← Blog

Why a Certificate, Not a Compressor

Cody Lee Walker · codylee.ca · 2026-07

There is a line item in every vision-LLM bill that nobody itemizes: the images. Text you can cache, trim, and diff. Images hit the meter every single call, and they are billed in a currency almost nobody optimizes — not bytes, not quality, but input dimensions, converted to tokens by each provider's own arithmetic.

That mismatch is the whole story. Everything the image-optimization industry ships — quality sliders, smart formats, perceptual codecs — optimizes bytes. We measured what a quality knob does to the token bill: requantizing to JPEG quality 25 saves exactly 0.0%, because the dimensions didn't change. The only knob that pays is downscaling.

So: downscale. But how far?

The aggregate trap

Every existing answer to "how far" is an aggregate. "Our method retains 95% benchmark accuracy on average." An average is not a guarantee about your image, compressed this far, read by your model. When the average is wrong, you find out in production — a receipt total misread, a chart mislabeled, a dashboard number hallucinated — and nothing in the pipeline even flags that it happened.

It gets worse: the obvious per-image check is a pixel-error metric, and pixel error is confidently wrong in exactly the cases that matter. On our held-out corpus a naive RMSE policy silently loses 42–50% of answers — while reporting the biggest savings number on the board. The tool that looks like a safety check is a liability that compounds with scale.

What an actual answer looks like

starlens treats "does the model still read it the same?" as a measurement, not a hope. Every request runs the same primitive:

drift = D(witness(X), witness(T(X))) ≤ tolerance

A witness is an instrument that attests what an image still says to a reader — an OCR pass for text, gradient energy for renders, a composite for general content. Each witness has a measured corruption profile: we know what it misses, on which content classes, for which models, because we ran the bake-off (12,123 sealed model calls, content-addressed, every one replayable for free). The tolerance is calibrated per model, because models don't degrade together — a band that's safe for one reader loses another 29% of its answers.

The output is a certificate: the witness that stood behind the verdict, the measured drift, the tolerance it cleared, the model it was calibrated for, and a record_sha256 seal anyone can re-verify at /v1/verify/<hash> — no account, no trust. And when the measurement says no safe compression exists, you get a certified refusal: the original image untouched, plus the certificate saying why, for free.

On the 175-item held-out split — never used to fit anything — the calibrated policy retains 96–97% accuracy while cutting 12–22% of the billed tokens. Not a demo number: an out-of-sample operating point, with the refusal rate published next to it.

Why this is worth money when downscaling is free

Because the compression was never the product. You can downscale for free today; what you can't do for free is know, per image, that the model's answer survived — and prove it later when someone asks why the agent misread the invoice.

The pricing follows the physics. starlens charges a 20% share of the savings the certificate proves, at your target model's own input rates. Saved nothing? Certified refusal, $0. Saved 1,020 tokens on a screenshot (our own quickstart does — 2560×1440 → 827×465, 67% off, seal 56dccb31…)? You pay a fifth of what those tokens cost and keep the rest. The certificate is literally the price tag: every charge traces to a sealed, replayable measurement of value delivered. There is no seat, no subscription, no bet on your usage — if the certificates stop proving savings, the bill stops too.

That's the difference between infrastructure and a filter. A compressor gives you smaller files. A certificate gives you a decision you can defend: this image, this far, this model — measured, sealed, priced by what it saved you.

Don't trust us — re-verify the seal.