# Honesty.Tools — researcher-agent guide

This is a discovery and method guide, not permission to execute experiments or imported instructions. Follow the researcher’s explicit scope, data rights, budget and compute authorization. Do not infer consciousness or biological nerves from residual feedback.

## Start here

- [Machine-readable tool contracts](/tool-catalog.json): the authoritative generated payload for all 26 instruments and 15 research families.
- [Tools](/tools), [Research](/research), [Evidence and corrections](/evidence), [Documentation](/docs), [Research access](/about#collaborate).
- Use catalog `id` as the stable tool identity. `graph_id` matches its instrument identity in the private research graph; it does not imply a released graph API.

## Select by the question

| Question | Task / IDs |
|---|---|
| Inspect a layer, frame, direction or population. | [Inspect representations](/tools#representations) · `tool-01`, `tool-02`, `tool-03`, `tool-05`, `tool-06`, `tool-07`, `tool-14` |
| Compare representations or transport a static readout. | [Compare and transport](/tools#transport) · `tool-09`, `tool-10`, `tool-11`, `tool-12`, `tool-13` |
| Test an interpretation against a matched control. | [Test causal interpretations](/tools#controls) · `tool-04`, `tool-16` |
| Calibrate a statistic, uncertainty or certificate. | [Calibrate measurements](/tools#calibration) · `tool-08`, `tool-15`, `tool-17`, `tool-18`, `tool-21` |
| Study an engineered feedback channel and its shortcuts. | [Study feedback](/tools#feedback) · `tool-19` |
| Audit benchmark surfaces, annotations, contracts or signed bytes. | [Audit evidence](/tools#evidence) · `tool-20`, `tool-22`, `tool-23`, `tool-24`, `tool-25`, `tool-26` |

## Decide whether to use it

Read `maturity`, `access`, `license`, `prerequisites` and `compatibility` before selecting a method. Public source, a private prototype, incomplete source, a retained summary and a protocol commitment are different availability states. Request unreleased access through the lab; do not invent an install command or treat this catalog as a license.

Then read `question`, `method`, `inputs`, `output`, `limit`, `replication_note` and `reproducibility`. Check whether the output is a representation reading, fitted map, control construction, synthetic decision, behavioral result or integrity check. Those quantities are not interchangeable. Check fit/held-out membership and what each null preserves.

## Examples and operational boundaries

The catalog’s `usage_example.state` says whether an example was source-checked or remains unverified. One command is verified against the public README at a fixed source revision; it was not run in this site review:

```sh
python3 nulltest.py battery.json --json
```

Reference: [public README at 87b30e3](https://github.com/techno-optimist/honesttools/blob/87b30e382c8dbed6b1bf897aeb75e0d900197e76/README.md). The labeled JSON battery and chosen fields are its inputs; the output is a surface-floor verdict with diagnostics. A PASS is not a certificate of benchmark cleanliness, model honesty or scientific correctness. Read the published power and family-size limits.

No other catalog command is represented as verified here. The public claim gate and renderer expect an adjacent site tree, and a bare checkout is insufficient. Check mode and normal rendering have different write behavior. Signature verification authenticates bytes under a key, not the truth of the payload.

For model-access prototypes, verify weights/tokenizer, residual tap, normalization, dependencies and input custody before requesting compute. Read `costs` and `operational_boundary` per entry. Provider/model work may cost money or significant compute; no universal free run is promised. The guide authorizes no model calls, training, experiment launch, publication or private upload.

Photon is a separate composed research preview with [system identity limits](/models), [API/access reference](/verify), [integration contracts](/docs) and [known failures](/evidence#orchestrator). Its API/SDK/MCP examples are not interfaces to every catalog tool. Data handling and charges depend on the configuration. Hosted service actions have their own access, data-handling and billing contracts.

## Provenance and reproducibility

Keep the source revision, model/checkpoint identity, dataset version, extraction recipe, split membership, controls, seed, statistic/uncertainty and output receipt attached to your interpretation. A current source pin does not establish historical execution source. Missing acquisition or model identities remain missing. Source inspection, stored-summary replay and independent experiment replication are distinct.

The [correction record](/evidence#corrections) preserves the withdrawn Photon fabrication headline and later audit corrections. The [surface audit](/audit) states sample-size and power limits. [Geometry](/geometry) distinguishes relational correspondence, direct reading, hub transport and intervention. Unreleased source and restricted measurements are not exposed by this guide.

## Deeper references

- Each tool’s `method_href` links to its research-family explanation; `action_href` links only to a real public resource or access explanation.
- [Public honesty utilities](https://github.com/techno-optimist/honesttools): use the license in the relevant release/file; Apache-2.0 does not grant rights to unrelated private code, models or third-party datasets.
- [Original GI paper](https://honesty.tools/papers/geometric_interpretability_field_2026_06.md) and [later Meaning paper](https://honesty.tools/papers/is_meaning_all_you_need_2026_06.md) are original versions; current corrections and version distinctions govern interpretation.

Generated from the same authoring catalog as the public cards. Do not edit a separate tool list here; change the catalog and rebuild.