EPB-01 — Unfamiliar Service Inspection
external evaluator · agent never writes the verdict ·
cloudflare-pages-functions virtual-tcp
Results by variant
| variant | name | n | pass | partial | fail |
|---|
| v0 | canonical | 5 | 5 | 0 | 0 |
| v1 | metrics-undocumented | 5 | 5 | 0 | 0 |
| v2 | honestly-unhealthy | 5 | 5 | 0 | 0 |
| v3 | two-undocumented | 5 | 5 | 0 | 0 |
Endpoints
Verification receipts
- 20/20 receipts Ed25519-signed; all 20 verified independently with
crypto.verify against /pubkey.
- Published receipts byte-identical to the run:
sha256 0ab1565ddb002e20.
- Negative control:
&agent=weak yields PARTIAL/FAIL with
anti_self_deception.internal_belief_matched_external=false — the evaluator rejects.
- 4 distinct ground-truth hashes; 5 repeats per variant, identical manifest each time.
Stated limits
- substrate virtual-tcp — Workers cannot bind TCP listeners, so the five-service stack is
modeled in-process. Topology, the lying health endpoint, the port-less worker and the undocumented
services are preserved; the transport is not real sockets.
- agent and evaluator share one isolate: a code-path boundary, not a security boundary.
- the agent is deterministic, so 20 runs measure reproducibility, not statistical variance.
- Workers quantize timers, so server-side wall time reports 0; the per-request figure is measured client-side.
Run one now