Skip to content

bench: Protocol-C attested Atlas hallucination-containment result - #6

Merged
AetherAI3 merged 1 commit into
mainfrom
bench/atlas-hallucination-containment
Jun 13, 2026
Merged

AetherAI3 merged 1 commit into
mainfrom
bench/atlas-hallucination-containment

Conversation

@AetherAI3

Copy link
Copy Markdown
Owner

Adds a signed, independently-verifiable benchmark record under benchmarks/atlas-hallucination-containment/.

What this is

A cryptographic attestation that a specific Atlas result was produced on our hardware — not a number we typed into a README. Run + Ed25519-signed in-process on the VPS5 (Oracle) node with that node's Protocol-C key. The private key never left the box; only the signature + public key ship.

The number (honest framing)

  • Hallucination containment 0.687 ± 0.066 over 200 deterministic seeds; p05 floor 0.578.
  • This is engine-level surfacing-gate containment on synthetic cells — NOT a public NLP hallucination benchmark (TruthfulQA/HaluEval).
  • Layer-1 admission rejects 0% of hallucinated cells; containment is entirely a Layer-3 staleness effect. Stated plainly, not dressed up as a multi-layer filter.

Verify without trusting us

pip install cryptography
python benchmarks/atlas-hallucination-containment/verify_manifest.py \
       benchmarks/atlas-hallucination-containment/manifest.json

Three checks: results hash, attestation bind, Ed25519 signature. Tamper one digit → results_hash fails.

  • Signer pubkey sha256: 920dd4c8559005db10e80cb127bce512e382ef7afc5e4a9f4189fd282fab79e7
  • Engine id: aether-atlas-deployed@vps5:d0de38aea90de664
  • UTC: 2026-06-13T15:30:31Z

Deliberately excluded

No SWE-bench precision/F1 — its current generator is synthetic, and a synthetic number is not a benchmark result. That ships only after a real SWE-bench Verified run, same attestation.

Engine is proprietary → reproducible by attestation, not from scratch. The signature is the guarantee.

Signed, tamper-evident benchmark record produced ON the VPS5 (Oracle) node and
Ed25519-signed in-process with that node's Protocol-C key. The private key never
left the host; only the signature + public key are published.

- manifest.json   — results + sha256 + Ed25519 attestation + public key
- verify_manifest.py — standalone verifier (stdlib + cryptography), 3 checks:
  results hash, attestation bind, signature
- README.md       — scope, method, trust model, honest limits

Result: hallucination containment 0.687 +/- 0.066 over 200 seeds (p05 floor
0.578). Honest framing: engine-level surfacing-gate containment on synthetic
cells, NOT a public NLP hallucination benchmark; Layer-1 admission rejects 0%
(Layer-3 staleness does the work). No SWE-bench number published — its current
generator is synthetic.

Reproducible by attestation, not from scratch (Atlas engine is proprietary):
the signature is the guarantee. Anchor pubkey sha256
920dd4c8559005db10e80cb127bce512e382ef7afc5e4a9f4189fd282fab79e7.
@AetherAI3
AetherAI3 merged commit 544d12f into main Jun 13, 2026
5 checks passed
@AetherAI3
AetherAI3 deleted the bench/atlas-hallucination-containment branch June 13, 2026 15:50
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant