Data files

Added 36 fixed-challenge answers for gemini-4-argon. The reference bank now contains 59 fingerprint identities and 2,200 samples. Collection requests google/gemini-4-argon through streaming OpenRouter Chat Completions, with reasoning_effort=low and the fixed openrouter/google-ai-studio route. Fallbacks are disabled. All 36 answers passed validation on the first attempt. Of these, 33 completed naturally. query-10, query-25, and query-36 stopped at 500 integers and retain their truncation markers.

Returned model IDs and Google AI Studio provider metadata passed checks against the raw responses. Sampling evidence and the enrollment receipt are saved in collections/gemini-4-argon-openrouter-google-ai-studio-low-20261014/. This enrollment also adds two independent fixed evaluation groups of three answers, with the same provider and low reasoning effort. Evaluation answers do not enter bank construction, training, or calibration.

After retraining, both new evaluation groups ranked gemini-4-argon first. Predictions on the previous 116 complete groups did not change: 105 were correct before and after enrollment. Evaluation covers 118 of 120 planned groups and 354 of 360 answers, with 107 of 118 complete groups correct. The two missing groups belong to kimi-k2.8-preview.

Added 36 fixed-challenge answers for muse-wonder. The reference bank now contains 58 fingerprint identities and 2,164 samples. Collection requests meta/muse-wonder through streaming OpenRouter Chat Completions, with reasoning_effort=low and the fixed openrouter/meta route. Fallbacks are disabled. All 36 answers completed naturally on the first attempt, without truncation. Returned model IDs and Meta provider metadata passed checks against the raw responses.

Sampling evidence and the enrollment receipt are saved in collections/muse-wonder-openrouter-meta-low-20261013/. This enrollment also adds two independent fixed evaluation groups of three answers, with the same provider and low reasoning effort. Evaluation answers do not enter bank construction, training, or calibration.

The detector parameters are now stored in shared_detector.bin: a JSON header plus float32 values. The website downloads 3.37 MB for the detector instead of 9.75 MB. Holdout predictions and rankings did not change (116/116 groups identical; maximum confidence change 1.8e-7). import_manifest.json was removed.

The maintainer confirmed that muse-spark-1.3-contributor and muse-spark-1.3 are the same model. The two reference batches of the former (36 answers, channel openrouter/unknown) were removed. The bank now contains 57 fingerprint identities and 2,128 samples. The two reference rows before the removal can be retrieved from the git history.

Before the removal, leave-one-environment-out checks on the reference bank separated only 33–37 of 72 answers between the two labels. This result did not differ from labels shuffled within each environment. The fixed evaluation set keeps the collection records of this model. The evaluator lists it as a retired model and does not count it in the planned groups.

shared_detector.json now uses shared-detector-v2. The ranking score gives equal weights to the frequency ranking and the position/order discriminant. See Ranking and calibration for the method. Leave-one-environment-out cross-validation on the 58-identity reference bank selected the weights and the shrinkage from a grid declared in advance. The previous ranker scored 631 of 696 groups, and the new ranker scored 642 of 696. A joint generation process model added only 1 group as a third component, and the nested selection did not choose it consistently. It is not part of the product.

Retraining after the removal of muse-spark-1.3-contributor gave nested top-1 of 641/684, binary NLL of 0.138 (constant baseline 0.235), AUC of 0.931, and temperature τ of 4.781. Fixed evaluation scored 103 of 114 complete groups. On the same 114 groups, the previous ranker retrained on the same bank scored 100 of 114. The fixed evaluation set was used only for evaluation, not for selection or calibration.

In fixed evaluation, 13 answers had a smallest valid number of at least 200 (from claude-haiku-4.5, gpt-6-sol, gpt-5.4, and gpt-5.6-terra). At the maintainer's request, these slots were retested with new prompts of the same count, with the official provider pinned. All earlier attempts are kept. The smallest values of the retested answers are 1–103.

