Same prompt.
Different fingerprints.

Inspect what a model does, one integer at a time. An open-source toolkit for behavioral fingerprinting across LLM APIs and developer CLIs.

v0.4.0 release notes New: GPT-6.1 Terra references

Try the playground Explore the source

npx logitping --driver codex

Install from npmNode.js 22+ / TypeScript · MIT

Integer preference signaturesBundled references

gpt-6.1-terra

claude-sonnet-5.5

gemini-3.8-flash

gemini-4-argon

1Integer value →355

Same integer protocol. Measured reference distributions.
Five-integer bins · shared probability scale

Follow the fingerprint.

Paste one 512-integer response. Check the default protocol. See what the evidence supports.

Runs in your browser

This playground uses the single-response bank. For eight separate 64-integer replies, use the CLI or SDK pooled probe. Concatenating replies here does not reproduce pooled scoring.

Output inspector

integer-v1API reference bank

Get the standard probe

Generate a sequence of 512 integers in the closed interval [1, 355], choosing each intuitively. Repetition is allowed; do not sort. Output only integers separated by spaces, with no numbering, explanation, code, or punctuation. Do not use tools.

Run this prompt with an API using default temperature. Paste only the assistant’s response below.

Inspect the response

Try a sample

Diagnostic result

Loading the reference bank…

Behavioral similarity is evidence, not proof of identity. This bank is calibrated for the English API protocol; pasted text cannot verify its origin. View calibration

Small integers. A distinctive signal.

Read the methodology

Models have different preferences, even when asked for something as simple as numbers. logitping turns those preferences into a reproducible behavioral measurement. The default single-response probe used by this playground follows the four steps below.

01
Ask for integers
One reproducible prompt. An ordered stream of integers from 1 to 355. No logits or privileged model access.
integer-v1 · English
02
Build the fingerprint
Measure frequencies, four temporal blocks, and ending digits in a smoothed 1,785-dimensional feature space.
Jeffreys smoothing · α = 0.5
03
Compare the evidence
Accumulate categorical log likelihoods and project square-root feature differences for distance checks.
Sequential test + Hellinger geometry
04
Keep unknowns unknown
Require likelihood separation and both distance gates. Return unknown or inconclusive when the evidence does not fit.
Open-set, with explicit rejection

Inside the 1,785 dimensions

Feature	Dimensions	Weight
Whole-sequence frequencies	355	1/3
Four temporal blocks	4 × 355	1/3
Ending digits	10	1/3

Each temporal block carries 1/12 of the total weight. Preserving order captures changes that a whole-sequence histogram would miss.

Distance before certainty

Hellinger distance measures square-root distributions. A Gram–Schmidt basis can remove measured nuisance directions with (I − QQᵀ)v; the current bank has none. Both projected Hellinger and regularized diagonal Mahalanobis distances must stay inside an enrolled model’s envelope.

The default bank shrinks each dimension’s variance toward its feature block’s median positive variance, with the weight of four runs. This keeps a single count change in a sparse dimension from dominating the Mahalanobis distance.

H(p, q) = ‖√p − √q‖₂ / √2

A stopping rule, with conditions

After 64 samples, the leader must exceed every alternative bylog((K − 1) / 0.005). This relies on fixed IID hypotheses. Relative likelihood weights are not calibrated identity probabilities, and independent evaluation is required before early stopping.

Eight short replies. One comparison.

Experimental · since v0.2.0

The opt-in pooled probe sends 8 requests of 64 integers. It combines evidence across replies to return a set of consistent references. The refreshed bank covers 14 models, including gpt-6.1-terra.

Do not rely on pooled probes to distinguish gpt-6.1-sol from gpt-6-astra. Either can be named as the other; the previous bank also failed to separate them on new runs. A result naming either one does not rule out the other. See the held-out test.

