← s2ars2ar

Certificates: the bands, the witness, the replay

Cody Lee Walker · s2ar · calibration seen 2026-10-10 · a living page: it follows the landing’s calibration line

The bands on this page are the ones the live landing serves (data/calibration/live.json, line sha 2c115bd3cae8…); the tower measures them and this page never carries an older set than the landing. A certificate is a signed Assay record that an image, compressed, still reads the same to a vision model, within a band measured per reader. MEASURED — not argued.

17 models measured; the landing shows 4 demo items at scales ×1.0, ×0.85, ×0.7, ×0.55, ×0.4, ×0.3, ×0.2, ×0.12, ×0.08, ×0.05; the certify walkthrough's record `f164152c1dd55564…` verifies free at `/v1/verify/<sha>`.

1. What a band is

certify compresses an image and has a witness measure what a vision model would still read from it. The witness is deterministic (recon: gradient-energy over the reconstruction, with RMSE for coarse images and OCR for text); the band is how far the witness’s score may drop before that model’s answers start to move, measured per model on the calibration items. Tolerance is per reader: how far the band must tighten so the witness never outruns the model. A strict tolerance is a model that loses detail early.

2. The bands, per model

Strictest tolerance first; typical and max are vision-token savings at the certified scale.

model provider tolerance typical savings max n
Gemini 3 6 Flash Google 0.35 −62% −99% 9
Qwen3 7 Flash OpenAI 0.5 −96% −99% 26
Qwen3 7 Plus OpenAI 0.5 −91% −99% 91
Llama 4 Scout Meta 0.5 −89% −99% 83
Llama 3.2 11B Meta 0.5 −88% −99% 77
Grok 4 5 OpenAI 0.5 −87% −99% 91
Claude Sonnet 4.6 Anthropic 0.6 −91% −99% 88
GPT-4o mini OpenAI 0.6 −89% −99% 76
Claude Fable 5 Anthropic 0.6 −88% −99% 28
Gemini 3.5 Flash Google 0.6 −88% −99% 77
GPT-5.4 mini OpenAI 0.6 −86% −99% 87
GPT 5 6 Luna OpenAI 0.65 −96% −99% 26
GPT 5 6 Terra Pro OpenAI 0.65 −96% −99% 26
Gemini 3 5 Flash Lite Google 0.65 −94% −99% 25
Claude Sonnet 5 Anthropic 0.65 −89% −99% 88
GPT 5 6 Luna Pro OpenAI 0.7 −97% −99% 26
Gemini 3 1 Flash Lite Google 0.75 −96% −99% 81

3. Where the witness over-certifies

The chips on the landing, as a table: the calibration items where recon would pass a scale the model fails. The certificate pulls back on those, so an over-certified item is a tighter band, not a wrong certificate.

model witness tolerance over-certified items which
Llama 4 Scout recon 0.5 7 tiny_code, which_color, dense_config, small_label_number
Grok 4 5 recon 0.5 8 count_circles, tiny_code, dense_config, small_label_number
Qwen3 7 Plus recon 0.5 5 tiny_code, dense_config, small_label_number, subscript_number
Llama 3.2 11B recon 0.5 6 tiny_code, which_color, dense_config, small_label_number
Claude Sonnet 4.6 recon 0.6 6 tiny_code, which_color, dense_config, small_label_number
Claude Fable 5 recon 0.6 7 count_circles, tiny_code, which_color, legend_lookup
Claude Sonnet 5 recon 0.65 7 count_circles, tiny_code, bar_value, dense_config
GPT-5.4 mini recon 0.7 7 tiny_code, legend_lookup, dense_config, small_label_number
Gemini 3.5 Flash recon 0.75 6 count_circles, tiny_code, dense_config, small_label_number
GPT-4o mini recon 0.75 3 tiny_code, dense_config, small_label_number
Gemini 3 1 Flash Lite recon 0.8 4 tiny_code, dense_config, small_label_number, subscript_number

4. The witness bake-off

On the render corpus the witnesses were raced for safety (band-ratio geomean; 1.0 = never over-certifies): RMSE over-certifies its blind spots about five times as often as the gradient-energy witness, which is why recon is primary and RMSE is kept for coarse images only. The full table, per item and per model, is in the calibration reports (the public corpus report).

5. How a band is replayed

A record keeps the two images’ hashes and dimensions, the witness, its score, the model’s band and the verdict, never the pixels. Anyone holding the two images can run the witness again and compare; anyone at all can check the seal and the signature at /v1/verify/<record_sha256> (the certify walkthrough does both). The record this site’s walkthrough issued: f164152c1dd555641a29b7fb2a297fc54c3d805114b426c3f3ea40983fd7bc1a.