Example records, by sector.
Each example shows what the operator writes, who checks it, and what the check does and doesn't show. The entries are made up.
A claims agent recommends a decision, and a person approves it
A benefits program uses an AI agent to read each claim and recommend a decision. An examiner approves or overturns every recommendation before a notice goes out.
- Who checks it
- The program's quality review, its inspector general, and an appeals reviewer.
- What the check shows
- Each entry is as it was written. The approval was recorded before the notice went out, and the model version and settings the operator recorded are on the record.
- What it can't show
- Whether the examiner read the file, or whether the decision was right.
- Where policy asks
- OMB M-25-21, human oversight of high-impact AI
- 0109:02:11input-received · claims-agent-3Claim 2026-88104 read: the application and 14 supporting documents
- 0209:02:12model-version · claims-agent-3Model version 2026.08, with the settings the program approved
- 0309:04:37recommendation-made · claims-agent-3Recommended a denial, reason code 4
- 0411:18:05approval-given · examiner-17Denial reviewed and approved
- 0511:18:40notice-sent · claims-agent-3Decision notice sent to the claimant
An agent releases a grant payment under a limit
A grants office lets an AI agent check payment requests against the award and release the ones under a limit. Anything above the limit waits for a grants officer.
- Who checks it
- The grants office, its inspector general, and the recipient's own auditor.
- What the check shows
- The limit in force when the money moved, and that the release came after the check.
- What it can't show
- Whether the documents behind the request were genuine.
- 0114:20:03input-received · payments-agentPayment request 9 received: $182,400
- 0214:20:05agent-configuration · payments-agentLimit in force: $250,000; above it, a grants officer approves
- 0314:20:41check-completed · payments-agentRequest matched budget lines 2 and 5 of the award
- 0414:21:02payment-released · payments-agent$182,400 released to the recipient
A documentation agent drafts a clinical note, and the clinician signs it
A federal health system uses an AI agent to draft visit notes from a clinician's dictation. The clinician edits and signs every note.
- Who checks it
- The health system's quality and patient safety office, and its compliance reviewers.
- What the check shows
- Which model drafted the note, that a clinician edited and signed it, and when.
- What it can't show
- What the note says, unless the operator reveals it, or whether it is clinically right.
- What stays out
- Patient data. Each entry points to the note by its reference.
- 0110:41:09model-version · notes-agentModel version 3.1, the build the health system approved
- 0210:44:52draft-created · notes-agentVisit note drafted from dictation, 412 words
- 0310:51:30draft-edited · clinician-208Note edited by the clinician, 3 changes
- 0410:52:02note-signed · clinician-208Note signed
An agent plans where supply trucks go first after a storm
An emergency management agency uses an AI agent to turn damage reports into a supply plan. A duty officer approves each plan before the trucks leave.
- Who checks it
- The agency's after-action review and its inspector general.
- What the check shows
- What the agent proposed, what the duty officer changed, and the order it happened in.
- What it can't show
- Whether the plan was the best one.
- 0105:10:00input-received · logistics-agent41 damage reports and 12 road closures read
- 0205:14:22plan-proposed · logistics-agent18 trucks: water to sites 3, 7 and 9 first
- 0305:31:47approval-given · duty-officer-2Plan approved with one change: site 12 moved up
- 0405:32:10dispatch-sent · logistics-agent18 dispatch orders sent
A model predicts a part failure, and an agent orders the part
A maintenance command uses an AI model to predict which parts will fail, and an agent to order replacements within a weekly budget the command sets.
- Who checks it
- The command's maintenance review, the depot, and the program office.
- What the check shows
- The prediction behind each order, and which model made it.
- What it can't show
- Whether the pump would actually have failed.
- 0106:00:00model-version · maintenance-modelModel version 5.0 in use, trained on data through August
- 0206:12:31prediction-made · maintenance-modelAircraft 22-114: fuel pump likely to fail within 60 flight hours
- 0306:12:48order-placed · parts-agent1 fuel pump ordered from depot stock, inside the weekly budget
- 0415:40:02maintenance-done · crew-chief-6Pump replaced; the old pump sent for inspection
A model passes test, and each site records the build it runs
A program fields an AI model after test and evaluation. The release pipeline, the test organization and each site write to the same record.
- Who checks it
- The test organization, the program office, and the authorizing official.