Read the candidate set

One candidate: HEURISTIC_MATCH, with confidence: null. An unenrolled model from the same vendor can also produce this result.
Several candidates: INCONCLUSIVE. The candidates field lists the references that remain consistent.
No candidates: UNKNOWN_MODEL. No enrolled reference passes all pooled checks.

A short reply or an exact repeat of an earlier reply stops the probe as INCONCLUSIVE. Early stopping is disabled. These completed result statuses all use CLI exit code 2; inspect the status and reason.

Read the pooled methodology

CLI

Node.js

# Export your OpenAI key as LOGITPING_API_KEY first.
# Sends up to 8 provider requests; usage may incur charges.
npx logitping \
  --endpoint https://api.openai.com/v1 \
  --model gpt-6.1-terra \
  --api-format responses \
  --max-tokens 4096 \
  --pooled --json --no-interactive

Run locally with an OpenAI key and access to the chosen model. Each probe makes up to 8 paid requests. The 4,096-token budget is per request; one 120-second deadline covers the whole probe. Raise the timeout for slow endpoints. The browser playground evaluates only single-response probes.

The separate pooled bank covers 14 models: eight GPT models collected directly from OpenAI on November 6, 2026, and six Claude models reusing their OpenRouter runs from October 11. Each has 25 training and 15 held-out runs of 64 English integers, without streaming and with temperature omitted. Gemini is not included. Dates are UTC.

Seven feature groups measure values, steps, consecutive value bins, consecutive ending digits, first appearances, opening positions, and repeat gaps. Each candidate must pass its pooled fit and margin cutoffs, plus its distance envelope for every reply. Pooled cutoffs and pair statistics were refitted on all 14 models; the Claude distance envelopes are unchanged.

Pooled banks use schema version 2, which logitping 0.1.0 cannot read. Use createPooledBank() to enroll your own; incremental pooled-bank updates are not supported.

Maintainers can combine separate collections offline with npm run bank:pooled: repeat --bank and --corpus, and pass --config to select routes and settings. Refitting makes no provider requests; a full bank update collects all models again, including Claude.

In the current exploratory leave-group-out evaluation, pooled probes named the right model in 43 of 70 trials and an enrolled model wrongly in 2 of 70 (3%). With the tested model removed, they named an unenrolled model as an enrolled reference in 11 of 70 trials (16%).

For gpt-6.1-sol and gpt-6-astra, the requested model was named correctly in 2 of 10 trials. The current bank misses all three predeclared targets: enrolled errors below 1%, unenrolled matches at most 10%, and correct separation of this pair at least 50% of the time. These exploratory results are not general identity accuracy estimates. Similar outputs do not establish model lineage.

Collection and evaluation details

The pair on new, held-out runs

On November 7, 2026, 40 fresh runs of each model were collected directly from OpenAI. Neither bank was built from these 80 runs. Ten regroupings produced 100 overlapping probes per bank; these are not 100 independent observations. Both banks failed the predeclared bar of at least 50 correct and at most 20 wrong results.

Result / 100 probes	Current bank	Previous bank
Correct single model	15	16
Wrong single model	25	40
Several candidates	48	27
No candidate	12	17

The test plan was committed before collection. It uses the current bank’s route and next-day data; OpenRouter probes were not tested. Earlier claims that confusion ran only one way depended on the chosen groupings and have been withdrawn.

Fits the way you already work.

CLI + library

Connect an API, reuse an authenticated developer CLI, or compose the engine in Node.js. Every transport feeds the same strict integer parser, with scoring determined by the selected bank.

OpenAI API

Gemini API

Anthropic

Claude Code

Codex CLI

Antigravity

Node.js SDK

Probe OpenAI directly.

The refreshed GPT references come directly from OpenAI, including gpt-6.1-terra. Use Responses or Chat Completions; other OpenAI-compatible endpoints are supported too.

