public-model.example is blocked (hard-block) under version 6 and sealed. For assistant.internal.example.local it is logged only (log-only), with nothing sealed. - public-model.example
- hard-block
- assistant.internal.example.local
- log-only
- sealedFor
- public-model.example
36 fields, including the sealed record and sample ciphertext
Input
- occurredAt
- 2026-11-03T09:14:00.000Z
- promptMasked
- Summarise the case notes for NRIC S•••••••D before the review.
- identifierMasked
- S•••••••D
Detection
- classesScanned
- 11
- detectorClass
- id.sg.nric
- tier
- deterministic
- confidence
- 1
- matchCount
- 1
Policy snapshot
- versionNumber
- 6
- compiledAt
- 2026-11-02T18:00:00.000Z
- snapshotHash
- 143059541e19c2746ba6f7db2e740e6ac6fcc4a5e79bc676ba96a0ad4b90db22
public-model.example
- sanctionedState
- unsanctioned
- ruleName
- National IDs to unsanctioned AI
- stage
- enforce
- action
- hard-block
- winningClause
- rule
- evidenceCapture
- full
assistant.internal.example.local
- sanctionedState
- internal
- ruleName
- National IDs to the approved internal assistant
- stage
- enforce
- action
- log-only
- winningClause
- rule
- evidenceCapture
- none
Sealed record Sealed
- sealedFor
- public-model.example
- evidenceClass
- prompt-payload
- aeadAlgorithm
- aes-256-gcm
- keyWrapAlgorithm
- aes-256-gcm-keywrap
- plaintextByteLength
- 62
- ciphertextBytes
- 78
- nonceBytes
- 12
- wrappedDekBytes
- 125
- standIn
- software-custody-token
- plaintextFoundInRecord
- false
Sample ciphertext Sealed
A reproducible sample of the masked prompt, under a throwaway key anyone can derive.
- nonce
- /grhlLC7NEDWFr8N
- ciphertext
- BRd/lpzLVH4W6JdEwhyOJMLBuGbl21ZRZVp8HBctML6ZKMBr9CgRVdKXKAdvQJ4ooXrENHg2dmI6/zbx7MW2/mmsgtJwKd6kp1h5zwlX
- ciphertextSha256
- 044680a506a47db5c39badd04bea36f3ff6afc2745580d480bfb5fc136ed5da6
09:14: the analyst presses send
At 09:14 UTC on 3 November 2026, an analyst asks a public model, public-model.example, for help with a case. Masked for this page, the prompt reads: “Summarise the case notes for NRIC S•••••••D before the review.” The prompt the system judged held a complete synthetic identifier with a valid checksum.
The prompt does not leave. In the browser extension, the analyst gets a dialog titled “Blocked”, which says “This can’t be sent to this destination. Use the approved destination instead.” and names the category “Singapore NRIC/FIN”. The analyst can close it. A dispute button exists, but in this version it goes no further than the browser. There is no button to send anyway.
Why it stopped
Two separate steps produced that block. This run scanned all 11 classes in the detector registry and returned one hit: id.sg.nric, from the deterministic tier, with one match at confidence 1. The live path scans only the classes that enabled rules name. No model took part. Detection only names a class of text.
The decision came from policy version 6, compiled at 18:00 the evening before. The rule National IDs to unsanctioned AI covers that class going to a destination listed as unsanctioned, and its action is hard-block. No group has its own rollout stage in this policy, so the deployment default, enforce, applies, and it keeps the rule’s action as written.
The same prompt, judged for the internal assistant
For comparison, the same prompt was also judged under the same version as if it were bound for assistant.internal.example.local, a destination listed as internal. Detection found the same class. A different rule, National IDs to the approved internal assistant, matched, and its action is log-only. Sent there, the prompt would go through and the analyst would see no dialog.
The stage is enforce here too. A stage caps an action and cannot raise one, so enforce does not turn this into a block. A team that wants a block here changes the rule and publishes a new version. Being internal earned no exemption: the text was scanned and a rule decided. In this policy, an NRIC reaching the internal assistant is a choice somebody wrote down, and it can be read and challenged.
What is recorded
The public-model decision names the rule, the stage, the action and evidence capture full, and the decision’s trace shows the rule clause setting the action. It is tied to version 6 by that snapshot’s hash, which begins 14305954. The prompt is sealed with AES-256-GCM: 62 bytes of plaintext become 78 bytes of ciphertext, and a search of the sealed record for the plaintext finds nothing.
The internal decision has evidence capture none, so nothing is sealed for it.
Nobody reads the sealed prompt in this case, because no reveal runs. A reveal is the officer ceremony that decrypts it. Officer keys in this POC are software keys, so hardware custody is not part of what it shows.
What it came to, and its limits
The prompt did not reach the public model. On a live path, the security team would hold that decision and one sealed record it cannot casually read. That is the whole result: one prompt, and no count of anything.
- The run called the detector, the policy evaluator and the sealing code directly. The verdict service, the database, the audit chain and the browser did not run.
- The dialog wording comes from the extension’s code. That path is proven on a synthetic site only, and it records the actor as unattributed.
- The version 6 hash comes from the control plane’s compile computation with a fixed id, so that it reproduces.
- The sealing code picks its own random key and nonce, so the byte counts come from a real seal and its bytes are not published. The sample ciphertext shown with this case seals the masked prompt under a throwaway key anyone can derive.
How to read this
| Read it as | Which means | Do not read it as |
|---|---|---|
| Public destination | id.sg.nric detected, rule action hard-block, prompt not sent | An incident count |
| Internal destination | Same detection, rule action log-only, nothing sealed | Exempt from scanning, or raised to a block by the stage |
| The evidence | Ciphertext under software keys, with no reveal in this run | The prompt in the clear, or hardware custody |