BeavousCampaign workshop
Evaluation / Judge guide

The fastest path through the proof.

This page separates the completed paid-provider campaign, deterministic recovery evidence, and the remaining launch limitations so the evaluation stays honest.

1. Start with the owned Field Bottle.

Open the public demo product, then enter the live judge workspace. Review the source references and approved Product Truth Lock.

2. Inspect the completed live campaign.

Choose “Open live judge workspace” on the sign-in page, then open “Field Bottle — Verified Live Campaign.” It completed 43 durable steps with four Gemini Flash Image masters, sixteen clean ratios, sixteen text variants, approved-claim copy, and a 10.005-second Gemini Omni portrait reel with audio. Open an asset and choose “View provenance” to inspect its B2 key, SHA-256, Genblaze run, evaluation, and integrity result.

3. Create and test the live workflow.

The credential-free workspace is writable: add an owned product or import the public Field Bottle URL, approve a Truth Lock, choose campaign direction and channels, then build and approve a production plan. Paid generation is enabled with one active run, four campaigns, and $20 of shared daily spend. Identity settings and account deletion remain protected.

4. Follow recovery and lineage.

The live run persists 38 verified Genblaze manifests plus a fetched-byte-verified 40-file campaign ZIP. Its first 16:9 detail crop remains visible as failed evidence, followed by the passing padded generation. “Trails in Four Frames” remains available as a clearly labelled deterministic recovery showcase with a genuine Genblaze from_result parent link.

Providers and models

WorkDefaultFallback
Truth extraction + planning + QAGemini 3.6 Flash3.5 Flash-Lite
Preview imageGemini 3.1 Flash-Lite Image3.1 Flash Image
Master + ratio imageGemini 3.1 Flash ImageGemini 3 Pro Image
Fast portrait reelGemini Omni Flash PreviewVeo 3.1 Fast
Premium reelVeo 3.1Feature flagged
Ratios + overlays + ZIPDeterministic transforms + packagerNo AI spend

Where Genblaze is meaningful

The implemented worker wraps Google image and video providers in Genblaze contracts, routes media through the B2 ObjectStorageSink, assigns stable tenant and run identities, and persists canonical manifests plus fetched-byte verification. The live campaign routes real Gemini image and Omni video outputs through named Genblaze pipelines and the strict B2 sink. All 38 manifests pass canonical verification and all 38 media objects pass an independent fetched-byte hash. Signed video GET, byte-range delivery, FFprobe, full FFmpeg decode, and ZIP integrity also pass. The separate recovery showcase preserves one genuine from_result parent link for a concise lineage walkthrough.

Known limitations

  • The credential-free judge account is shared. Creative workflows are live and bounded; identity settings and account deletion are protected, and daily quota may be consumed by another evaluator.
  • Product Truth scores are model-assisted checks, not certainty guarantees; human approval remains authoritative.
  • All configured model IDs passed catalog probes. Text, Flash Image, and Omni live calls passed; premium image and video fallbacks remain intentionally disabled.
  • Integrity verification does not prove authorship or signer identity.
  • Deterministic showcase media and human-authored QA are labelled; neither is presented as a paid Gemini call or model evaluation.
  • The live campaign did not require a paid evaluator-led master correction. Historical and deterministic parent-linked recovery evidence is identified separately.
  • The separate public Backblaze showcase bucket could not be created because the B2 account is still inside Backblaze's payment-history restriction. The private canonical B2 campaign is complete; reviewed public copies are served by the web deployment and identify their private B2 canonical source.

Technical links

The repository, architecture, test instructions, model list, deployment notes, and SDK feedback are maintained together. The one-click judge action keeps credentials server-only; no password is published in this page or repository.