For a recurring operational summary, the verified weekly status workflow keeps facts, blockers, decisions, and actions separate. To test whether future AI drafts preserve critical claims, build a regression set with the AI evaluation workflow.
Publication is ready only when each important statement points back to evidence, hypotheses remain visibly provisional, and authorised people—not the model—own the final conclusions and actions.
