Operation AngelsBorn Between 2 Generals

CreateConnectProtect

Safe PerimeterOperation Angels · Protect

Verify it yourself

Do not take the site's word for it

The checks below are running in your browser as you read this. They are not a recording of a build that passed somewhere else.

Footage: generated (OpenAI Sora 2) - abstract, depicts nothing and no one

Every claim this site makes about its own instruments is checkable here. The classifier is run against its vector table and fuzzed against its invariants; the token scheme is checked against a published cryptographic test vector; the models are checked for the properties they are supposed to have.

running...

What each group is for

  • Tier vectors - the ten encounters the ladder exists to get right, from passing on a pavement to sustained contact in a school zone. Two of these failed when they were first written, and both failures were real faults in the classifier: a crowded-venue encounter that tipped into Notify at twenty seconds, and a dispatch tier that could never actually be reached.
  • Tier invariants - thousands of randomized encounters checking the three rules that must hold for every possible input: a lost fix never alerts, a degraded fix never exceeds Notify, and dispatch never happens without sustained dwell plus a zone or a third repeat.
  • Crypto - the HMAC used to derive the rotating identifiers is checked against RFC 4231's published test vector, so the derivation is not merely self-consistent but demonstrably the standard algorithm.
  • Model properties - coverage never exceeds the sum of its parts, de-anonymization risk never falls as more people watch, the graduated ladder never refers more cases than naive dialling.
  • The console - the query engine is run against its own vectors, including the ones that matter most: a query that would resolve an identity is refused; one approval alone cannot resolve; two approvals plus an order can; and every refusal and every grant lands in an append-only log that cannot be rewritten.
  • Sensing - the admissibility rules, including that a single modality can never license escalation and that two modalities disagreeing beyond tolerance is a refusal rather than an average.
  • The warning - the state machine, and the anti-inversion invariant: the warning text is byte-identical whether or not a child is present, because a warning that varied would be a detector.
  • The bracelet - real Merkle inclusion proofs over a day of events, verified in your browser, including that a tampered leaf fails and that a disclosure carries no other event's salt.
  • The anchor coverage model - the arithmetic behind the largest unknown in the concept: that doubling target coverage cannot lower cost, that zero sites means zero anchors, and what the assessment's own forty-thousand-anchor figure would imply.
  • The page itself - no third-party requests, no storage, and the full-bleed loops verified by their decoded dimensions rather than by a flag that says they are playing.

If any row reads FAIL, the site is wrong and the failure is the interesting part. The checks are in /assets/check.js and the engines they exercise are in /assets/engine.js; both are plain readable JavaScript.

How this works

0:51 · narrated · captioned · nothing here is sent anywhere when you press play. Prefer to read it? The transcript is below.

Transcript
AtSectionWhat it says
0:02Running nowThe checks on this page are running in your browser as you read this. They are not a recording of a build that passed somewhere else.
0:09What is checkedThe classifier is run against its vector table and fuzzed against its invariants. The token derivation is checked against a published cryptographic test vector.
0:18The refusalsThe console's query engine is checked as well, including the parts that matter most. A query that would resolve an identity is refused, and one approval cannot resolve it.
0:29The page itselfThe page checks itself too. No third-party requests, no storage, and the loops verified by their decoded dimensions rather than by a flag that says they are playing.
0:39If a row failsIf a row reads FAIL, the site is wrong, and the failure is the interesting part. The checks and the engines they exercise are plain, readable JavaScript.