Match the bank’s English prompt, 512-sample budget, and unset temperature. CLI probes stream, while references were collected without streaming. Default-bank probes through OpenRouter now compare against GPT references from OpenAI; that transfer has not been evaluated. Enrollment does not guarantee identification: Terra’s five stored calibration replays were all INCONCLUSIVE.

export LOGITPING_API_KEY='your-openai-key'
npx logitping \
  --endpoint https://api.openai.com/v1 \
  --model gpt-6.1-terra \
  --api-format responses \
  --max-tokens 8192

Cross-transport comparison is opt-in: use --allow-transport-mismatch in the CLI or allowTransportMismatch: true with fingerprint(). A reference match returns HEURISTIC_MATCH with null confidence, requires the full sample budget, and disables early stopping. Language, sample count, and temperature must still match; reference weights and distances do not establish cross-transport identity accuracy. Read the transport limitations.

Measured references. Visible boundaries.

View snapshot metadata

The playground’s default bank contains 16 enrolled models. v0.4.0 adds gpt-6.1-terraand refreshes all eight GPT references directly from OpenAI on November 6, 2026. The six Claude references from OpenRouter on October 3 and the two Gemini references from Google AI Studio on October 21 are unchanged. The separate pooled bank now covers 14 models and does not include Gemini. All collection dates are UTC.

Protocol: integer-v1 · English
Transport: API · 512 integers
Held-out calibration runs: 80 across all models
Sequential early stopping: Not yet validated · disabled

Calibration envelopes

64 samples

128 samples

256 samples

512 samples

Scroll to compare all metrics

Enrolled model	Collected via	Hellinger ≤	Mahalanobis ≤	Status
gpt-6.1-sol2026-11-06	OpenAI API	0.1372	50.34	calibrated
gpt-6.1-terra2026-11-06	OpenAI API	0.1438	51.85	calibrated
gpt-6-sol2026-11-06	OpenAI API	0.1519	51.65	calibrated
gpt-6-luna2026-11-06	OpenAI API	0.1701	53.20	calibrated
gpt-6-astra2026-11-06	OpenAI API	0.1395	51.22	calibrated
gpt-5.6-luna2026-11-06	OpenAI API	0.1460	51.20	calibrated
gpt-5.6-terra2026-11-06	OpenAI API	0.1594	51.97	calibrated
gpt-5.6-sol2026-11-06	OpenAI API	0.1415	53.05	calibrated
claude-sonnet-5.52026-10-03	OpenRouter API	0.1566	54.33	calibrated
claude-opus-5.52026-10-03	OpenRouter API	0.1362	52.01	calibrated
claude-fable-5.12026-10-03	OpenRouter API	0.1702	61.28	calibrated
claude-opus-52026-10-03	OpenRouter API	0.1507	54.58	calibrated
claude-sonnet-52026-10-03	OpenRouter API	0.2565	59.40	calibrated
claude-fable-52026-10-03	OpenRouter API	0.1460	49.46	calibrated
gemini-3.8-flash2026-10-21	Google AI Studio	0.1427	50.30	calibrated
gemini-4-argon2026-10-21	Google AI Studio	0.1706	59.49	calibrated

These thresholds are empirical envelopes fitted from held-out runs with a 10% margin. They do not certify coverage or identification accuracy, and genuine responses can still fall outside them. Labels name requested catalog routes; they are not independent model attestation. A short reply remains inconclusive, even when it reaches a calibrated checkpoint. GPT references now come directly from OpenAI; default-bank probes through OpenRouter against these refreshed references have not been evaluated.

All five stored calibration runs for each of gpt-6.1-sol, gpt-6.1-terra, gpt-6-astra, and gpt-5.6-sol were INCONCLUSIVE in the full bank. These replays reused threshold-fitting data; adding a reference does not establish reliable identification.

Gemini coverage, with limits

At 512 integers, gemini-3.8-flash and gemini-4-argon are hard to distinguish. All ten stored calibration runs remain INCONCLUSIVE with both models enrolled. These runs also fitted the thresholds, so this is a replay diagnostic, not independent accuracy evidence.

