Compress the image. Prove the model still reads it.
Vision models bill by dimensions, not bytes. certify shrinks an image and seals a per-model record that the answer survived, or a refusal for free. The lab replays the real calibrations; the sections after it are how the record is made and checked. The tools themselves are on the front page.
Drag the compression. Watch each model break.
The measurements.
- JPEG quality 25 saves 0.0% — providers bill dimensions, not bytes. The billing math →
- Naive pixel-error silently loses 42–50% of answers on held-out content — while reporting the best savings number. The RMSE trap →
- A band that holds 100% on Claude drops GPT to 71% — transfer is one-directional. Name the model →
- 12,123 sealed VLM calls behind the measurements; 0.96–0.97 answer retention at 12–22% savings on a 175-item held-out split. Read the essay →
Don't trust the compressor — re-run the seed.
Every response carries a sealed record: the exact bytes in and out, the witness that measured them, the drift it saw, the tolerance it held, and the model the guarantee names. The seal is a hash of all of it. Anyone can re-verify — free, no account, deterministic down to the seed.
# re-check any seal, yours or ours curl https://api.s2ar.dev/v1/verify/<record_sha256> # → re-runs the witness, recomputes the hash, returns the verdict
- source / compressed — bytes, dimensions, sha256 of each
- tokens — before / after / saved, at your provider's billing shape
- fidelity — witness, drift, tolerance, verdict: isomorphic | moved
- target — the model this guarantee names (transfer is one-directional)
- seed + harness — everything needed to replay the measurement
- record_sha256 — the seal. GET /v1/verify/… re-checks it
Pixel-error is blind to what the model reads.
Plain RMSE calls a 640×360 render "identical" at 32×18 — because the image is mostly smooth. But the signal a model needs is high-frequency: small text, 1-cell texture, thin edges. starlens compresses against a witness of the reader's content — RMSE for coarse images, OCR for text, gradient-energy for render detail — so the certified band ends where the answer would move, not where the average pixel does. Every band is measured per model and sealed per request; re-verify it from the hash.
- Measured, not asserted. Calibrated on claude-sonnet-4.6, gpt-4o-mini and llama-3.2-11b across text, UI, render and game corpora — where each model's answer actually breaks.
- RMSE over-certifies its blind spots ~5×. The gradient-energy witness never does — that gap is the whole reason the witness exists.
- The band is model-relative. A frontier reader isn't always the most robust — on some images the 11B open model reads deeper. The certificate names its target.
- The witnesses are cheap. Millisecond, deterministic, re-runnable — the guarantee costs almost nothing to make or to check.
Token compression research needs the model. This works on yours.
Prune, merge, resample — inside the model
Visual-token pruning, merging and resamplers reach 4–16× fewer tokens in the lab — by changing the vision stack itself. Powerful, and out of reach unless you own the weights and the serving path. Against a closed API, none of it applies.
Compress the input, certify the answer
starlens never touches the model. It shrinks what you send — to Anthropic, OpenAI, Google, anyone's billing meter — and proves per request, per model, that the answer held. Model-agnostic, deployable today, honest when it can't help.
The technique adapts — by content, and by model.
Two dials tune every certificate. The witness is chosen by content — RMSE for coarse images, OCR for text, gradient-energy for render detail (the bake-off above). The tolerance is measured per reader: how far recon's band must tighten so it never outruns that model. Below are the models we've measured — strictest tolerance first; the chips are where recon over-certifies each one, so the certificate pulls back.
It also speaks OpenAI. Point a base URL at it.
The compressor is also a model on an OpenAI-compatible endpoint, not a proxy into your conversation — it sees only the image. Call it, get a smaller image plus a sealed certificate back, then send the small image to your real model every turn. Or self-host the same audited engine.
# one call compresses; every later turn is certified-cheaper client = OpenAI(base_url="https://api.s2ar.dev/v1") r = client.chat.completions.create( model="stardata/starlens", messages=[{"role":"user","content":[ {"type":"image_url","image_url":{"url":data_url}}, {"type":"text","text":"target:claude-sonnet-4-6 probe:auto"}]}]) # → r.choices[0].message.images[0] = the compressed image # → r.choices[0].message.content = the sealed certificate (verify by hash)
utility model
Sits on any OpenAI-compatible base URL. Never holds your key or sees your prompt — only the image. Not in your critical path: if it's down you lose the savings, not availability.
compressing meta-model
Call starlens/claude-sonnet-4-6; it compresses, forwards, returns the real answer, and bills on the tokens it saved. Same certificate, no workflow change.
Cents per call. Nothing for a call that fails.
assert: $0.005. inspect: $0.005. certify: $0.01. score: $0.005. sym: a repo map $0.01; a file's skeleton, one function's source, a search by name $0.005; a search by meaning $0.01. Three keyless calls a day are free on each route (one for certify); signing in with an email grants $0.50. No subscription, no per-seat fee, no minimum.
Pay three ways. Per call: USDC on Base over x402 — every unpaid request answers a 402 with the amount and the payee, sign it, resend; no key, no account. By card: a pack of $5 · $25 · $100; the key is shown once after checkout and goes in Authorization: Bearer. Through an agent wallet: Stripe's Machine Payments Protocol at POST /v1/credits/buy. One ledger for every tool.
Money moves only after the answer exists. Every paid answer carries a receipt (an evidence hash of the answer, the cost, the determinism of the endpoint); certify answers are signed Assay records. The issuer signs its own day — calls paid, delivered and credited — at the proceedings, under the charter: no placement for sale, no different verdict for money, nothing kept about a caller.
A deterministic world grades the models.
Our element simulation is seed-deterministic with an exact matter census — so every question about a frame has a machine-checkable answer, no human labels. We compressed the frames until the models broke: the certificate held 100% for its target model at the certified rung, the industry-standard RMSE witness said "isomorphic" all the way down to 32 pixels wide — where accuracy collapses to a coin flip. See the full ladder →
What we can save, by model.
Typical vision-token savings. Measured where we've calibrated; estimated from robustness tiers elsewhere — the list grows every calibration.