At the maintainer's request, gpt-6-astra and gpt-6.1-sol share one fingerprint class. Its internal ID is gpt-6-astra, and the interface shows it as gpt-6-astra/6.1-sol (batch field model.display_name). Both batches contribute their 36 answers to the bank and ranker training. The bank now contains 58 fingerprint identities and 2,164 samples. Requested models, returned models, sample IDs, collection channels, and raw evidence retain their original values. The two reference rows before the rename can be retrieved from the git history.

Detection reports the combined class and no longer ranks the two names separately. When this class ranks first, the website notes that the number fingerprint cannot reliably tell the two models apart. Fixed evaluation retains all four groups and 12 answers from the two requested models. It displays their original collection labels and maps both truths to the combined class. Predictions that confused these names count as correct under the combined labels. A change in hit count therefore includes the effect of merging labels and does not directly show better model separation.

Added 36 fixed-challenge answers for claude-haiku-5.5. The reference bank now contains 59 models and 2,164 samples. The requested model is anthropic/claude-haiku-5.5. Collection uses streaming OpenRouter Chat Completions, reasoning_effort=low, and the fixed openrouter/anthropic channel with fallbacks disabled. Prompts and system prompts come from the fixed challenge suite.

The first three local requests returned HTTP 403 because of a regional restriction. Collection then resumed the same batch on Ikaleio-TYO. All 36 answers completed naturally, without truncation. The returned model and Anthropic provider metadata passed checks against the raw responses. Sampling evidence, the three failed attempts, and the enrollment receipt are saved in collections/claude-haiku-5.5-openrouter-anthropic-low-20261008/. The existing fixed evaluation set does not yet cover this new model.

Added seed-2.1-turbo, seed-2.0-lite, seed-2.0-mini, qwen3.8-flash, and qwen3.8-omni-flash, each with 36 fixed-challenge answers. The reference bank now contains 58 models and 2,128 samples. All five batches use the same fixed prompts and streaming OpenRouter Chat Completions. Seed is pinned to openrouter/seed/fp8 and Qwen to openrouter/alibaba, with fallbacks disabled.

Collection initially used default reasoning effort. Turbo's seventh attempt at query-15 used low. At the user's request, remaining pending challenges then used none to disable reasoning: 13 Turbo answers, 29 Flash answers, and 6 Omni-Flash answers. Previously accepted answers retain their settings; prompts and prefixes did not change. Actual parameters, selected attempts, original responses, failures, and enrollment receipts are saved in collections/seed-qwen-20261003/. The earlier rate-limited Flash batch is retained but was not enrolled.

All batches stop automatically at 500 integers and retain the truncated status: 7 Turbo answers, 7 Lite answers, 4 Mini answers, 4 Flash answers, and 4 Omni-Flash answers. All other answers completed naturally. Enrollment checks returned models and provider metadata in the raw responses. Detection rankings do not select the answers. The existing fixed evaluation set does not yet cover these five new models.

The OpenAI o200k_base mapping in tokenizer_bank.json now covers every model whose name starts with gpt-5 or gpt-6 (confirmed by the maintainer). gpt-5 already covered gpt-5.x. Before this change, the only gpt-6 model with a mapping was gpt-6-astra, through an alias of the o200k class; that alias is removed. These models are checked as a vendor-published tokenizer, so only an exact match is consistent. The ranking does not change. This is a manual edit of a generated file: before you run token-fp export-api-bank again, add these mappings to the generator.

All 16 codex-subscription batches (216 answers, reasoning effort not recorded) of gpt-5.4, gpt-5.5, gpt-5.6-sol, gpt-5.6-terra, gpt-5.6-luna, and gpt-6-astra left the official bank. The original rows can be retrieved from the git history. A paired study (research/studies/astra-sol-separation/) found that the collection channels and reasoning settings of the earlier reference batches were inconsistent, which interfered with the fingerprint comparison of gpt-6-astra and gpt-6.1-sol. The new batches call the Responses API of api.openai.com directly from a local machine with an OpenAI key, streaming with reasoning_effort=low, the same as gpt-6-sol. Each completed the 36 fixed challenges with the same prompts and system prompts as the earlier batches, and the evidence is saved in collections/<model>-openai-direct-20260930/. The returned models are gpt-5.4-2026-03-05 and gpt-5.5-2026-04-23; the others match the requested names. Truncated at the sampling limit: 7 answers of gpt-5.6-luna, 1 of gpt-5.6-terra, and 1 of gpt-5.4; all others completed naturally. All failed attempts are kept: 4 dropped connections for gpt-6-astra and 3 for gpt-5.4, and too few valid numbers once for gpt-5.6-terra and twice for gpt-5.6-luna. Each of them succeeded on the next attempt.

