Detection

ExamplesCopy Anchor Link

# Pass the connection settings explicitly
npx lmfpd@latest -b https://api.example.com/v1 -k sk-xxx -m gpt-6-astra

# Read the connection settings from the environment; use Chat Completions and set the reasoning effort
npx lmfpd@latest -a cc -e high

# Run 5 detection rounds
npx lmfpd@latest -n 5

# Send only one request per round
npx lmfpd@latest -b https://api.example.com/v1 -k sk-xxx -m gpt-6-astra --count 1

# Strict mode without streaming; save the full report (credentials excluded)
npx lmfpd@latest -s -ns --timeout 180 --output result.json

# Analyze a saved report again offline without calling a model
npx lmfpd@latest --input result.json --json

# Also run the tokenizer probe (about 12 more requests)
npx lmfpd@latest -a cc --tokenizer

OptionsCopy Anchor Link

An origin without a path, such as https://api.example.com, uses /v1. A base URL with a path preserves that path and appends the selected endpoint. Messages requests use x-api-key and anthropic-version headers. Messages sends the service tier as speed with the anthropic-beta: fast-mode-2026-02-01 header; Responses and Chat Completions send it as service_tier. Detection requests omit the optional output token limit for Responses and Chat Completions. Messages requests keep max_tokens: 8192, which that API requires.

Sampling behaviorCopy Anchor Link

Each round generates --count random challenges with the shared challenge generator (default: 3). The prompts stay unchanged to keep the sampling method consistent. -p 1 only serializes the requests; -n 1 runs only one round. For a fixed single challenge, use --count 1 --challenges one-challenge.json. The file must contain a one-element array whose element has id, prompt, and expected_count. A mismatched array length is rejected before any requests are sent.

A naturally completed sample must contain at least max(80, ceil(expected_count × 0.55)) valid numbers, matching the scorer. A refusal, provider error, or unfinished response is rejected unless relaxed mode already stopped it at the requested count. Failed samples remain visible and are preserved in the results.

All samples in a round, including failed requests, must settle before the next round starts. The concurrency limit applies per round and is capped at the sample count. Detection does not retry automatically. A failed round does not prevent later rounds from running.

The timeout starts when each request is dispatched. For SSE, the first nonempty response-body chunk clears the timer, including a heartbeat or metadata event. Response headers alone do not clear it. No client deadline or Bun idle timeout remains after SSE starts. In JSON mode, the entire response must arrive before the deadline.

Press q or Ctrl+C to cancel active requests. Queued requests and later rounds do not start. With --output, samples received before cancellation are saved. If a tokenizer probe is still running after the last round, the same keys stop only the probe, and all results stay. Redirected output contains a static report, and progress goes to stderr: each sample, each round's leading candidate as soon as the round is scored, and the probe result with --tokenizer. --json keeps stdout limited to machine-readable results.

Results and offline recomputationCopy Anchor Link

bun run fpd --output result.json
bun run fpd --input result.json
bun run fpd --input result.json --strict --json

--input accepts a saved report, a report with outputs, an array of one to three { "text": "...", "expected_count": 300 } objects, or the result.json of a sampling batch. Sampling batches are grouped by condition into rounds of up to three answers. Reports with several rounds are analyzed one round at a time. Strict recomputation rejects samples recorded as truncated or with unknown completion. A plain output array cannot verify its original network completion status. --count applies only to online detection and cannot be combined with --input.

When all three samples are valid and the calibration matches the detector, the result is labeled Calibrated ranking and each candidate has a confidence value. When the calibration does not match, the result is labeled Uncalibrated ranking: it gives the ranking only, and confidence is unavailable. The terminal interface shows the latest ranking and summaries of recent rounds. After at least one scored round, a dimmed final line links the GitHub repository. JSON retains every round. The "most frequent candidate" counts round winners; it is not a combined probability. Confidence is relative to models in the reference bank and does not establish the upstream model's identity. When a custom reference bank does not match its detector, detection stops with an error that asks you to run fpd retrain --data-dir DIR.

Each round is scored from the number fingerprint alone, when its samples finish. With --tokenizer, the tokenizer probe result shows in a separate block as reference only. It does not change the candidates, their order, or the confidence. With --input, a tokenizer probe saved in the report shows in the same block. Reading it sends no requests.

If the smallest valid number of a sample is 200 or more, the interface warns that the sample has an abnormal distribution and the result is unreliable. The prompt causes it: with random challenges, a new run draws new prompts; with --challenges, replace the matching prompts in the file.

Exit codesCopy Anchor Link