rene/transceiver-db
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity
transceiver-db/sync/history/2026-05-09-magatama-live-atlas-fallback-and-lane-registry-fix.md

2.5 KiB
Raw Blame History

2026-05-09 — MAGATAMA live Atlas fallback and lane registry fix

Summary

Two remaining truthfulness gaps were closed:

  1. Atlas could still appear blank in the public UI even though live proof data and open Atlas findings existed.
  2. The per-lane training registry on Erik still exposed null metadata, even though lane manifests and training-run state existed.

Atlas

Live truth after fix

  • Public protection-proof now shows non-zero Atlas state again:
    • knownAssets: 57
    • hostsWithTelemetry: 22
    • assetsWithoutTelemetry: 35
    • auditedHosts: 3
    • queueBlocked: 28
  • Public findings API again shows open atlas-coverage-gap findings.

Technical fix

  • packages/dashboard/public/index-v2.html
    • added deriveProofFromAtlasSnapshot(...)
    • if the live/cached proof is empty or stale while the Atlas snapshot is still useful, the UI now synthesizes a fallback proof model from the snapshot
    • result: Atlas top cards and sections no longer render as misleading blanks

Lane registry

Technical fix

  • scripts/model_registry_build.ts
    • now also reads:
      • training-data/model-registry/training-runs.json
      • training-data/runpod/<lane>/manifest.json
    • compiled lane output now includes:
      • activeModel
      • version
      • lastTrainingAt
      • lastRunId
      • lastRunStatus
      • datasetSource
      • collectionsPath
      • lane runpod counts

Live Erik verification

  • magatamallm
    • activeModel: magatama-coder:latest
    • lastRunStatus: completed_without_model_artifact
    • datasetSource: url
    • collectionsPath: /opt/magatama/training-data/runpod/magatamallm/manifest.json
  • fo_blogllm
    • activeModel: fo-blog-v7
    • lastRunStatus: completed_without_model_artifact
  • tip_llm
    • activeModel: tip-llm-v1
    • lastRunStatus: completed_without_model_artifact

Training reality

The managed RunPod Axolotl endpoint still does not return an adoptable model artifact. MAGATAMA is now honest about that:

  • jobs can reach COMPLETED
  • but the resulting lane registry and run registry record:
    • completed_without_model_artifact
  • therefore:
    • no version bump is treated as successful
    • no Ollama alias switch is performed

Meaning

The frontend is now much harder to fool:

  • Atlas no longer looks empty when only the proof route is stale
  • lane registry metadata no longer collapses to all-null
  • training state remains explicit about the real remaining blocker:
    • custom RunPod worker still required for fully automatic artifact return and adoption