Added 36 fixed-challenge answers for gpt-6.1-sol, with the evidence in collections/gpt-6.1-sol-openrouter-openai-20260930/. The requests used the OpenRouter model openai/gpt-6.1-sol with streaming Chat Completions, reasoning_effort=low, and the fixed openrouter/openai channel (no fallback). All 36 answers were checked against the returned model and OpenAI provider metadata; all completed naturally, with no failures or truncation.

On the same day, the Claude reference labels switched to dotted version numbers (for example claude-opus-5.5), and the label of claude-haiku-4-5-20251001 became claude-haiku-4.5. Requested model names and raw sampling evidence keep the upstream values; the 25 rows before the rename can be retrieved from the git history. The gpt-6-sol-20260922 batch left the reference bank; its raw row can also be retrieved from the git history.

Added 36 fixed-challenge answers for claude-sonnet-5.5. The evidence is saved in collections/claude-sonnet-5.5-openrouter-anthropic-20260929/. The requests used the official OpenRouter model anthropic/claude-sonnet-5.5 with Chat Completions, the default reasoning effort, and the fixed openrouter/anthropic channel (fallback disabled). All 36 answers were checked against the returned model and the Anthropic provider metadata. All of them completed naturally, with no failures or truncation.

The original gpt-6-sol reference batch (36 answers collected through openrouter/openai on 2026-09-22) was relabeled gpt-6-sol-20260922. It keeps the original requested model name, channel, and sampling evidence. The new gpt-6-sol batch was collected on Ikaleio-TYO by calling the Responses API of api.openai.com directly with an OpenAI key. It completed the same 36 fixed challenges with reasoning_effort=low. Of these answers, 34 completed naturally and 2 were truncated at the sampling limit. The raw evidence is saved in collections/gpt-6-sol-openai-direct-20260928/. The first Chat Completions requests returned HTTP 400 because the API does not support max_tokens. The 3 failed records are saved in collections/gpt-6-sol-openai-direct-cc-20260928/ and were not enrolled. From then on the two labels were scored separately, until gpt-6-sol-20260922 left the reference bank on 2026-09-30. Because the channels and APIs differ, a difference between model versions cannot be inferred from the reference bank ranking alone.

All 15 batches without a pinned provider (216 answers) for six DeepSeek models were retired from the official bank. The original rows can be retrieved from the git history, and the original sampling attempts stay in the research directory. The new reference batches are saved per model in collections/deepseek-*-20260926/. V4.1 Flash and V4 Pro 0813 each have 36 answers pinned to openrouter/deepseek. Flash 0731, Pro Preview, Flash Preview, and V3.2 each have 36 answers pinned to openrouter/novita. For Flash 0731, the 11th attempt of query-11 and the 12th attempt of query-23 override only reasoning_effort=low, and all original failures are kept. The discontinued SiliconFlow backup batches stay in collections/ but were not enrolled. The official bank includes only complete batches that pass the provenance check. It does not let OpenRouter fall back automatically or mix answers from different providers.

The K3 update replaces the earlier 36 answers without a pinned provider with 36 answers pinned to the MoonshotAI source, with thinking disabled. The new batch is in collections/kimi-k3-moonshotai-20260914/. The old batch can be retrieved from the git history. See the manifest of the new batch for the prefix versions and the effective attempts.

The 1,804 reference rows from before the schema cutover can be retrieved from the git history. The new format keeps all sample IDs, answers, prompts, and their order. An unknown completion status is not turned into success. The runtime no longer reads the old format. Historical requests and failed attempts are still kept.