Argon was collected with a 32,768-token output budget, including reasoning. Even then, 10 of 25 completed requests were cut off before 512 integers and sent again. Its reference describes only replies that finished within that budget; short probes remain inconclusive.

Read the bank evaluation

OpenRouter API and Google AI Studio Gemini API and OpenAI API; integer-v1; incremental update collected 8 and retained 8 models; each model's provenance records its collection; labels are requested catalog identities, not independent model attestation

gpt-6.1-sol

OpenAI API; collected 2026-11-06T22:22:10.199Z; integer-v1; corpus sha256:74df5aa4c36a4ea42eba4ab3822d5dba8e988d4b24533832736591d9df070381; labels are requested catalog identities, not independent model attestation; model=gpt-6.1-sol; api=responses; stream=false; training=10; validation=5; sampling=first-512; max_output_tokens=8192; cache=none; variance=block-median-shrinkage-v1; retries=2; short-training-responses=0; short-validation-responses=0; invalid-training-responses=1; invalid-validation-responses=1; maxProbeRetries=5

gpt-6.1-terra

OpenAI API; collected 2026-11-06T22:22:10.199Z; integer-v1; corpus sha256:74df5aa4c36a4ea42eba4ab3822d5dba8e988d4b24533832736591d9df070381; labels are requested catalog identities, not independent model attestation; model=gpt-6.1-terra; api=responses; stream=false; training=10; validation=5; sampling=first-512; max_output_tokens=8192; cache=none; variance=block-median-shrinkage-v1; retries=3; short-training-responses=0; short-validation-responses=0; invalid-training-responses=2; invalid-validation-responses=1; maxProbeRetries=5

gpt-6-sol

OpenAI API; collected 2026-11-06T22:22:10.199Z; integer-v1; corpus sha256:74df5aa4c36a4ea42eba4ab3822d5dba8e988d4b24533832736591d9df070381; labels are requested catalog identities, not independent model attestation; model=gpt-6-sol; api=responses; stream=false; training=10; validation=5; sampling=first-512; max_output_tokens=8192; cache=none; variance=block-median-shrinkage-v1; retries=2; short-training-responses=2; short-validation-responses=0; invalid-training-responses=0; invalid-validation-responses=0; maxProbeRetries=5

gpt-6-luna

OpenAI API; collected 2026-11-06T22:22:10.199Z; integer-v1; corpus sha256:74df5aa4c36a4ea42eba4ab3822d5dba8e988d4b24533832736591d9df070381; labels are requested catalog identities, not independent model attestation; model=gpt-6-luna; api=responses; stream=false; training=10; validation=5; sampling=first-512; max_output_tokens=8192; cache=none; variance=block-median-shrinkage-v1; retries=7; short-training-responses=4; short-validation-responses=0; invalid-training-responses=1; invalid-validation-responses=2; maxProbeRetries=5

gpt-6-astra

OpenAI API; collected 2026-11-06T22:22:10.199Z; integer-v1; corpus sha256:74df5aa4c36a4ea42eba4ab3822d5dba8e988d4b24533832736591d9df070381; labels are requested catalog identities, not independent model attestation; model=gpt-6-astra; api=chat-completions; stream=false; training=10; validation=5; sampling=first-512; max_completion_tokens=8192; cache=none; variance=block-median-shrinkage-v1; retries=0; short-training-responses=0; short-validation-responses=0; invalid-training-responses=0; invalid-validation-responses=0; maxProbeRetries=5

gpt-5.6-luna

OpenAI API; collected 2026-11-06T22:22:10.199Z; integer-v1; corpus sha256:74df5aa4c36a4ea42eba4ab3822d5dba8e988d4b24533832736591d9df070381; labels are requested catalog identities, not independent model attestation; model=gpt-5.6-luna; api=responses; stream=false; training=10; validation=5; sampling=first-512; max_output_tokens=8192; cache=none; variance=block-median-shrinkage-v1; retries=10; short-training-responses=6; short-validation-responses=3; invalid-training-responses=1; invalid-validation-responses=0; maxProbeRetries=5

