# Analyze Local Model Reasoning with PTS

Reasoning traces do not show which events change a model's chance of solving a task. This skill guides PTS searches, event interpretation, and local dataset exports with explicit measurement caveats.

## Install

```bash
npx skillstore add lambdasec/pts-pivotal-token-search
```

## Metadata

- Status: approved
- Slug: lambdasec-pts-pivotal-token-search
- Skillstore revision: r1
- Version status: missing
- Tree hash: 4f5b1a9eb19f90cac7fe48152269e9a2f9da78418da770e23ab3ea047b383db2
- Author: lambdasec
- GitHub username: lambdasec
- License: Apache-2.0
- Repository: https://github.com/lambdasec/skills/tree/7d791206069498962a85a0f1ba95f7468dcf82f7/skills/pts-pivotal-token-search
- Ref: 87c9669baf85385959f9b0a344dc9dfc8d978f9c
- Supported tools: Claude, Codex, Claude Code
- Audit status: complete
- Agent install advisory: allowed
- Manual install advisory: allowed
- Artifact signature: available
- Audit attestation: unavailable
- Human verification: not\_verified
- Risk factors: external\_commands, network
- Quality score: 78
- Quality tier: bronze
- Public page: https://skillstore.pages.dev/skills/lambdasec-pts-pivotal-token-search
- Manifest: https://skillstore.pages.dev/api/skills/lambdasec-pts-pivotal-token-search/manifest

## Capabilities

- Check PTS availability and plan token or sentence searches around the selected model, dataset, device, and budget.
- Guide optional Jacobian lens calibration and latent enrichment with shuffled controls.
- Interpret JSONL event fields and rank token or sentence events by absolute probability change within one scale.
- Explain model-specific results, heuristic event links, and limits on causal interpretation.
- Guide hosted or local visualizer use while requiring permission before uploading private events.
- Link events and export local datasets using documented formats, including DPO and steering.

## Use Cases

- Investigate pivotal reasoning events: Plan a small token or sentence search to identify events associated with changes in task success.
- Inspect a local evaluation trace: Review existing event files, compare events within one scale, and identify examples that need manual checking.
- Prepare a local training dataset: Link existing events and test a filtered DPO export before processing a larger collection.

## Prompt Templates

### Understand an events file

```
Explain the fields in my local PTS events file at [path]. Summarize its scales and model identifiers without running searches or uploading data.
```

### Plan a small search

```
Plan a token search for [model] on [dataset] using [device] within [budget]. Check available flags and ask before downloads or execution.
```

### Review latent evidence

```
Review [events path] and plan J-lens enrichment for [model]. Include shuffled controls, consent for downloads, and separate reporting of latent scores.
```

### Design a controlled export workflow

```
Design repeated-seed checks and an intervention for [events path], then propose a filtered local DPO export. Separate observational findings from causal evidence.
```

## Limitations

- PTS must already be installed; this skill does not install dependencies or provide the underlying implementation.
- Searches require substantial sampling, and token or sentence scores depend on reliable answer checking.
- Latent scores are observational and cannot be ranked directly against token or sentence probability changes.
- Results depend on the model, task, and sampling settings; hosted-model analysis and general prompt tuning are outside scope.

## Best Practices

- Confirm the model, dataset, device, budget, and download permissions before starting a small pilot search.
- Compare events within one scale and model, inspect examples manually, and repeat measurements with other seeds.
- Use shuffled controls for latent enrichment and require interventions before making causal claims.

## Anti Patterns

- Do not combine latent readout scores and emitted-event probability shifts into one ranking.
- Do not treat heuristic event links or category labels as verified causal mechanisms.
- Do not install dependencies, download assets, or publish private events without the required user consent.

## Security Audit

- Audited at: 2026-10-28T06:44:03.665\+00:00
- Summary: All 30 static findings are false positives: Markdown backticks, legitimate local PTS workflows, and relevant project links. The skill requires consent for downloads and publishing, prohibits automatic installation, and cautions against unsupported causal claims. No evidence found of prompt injection or unauthorized data transfer; external PTS implementations and hosted services were not audited.

## Stats

- Views: 0
- Downloads: 2
- Favorites: 0
- Popularity score: 0