# Track eval results over time

Every score, for every model and checkpoint, on every dataset and harness version. The eval job writes with add-only access, so old scores can’t be quietly changed.

## An example

**You:** Eval job finished for r-0412 at 40k: GSM8K 61.3, MATH 24.8, HumanEval 38.4. Eval harness v2.3.

**ADoor:** Added 3 entries to Eval results

**Eval results**

| Model | Dataset | Score | Harness |
| --- | --- | --- | --- |
| r-0412@40k | GSM8K | 61.3 | v2.3 |
| r-0412@40k | MATH | 24.8 | v2.3 |
| r-0412@40k | HumanEval | 38.4 | v2.3 |

## What you can ask later

- “Which checkpoint has the best GSM8K on harness v2.3?”
- “Did anything regress more than two points since last week?”
- “Show scores for every model on MATH, newest first.”
- “Which results were run on the old harness and need rerunning?”
- “Chart GSM8K and MATH scores for every checkpoint of r-0412.”

## Who can see what

- **The eval job** writes results with add-only access, so it can’t change old scores.
- **An agent** like Cursor or Claude can read results and can’t write them.
- **Researchers** read everything and add notes.
- **A partner lab** can read results for one shared model only.

## How it compares with a spreadsheet or an eval platform

Eval platforms like Braintrust or LangSmith are better at running the evals themselves, per-example traces and side-by-side output review.

As the long-term scoreboard, ADoor holds up: add-only results, plain SQL for regressions, and read-only access for the agents that decide what to train next.

## How to start

1. Open an assistant like ChatGPT, Claude, Grok or Cursor and connect ADoor through MCP.
2. Tell it: “Set up eval results: Eval results (model, checkpoint, dataset, dataset version, score, harness version, run at).”
3. Give the eval job add-only access, and share one model’s results with a partner lab if you need to.

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

## Questions people ask

### Can results be changed after the fact?

Give writers add-only access, and the history records each change and who made it.

### Can I store per-example outputs?

Yes, as entries in a linked list. For large volumes, store them elsewhere and link them.

### Is ADoor free?

Yes. The free plan 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 someone leaves?

Remove their access and it stops working on their next request. The history still shows what they added and changed.

## Related

- [Keep a model training log](/use/model-training-log/)
- [Log benchmarks and latency measurements](/use/benchmarks/)
- [Start a dataset registry](/use/dataset-registry/)
- [ADoor vs Supabase](/compare/adoor-vs-supabase/)
- [ADoor vs Google Sheets](/compare/adoor-vs-google-sheets/)
- [All use cases](/use/)

Updated October 20, 2026.