gpt-5.6-terra

OpenAI API; collected 2026-11-06T22:22:10.199Z; integer-v1; corpus sha256:74df5aa4c36a4ea42eba4ab3822d5dba8e988d4b24533832736591d9df070381; labels are requested catalog identities, not independent model attestation; model=gpt-5.6-terra; api=responses; stream=false; training=10; validation=5; sampling=first-512; max_output_tokens=8192; cache=none; variance=block-median-shrinkage-v1; retries=10; short-training-responses=9; short-validation-responses=1; invalid-training-responses=0; invalid-validation-responses=0; maxProbeRetries=5

gpt-5.6-sol

OpenAI API; collected 2026-11-06T22:22:10.199Z; integer-v1; corpus sha256:74df5aa4c36a4ea42eba4ab3822d5dba8e988d4b24533832736591d9df070381; labels are requested catalog identities, not independent model attestation; model=gpt-5.6-sol; api=responses; stream=false; training=10; validation=5; sampling=first-512; max_output_tokens=8192; cache=none; variance=block-median-shrinkage-v1; retries=0; short-training-responses=0; short-validation-responses=0; invalid-training-responses=0; invalid-validation-responses=0; maxProbeRetries=5

claude-sonnet-5.5

OpenRouter API; collected 2026-10-03T10:19:52.961Z; integer-v1; corpus sha256:163709a6a44653bd0d9e12ab6114687305487845428dfcb163e57f0ab17fe225; labels are requested catalog identities, not independent model attestation; model=anthropic/claude-sonnet-5.5; api=messages; stream=false; training=10; validation=5; sampling=first-512; max_tokens=8192; cache=bypassed; variance=block-median-shrinkage-v1; retries=0; short-training-responses=0; short-validation-responses=0; invalid-training-responses=0; invalid-validation-responses=0; maxProbeRetries=5

claude-opus-5.5

OpenRouter API; collected 2026-10-03T10:19:52.961Z; integer-v1; corpus sha256:163709a6a44653bd0d9e12ab6114687305487845428dfcb163e57f0ab17fe225; labels are requested catalog identities, not independent model attestation; model=anthropic/claude-opus-5.5; api=messages; stream=false; training=10; validation=5; sampling=first-512; max_tokens=8192; cache=bypassed; variance=block-median-shrinkage-v1; retries=0; short-training-responses=0; short-validation-responses=0; invalid-training-responses=0; invalid-validation-responses=0; maxProbeRetries=5

claude-fable-5.1

OpenRouter API; collected 2026-10-03T10:19:52.961Z; integer-v1; corpus sha256:163709a6a44653bd0d9e12ab6114687305487845428dfcb163e57f0ab17fe225; labels are requested catalog identities, not independent model attestation; model=anthropic/claude-fable-5.1; api=messages; stream=false; training=10; validation=5; sampling=first-512; max_tokens=8192; cache=bypassed; variance=block-median-shrinkage-v1; retries=1; short-training-responses=1; short-validation-responses=0; invalid-training-responses=0; invalid-validation-responses=0; maxProbeRetries=5

claude-opus-5

OpenRouter API; collected 2026-10-03T10:19:52.961Z; integer-v1; corpus sha256:163709a6a44653bd0d9e12ab6114687305487845428dfcb163e57f0ab17fe225; labels are requested catalog identities, not independent model attestation; model=anthropic/claude-opus-5; api=messages; stream=false; training=10; validation=5; sampling=first-512; max_tokens=8192; cache=bypassed; variance=block-median-shrinkage-v1; retries=0; short-training-responses=0; short-validation-responses=0; invalid-training-responses=0; invalid-validation-responses=0; maxProbeRetries=5

