@indexyz/pi-model-trace

Attribute the model actually serving pi via ModelTrace numeric fingerprinting.

Install

pi install npm:@indexyz/pi-model-trace

Releases are published to npm as @indexyz/pi-model-trace; each release is tagged @indexyz/pi-model-trace@<version>.

Requires Pi 1.5.0 or newer.

To try a repository checkout (after npm install at the repository root):

pi -e ./pi-model-trace/index.ts

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. Probes use the current Pi installation and preserve the session's thinking level, including off; PI_MODEL_TRACE_BIN explicitly overrides the probe executable. Each probe has a ten-minute timeout and a one-MiB combined stdout/stderr limit. Failed or truncated output does not contribute to attribution. Switching or closing the session cancels active probes without retries or a late report.

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, 34 candidates across GPT, Claude, DeepSeek, Doubao, Gemini, GLM, Grok, and Kimi), 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.
• Probe processes load persisted and project resources independently. Providers registered only in the parent session's memory are not available in probes.
• 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 run check --workspace @indexyz/pi-model-trace
npm run test:e2e -- --package @indexyz/pi-model-trace

The command regression tests run real Pi RPC sessions and probe subprocesses against a local OpenAI-compatible HTTP endpoint; they need no model credentials.