Verifier reference
How verification works
Techtree checks a published Result bundle’s integrity, controlled comparison, score consistency, and publication policy. Verification makes the bundle internally checkable; it is not independent observation of the run.
What Techtree verifies
Internally checkable evidence
- Integrity. Stored files match their digests and signatures.
- Controlled comparison. Both branches use the same published Climb and ordered tasks, with only the permitted Skill changed.
- Score consistency. The summary recomputes from task-level results.
- Publication policy. The bundle contains no episodes, transcripts, or machine-local paths.
What remains unproven
Verification is not observation
- The site did not witness the execution.
- The participant’s machine is not independently attested.
- The result does not establish generalization beyond the Climb.
- Nobody else reproduced it unless a separate reproduction says so.
Method
The controlled comparison
Techtree runs the same fixed tasks twice.
The model stays the same. The Hermes harness stays the same. The runtime, tools, scorer, task membership, sampling, and budget stay the same. Only the Skill may change.
Same agent
Same tasks
Same evaluation
Skill v1 → Skill v2
Prime Intellect’s Verifiers library runs and scores the tasks. Techtree checks that the comparison stayed controlled and packages the Result into a signed proof bundle.
The Result may be an uplift, a tie, a regression, a failed Run, or an invalid comparison. Techtree does not turn every attempt into a success.
Check it yourself
Verify a Result offline.
Anyone holding the participant’s bundle can run the same verifier on their own machine.
Verify offline
techtree proof verify path/to/result-bundle
The offline verifier currently performs 17 checks
- the submission is small enough to be a proof bundle
- the submission is the four-member document and carries nothing else
- the bundle carries no more files than a proof bundle has
- every path is one relative path inside the bundle, written once
- every file is canonical base64, and none of them is empty
- the bundle carries the signed list of its own files
- every file hashes to the digest the bundle claims for it
- every signed document hashes to the digest written beside it
- every signature verifies under the key the bundle carries
- the fingerprint names the key by hashing it rather than by claiming it
- the bundle carries the signed result summary it commits to
- the campaign it names is one this site publishes
- the wins, losses and ties recompute from the task results
- the tasks are the list the campaign committed to, in order
- the terms the run was carried out under permit publishing this
- no file in it holds an episode, a transcript, or a path on a machine
- what the submission claims about the bundle is what the bundle says
For detailed information, read the Docs, copyable as Markdown to your agent.