@indexyz/pi-model-trace

Attribute the model actually serving pi via ModelTrace numeric fingerprinting.

Usage

/model-trace                # probe the currently selected model
/model-trace provider/id    # probe a specific registered model

The command runs three independent probes against the target model. Each probe is executed in its own non-persisted pi session — a fresh pi -p --no-session --no-tools -a child process with an empty context, no session persistence, and all tools disabled, so the model has to answer the numeric challenge directly and cannot reuse prior context or call tools. Transport-level probe failures are retried once; the pi binary defaults to pi on PATH and can be overridden with the PI_MODEL_TRACE_BIN environment variable.

Each challenge asks the model to emit ~300 integers between 1 and 355. The raw answers are scored locally against the bundled ModelTrace unified fingerprint bank (data/unified_bank.json, 13 GPT/Claude candidates), which computes:

0.75 × nuisance-projected Hellinger centroid similarity
+ 0.25 × ordered-block digit-sequence feature

The three scores are averaged and converted to closed-set model and family probabilities through the bank's calibrated temperature. The result is appended to the transcript as a model-trace message.

Notes

• Results are closed-set probabilities within the bundled bank only — an unlisted model is attributed to its nearest candidate. Treat them as a reference signal, not definitive proof.
• The model still sees pi's system prompt during probes; providers or harnesses whose prompts differ strongly from the bank's采集环境 can skew attribution.
• Probes consume the target model's inference quota (three ~300-token answers).
• The fingerprint core and challenge generator are TypeScript ports of upstream fingerprint-core.mjs / fingerprint.py (MIT License, © 2026 xqy2006).

Development

npm test
npm run typecheck