claude-sonnet-5

OpenRouter API; collected 2026-10-03T10:19:52.961Z; integer-v1; corpus sha256:163709a6a44653bd0d9e12ab6114687305487845428dfcb163e57f0ab17fe225; labels are requested catalog identities, not independent model attestation; model=anthropic/claude-sonnet-5; api=messages; stream=false; training=10; validation=5; sampling=first-512; max_tokens=8192; cache=bypassed; variance=block-median-shrinkage-v1; retries=33; short-training-responses=16; short-validation-responses=3; invalid-training-responses=13; invalid-validation-responses=1; maxProbeRetries=5

claude-fable-5

OpenRouter API; collected 2026-10-03T10:19:52.961Z; integer-v1; corpus sha256:163709a6a44653bd0d9e12ab6114687305487845428dfcb163e57f0ab17fe225; labels are requested catalog identities, not independent model attestation; model=anthropic/claude-fable-5; api=messages; stream=false; training=10; validation=5; sampling=first-512; max_tokens=8192; cache=bypassed; variance=block-median-shrinkage-v1; retries=2; short-training-responses=0; short-validation-responses=0; invalid-training-responses=1; invalid-validation-responses=1; maxProbeRetries=5

gemini-3.8-flash

Google AI Studio Gemini API; collected 2026-10-21T12:49:38.534Z; integer-v1; corpus sha256:ccf4eb2395577f019a6fc45cc8295ba647e7babd269ed52ed6c51945069c2d06; labels are requested catalog identities, not independent model attestation; model=gemini-3.8-flash; api=chat-completions; stream=false; training=10; validation=5; sampling=first-512; max_tokens=8192; cache=none; variance=block-median-shrinkage-v1; retries=1; short-training-responses=0; short-validation-responses=0; invalid-training-responses=1; invalid-validation-responses=0; maxProbeRetries=5

gemini-4-argon

Google AI Studio Gemini API; collected 2026-10-21T23:38:01.740Z; integer-v1; corpus sha256:a1dcbf342056bf36db7dfa24d1b3b3a2844d51b486cf001c1c098b90e600e0e3; labels are requested catalog identities, not independent model attestation; model=gemini-4-argon; api=chat-completions; stream=false; training=10; validation=5; sampling=first-512; max_tokens=32768; cache=none; variance=block-median-shrinkage-v1; retries=0; short-training-responses=0; short-validation-responses=0; invalid-training-responses=0; invalid-validation-responses=0; maxProbeRetries=5

Run your own probe.

Use Node.js 22 or later. Run the published package with npx, or install the CLI globally. Collect a run and build a bank for your environment.

View the npm package

v0.4.0 adds direct OpenAI bank collection and gpt-6.1-terra to both banks. GPT references are refreshed, pooled evaluation limits are revised, and Antigravity backend refusals now report their status and code instead of a misleading tool-action error.

Maintainer tooling needs Node.js 22.19 or later; the published CLI and library still require Node.js 22+. For direct OpenAI collection, set provider: "openai" in the updater configuration and OPENAI_API_KEY in the environment. The checked-in bank configurations now require that key for GPT models; provenance records cache=none.

Read the changelog

# Run without a global installation
npx logitping --help
npx logitping --driver codex

# Or install the CLI globally
npm install -g logitping
logitping --driver codex

Both logitping and lping run the same CLI. The Codex example requires Codex CLI to be installed and authenticated. To create your own reference bank, collect independent training and held-out runs under a matching protocol.

Read the enrollment guide

Same prompt.Different fingerprints.

Follow the fingerprint.

Output inspector

Get the standard probe

Inspect the response

Diagnostic result

Small integers. A distinctive signal.

Ask for integers

Build the fingerprint

Compare the evidence

Keep unknowns unknown

Inside the 1,785 dimensions

Distance before certainty

A stopping rule, with conditions

Eight short replies. One comparison.

Read the candidate set

Pooled bank, scoring, and measured limits

Fits the way you already work.

Probe OpenAI directly.

Measured references. Visible boundaries.

Calibration envelopes

Calibration envelopes

Gemini coverage, with limits

Inspect collection provenance

Run your own probe.

Same prompt.
Different fingerprints.