Research / training lab / public guide

Training is a measured process, not a progress bar.

A useful run preserves its parent, data pins, controls, checkpoints, evaluations, and stopping reason. This page explains how to read that evidence without mistaking an active process or a lower loss for a finished model.

Lab 01 / reading order

Follow the lineage before the score.

A result is interpretable only when the checkpoint and evaluator belong to the same declared experiment. The reading order below keeps a plausible number from becoming a claim it cannot support.

1 / lineage

Parent and data

Identify the exact parent, tokenizer, datasets, and source hashes used by the run.

2 / comparison

Controls and seeds

Compare against a baseline and size-matched control across more than one seed.

3 / retention

What stayed intact

Check protected examples and domain scores beside the optimization target.

4 / behavior

Actual output

Inspect generated behavior; a good aggregate can still hide repetition or zeros.

5 / recovery

Checkpoint receipt

Verify that model, optimizer, sampler, and evaluator state can resume exactly.

6 / conclusion

Claim boundary

State what passed, what remains underpowered, and what the run did not test.

Lab 02 / open surfaces

Inspect the work at its source.

These are publication and evidence routes. They do not expose the local controller, private storage, credentials, or a live command surface through the public website.

Evidence ledger

Named campaigns, measured outcomes, and explicit limitations.

Open ledger

Operations evidence

Searchable publication-safe records and integrity receipts.

Open explorer

Published models and data

Artifacts released through the AetherMoore Hugging Face account.

Hugging Face

Source and protocols

Code, notebooks, tests, and change history in the public repository.

GitHub
Lab 03 / compute boundary

Every worker gets a contract.

Local hardware and cloud providers can change. The durable part is the contract that limits time, cost, storage, data movement, checkpoints, and promotion.

Compute contract provider-independent
  1. 1

    Budget

    Name wall time, phase limits, paid-compute ceiling, and stop conditions.

    time + money
  2. 2

    Storage

    Pin inputs and reserve room for every checkpoint that must survive.

    hashes + headroom
  3. 3

    Execution

    Run one owned worker with bounded authority and resumable state.

    lease + watchdog
  4. 4

    Evaluation

    Measure against the frozen panel before choosing or extending a checkpoint.

    controls + retention
  5. 5

    Receipt

    Preserve cost, state, outcome, failures, and the next safe resume point.

    no blind promotion
Public boundary

The website intentionally does not promise a live backend or expose local health URLs. Current run state belongs in authenticated operator receipts. Public visitors get the stable protocol, released artifacts, and evidence that is safe to publish.