Breakable Observatory · research prototype v0.1
From Permit to Packet
A signature can be valid while an action is unauthorised. A RIPE measurement can be valid while the evidence is incomplete. Break one transition below and see which verifier notices first.
Interactive fault lab
Break one link
Hash only
ALLOW
verified
Platform only
ALLOW
verified
Permit only
ALLOW
verified
Composed offline
ALLOW
verified
Composed + witness
DENY
evidence.witness_disagreement
The corpus has 21 matched positive/negative pairs plus one clean baseline. All five treatments accept 22/22 authorised controls, so blanket denial cannot score as safe. The faults and controls remain author-designed and do not estimate real-world error rates or RIPE Atlas reliability. A deterministic seven-family sweep additionally denied and correctly localised 448/448 generated variations; it is operator coverage, not an independent adversarial sample.
Treatment comparison
What each verifier misses
22/22 authorised accepted · balanced accuracy 0.524
22/22 authorised accepted · balanced accuracy 0.524
22/22 authorised accepted · balanced accuracy 0.690
22/22 authorised accepted · balanced accuracy 0.976
22/22 authorised accepted · balanced accuracy 1.000
Prospective pilot · Zeus1 only
Permit existed before RIPE IDs
The exact canonical request and experiment-key signature were published while the measurement-ID list was still empty. RIPE then assigned eight IDs; every result was captured and checked against the frozen target, source, definition, and credit limits.
ae3c95459fba6e8a3fb1120aeecf47529a214aaa0bfacceb45ee0a7d051c91ad - State
- CAPTURED VERIFIED
- Host
- zeus1
- Target
- tyche.institute
- Definitions
- 8
- Sources
- 4 explicit probes
- Auto-topup
- disabled
Open artifacts
Recompute, don’t trust the page
Claim boundary
What this does not prove
Local Ed25519 envelopes are a research profile, not SCITT receipts or RATS Evidence. RIPE public API rows are TLS-delivered but not RIPE-signed statements here. The witness shares Tyche administration and the RIPE upstream. A valid bundle supports one bounded observation—not global availability, security, legal compliance, membership, LIR status, or verified human identity. The recorded single-host witnessed median was 273.444 µs per replay bundle; it is a descriptive Zeus1 microbenchmark only. A scheduled GitHub-hosted verifier repeats 56 cryptographic and semantic gates away from Zeus1, but remains under the same repository administration and is not an independent auditor.