Recompute the claim
Anchor reruns the claimed benchmark path and checks the result against pinned protocol fields.
ALETHOR BUILDS ANCHOR
Know which model results are real.
Anchor is Alethor’s product for proving model results. When someone says a model is better, Anchor reruns the benchmark, checks the evidence, and shows whether the result is trustworthy enough to become your team’s baseline.
invite-only private alphabaseline comparisonsigned evidence
Verification detail
Verification state, baseline status, and evidence stay on one record.
Verdict readyBaseline eligibleEvidence bundled
The decision surface keeps the benchmark claim, pinned protocol fields, compare status, and evidence bundle close to the run.
What Anchor Does
Anchor keeps benchmark decisions readable.
Verify a claimed benchmark result, accept the passing run as the official baseline, compare later candidates against the same protocol, and keep signed evidence attached to the decision.
Current private-alpha scope
verify claimed resultpin official baselinecompare later candidateexport signed evidence
Accept one baseline
A passing canonical run becomes the official baseline for later candidate comparison.
Carry signed evidence
The verdict, artifacts, manifest, and signature stay attached to the acceptance decision.
Workflow
A fixed path from claim to signed evidence.
Each step stays anchored to one run, one benchmark path, and one accepted baseline.
1
Verify
Recompute the claimed benchmark result on the fixed protocol path.
2
Inspect evidence report
Read the verdict, checks, artifacts, and pinned protocol fields.
3
Pin baseline
Accept a passing canonical run as the official baseline.
4
Compare candidate
Evaluate a later run against the same pinned baseline protocol.
5
Export signed evidence
Package the bundle, archive the record, and export signed evidence.
Proof
Real product states keep the decision legible.
Verification verdict, baseline readiness, compare validity, and package outputs stay attached to the same run record.
Runs
Runs separate job status from verification verdict.
Candidate and baseline records stay readable before a run is accepted or compared.
Compare
Compare stays anchored to the pinned protocol.
Outcome, delta, and protocol-match fields stay explicit before a candidate is accepted.
Private Alpha
The current private-alpha contract is deliberately fixed.
Anchor currently runs on one benchmark path with pinned protocol fields for verification, baseline acceptance, compare, and signed evidence export.
Invite-only private alpha. Current scope: verify a claimed result, pin the baseline, compare later candidates, and export signed evidence.
- benchmark_id
- mmlu_pro
- task
- leaderboard_mmlu_pro
- few-shot
- 5-shot
- metric
- acc
- canonical batch_size
- 1
Read A Run
Four statuses explain the current state.
These terms orient a run quickly without turning the product into a general dashboard.
Job status
queued / running / succeededWhere the verification run is in execution.
Verification verdict
pass / failWhether the recomputed benchmark result passed the checks.
Baseline eligibility
eligible when canonicalWhether the run can be accepted as the official baseline.
Package status
ready / packagedWhether the signed evidence bundle is ready or already exported.
Why Teams Use Anchor
Trust improves when the evidence stays attached.
Anchor keeps the narrow acceptance path readable instead of scattering it across trackers and dashboards.
Explicit checks
Verification steps stay named, visible, and tied to the verdict.
Reproducible record
Artifacts, manifests, and replay inputs stay attached to the accepted run.
Stable comparison
Later candidates are compared against one pinned baseline protocol.
Private Alpha
Request private alpha.
Anchor is currently available as an invite-only private alpha for teams that need to verify model claims before they set a baseline.