# Start an experiment tracker

The hypothesis, what changed, the result and whether it held up. Before anyone tries a new warmup schedule, they can ask what the team already learned.

## An example

**You:** Experiment e-31: does adding 10% code data help math evals? Compare r-0412 (with code) against r-0409 (without).

**ADoor:** Added 1 entry to Experiments, linked to both runs

**Experiments**

| Experiment | Change | Result | Verdict |
| --- | --- | --- | --- |
| e-31 | +10% code data | GSM8K +2.1, MATH +0.8 | Keep |
| e-30 | Warmup 2k → 500 steps | Loss spike at 1k | Revert |

## What you can ask later

- “Which changes improved GSM8K by more than a point?”
- “What have we already tried with warmup?”
- “Show every experiment where the verdict was revert, and why.”
- “Compare the configs of the two best runs field by field.”
- “Chart GSM8K against training tokens for every run in e-31.”

## Who can see what

- **Researchers** add experiments and verdicts.
- **A planning agent** reads all experiments and can propose new ones, but can’t edit verdicts.
- **A collaborator from another team** reads one project’s experiments only.

## How it compares with Weights & Biases, Neptune or Comet

Weights & Biases, Neptune and Comet lead on automatic logging from your training code, rich comparison dashboards, sweeps and artifact versioning.

The reasoning around experiments belongs in ADoor: what you were testing and what you concluded, compared with plain SQL, with an agent able to check what’s been tried before it suggests something.

## How to start

1. Sign in to ADoor, then add it to an assistant like ChatGPT, Claude, Grok or Cursor through MCP.
2. Tell it: “Set up an experiment tracker: Experiments (id, hypothesis, change, runs, result, verdict, notes), linked to Runs.”
3. Give a planning agent read access to Experiments and add access, without edit access to Verdict.

[Get your account](https://app.adoor.halmoslabs.com/signup)

## Questions people ask

### Should I still use W&B?

For live metric dashboards, yes. ADoor keeps the record of what you tried and what you concluded.

### Can I join experiments with runs?

Yes. Experiments link to runs, and plain SQL joins them.

### What does it cost?

Nothing on the free plan, which holds 3 tables and 100 rows. Pro is $10 a month per workspace, with 1 GB of storage and no per-GB charges.

### What if I want to leave?

Export each list as CSV or JSON and take it with you.

## Related

- [Keep a model training log](/use/model-training-log/)
- [Track eval results over time](/use/eval-results/)
- [Share research notes](/use/research-notes/)
- [ADoor vs Neon](/compare/adoor-vs-neon/)
- [ADoor vs Airtable](/compare/adoor-vs-airtable/)
- [All use cases](/use/)

Updated October 20, 2026.