- What the check shows
- That the build each site recorded is the build the test organization recorded as passing.
- What it can't show
- That a site runs what it reports. That rests on the site's own write.
- Where policy asks
- DoD AI Ethical Principles, Traceable
- 01Sep 02model-build · release-pipelineBuild 2026.09.3 created
- 02Sep 09test-completed · test-orgBuild 2026.09.3 passed 412 of 412 tests
- 03Sep 16model-fielded · release-pipelineBuild 2026.09.3 fielded to 3 sites
- 04Sep 16model-version · site-2Site 2 running build 2026.09.3
An uncrewed aircraft changes its own route inside set limits
An uncrewed aircraft inspects a levee and reroutes around weather on its own, inside limits loaded before the flight. A remote pilot can take over at any time.
- Who checks it
- The operator's safety office, the program office, and an investigator if something goes wrong.
- What the check shows
- The limits loaded, each route change the aircraft wrote, and what the pilot did.
- What it can't show
- Anything the aircraft didn't write.
- 0107:58:40agent-configuration · uas-17Limits loaded: geofence 12, ceiling 120 m
- 0208:31:05route-changed · uas-17Route changed around a storm cell: 2.4 km detour
- 0308:31:06limit-held · uas-17Detour inside geofence 12
- 0408:52:19operator-action · pilot-3Flight ended early by the remote pilot
An imagery model flags a site, and an analyst confirms it
An analysis program runs an AI model over overhead imagery. The model flags changes, and an analyst reviews every flag before a report goes out.
- Who checks it
- The program's analytic standards reviewers, its oversight, and the requesting unit.
- What the check shows
- Which model and threshold produced the flag, and that an analyst confirmed it before the report went out.
- What it can't show
- Whether the flag was right.
- 0106:20:11model-version · imagery-modelVersion 4.2 in use, confidence threshold 0.85
- 0206:20:14input-received · imagery-model46 images read for area 12, collected 04:00 to 06:00
- 0306:27:52detection-flagged · imagery-modelSite 17 flagged: vehicles grouped where none were the day before
- 0407:05:33analyst-review · analyst-04Flag confirmed
- 0507:41:10report-sent · analyst-04Report sent to the requesting unit
A defense agent cuts off a host on its own
A network defense agent watches sensor alerts and can isolate a host without waiting for a person. The watch officer decides what happens next.
- Who checks it
- The security operations lead, the incident responders, and the inspector general.
- What the check shows
- What the agent did and when, even if an intruder later reaches the network's own logs.
- What it can't show
- Who the intruder was. PacSpace attributes nothing.
- 0102:13:07alert-received · defense-agent-2Sensor 12: outbound traffic to an address never seen on this network
- 0202:13:09agent-action · defense-agent-2Host ws-0418 cut off from the network
- 0302:13:10agent-handoff · defense-agent-2Incident 5521 handed to the watch officer
- 0402:26:44decision-made · watch-officer-1Host kept isolated; investigation opened
A report shared with another agency and a partner arrives unchanged
An agency shares a report with another agency and with a coalition partner. Each side writes what it sent or received to the same record.
- Who checks it
- Both receiving parties, and the sending agency's oversight.
- What the check shows
- That what arrived is what was sent, and that the correction is a new entry beside the original.
- What it can't show
- The report's contents to anyone the sender didn't share them with.
- 01Oct 01delivery-sent · agency-aReport 26-1143 and 2 annexes sent to agency B and partner C
- 02Oct 01delivery-received · agency-b3 files received; each matches what was committed
- 03Oct 02delivery-received · partner-c3 files received; each matches what was committed
- 04Oct 03correction-sent · agency-aAnnex 2 corrected in a new entry with its reason; the original stays
The operator decides what to write.
That is the same limit every log has. What the operator gives up is changing the record afterward. If writing stops, the records on either side put a start and an end on the gap.
PacSpace records and never decides. Whether an action was right is for the people who check the record. How the record holds
Where an example points to a policy, the record answers part of it, never the whole. PacSpace is not a compliance product and doesn't make a system compliant or authorized. OMB's two memos do not cover AI used as a component of a National Security System. The scope of each rule
Bring the case you think breaks it.
We would rather be evaluated by use than by description. Talk to us and we'll put you in a live environment: commit a record, do your best to change it, then check it yourself, with us out of the loop. The change shows.
The record must exist.