compression with a quality contract

The Benchmark

Every compression method shrinks tokens. The only number that matters is how much it saves without changing what the agent decides — and what it truly costs once prompt caching is priced in. This is a head-to-head on exactly that axis, where every technique runs through the same decision-equivalence gate and the same cache-aware cost model. The winner is computed, not assumed.

The result, in one line. On the bundled 7-domain corpus, Distil leads on certified savings — 37.4% cheaper at 100% decision-equivalence. Lossy methods reach higher raw token cuts, but drop decisions and are disqualified by the gate. Raw savings that change the answer aren't savings.

The standings

Ranked by cache-aware dollar savings. Certified means the method passed a statistical non-inferiority test and preserved 100% of decisions; anything less is disqualified, however much it saved.

TechniqueFamilyTokens saved$ savedDecision equiv.VerdictFidelity
distil-causalcache-aware + causal pruning 33.8%37.4% 100%✔ certified · leaderlossy*
truncate-tailsliding-window / truncation 26.9%33.0% 25%✘ fails gatelossy
summarizeabstractive / rolling summary 19.7%24.2% 0%✘ fails gatelossy
distil-losslesscache-aware lossless 8.4%10.3% 100%✔ certifiedbyte-exact
extractive-pruneextractive importance (LLMLingua family) 7.4%9.0% 68%✘ fails gatelossy
minify-alllossless minification 0.0%0.0% 100%✔ certifiedbyte-exact
baselinecontrol (no compression) 0.0%0.0% 100%✔ certifiedbyte-exact

*distil-causal drops context that ablation proves never changed a decision — so it isn't byte-reversible, but it is certified decision-equivalent. distil-lossless is the byte-exact operating point. Numbers: claude-opus-4-8 pricing, deterministic runner, bundled corpus — reproduce them yourself below.

Why this is the honest comparison

Same gate for everyone

No method is special-cased

Every technique — including Distil's own — is scored by the identical decision-equivalence + non-inferiority gate and the identical cache-aware cost model. The harness will happily rank a competitor above Distil if it earns it. It doesn't, because no other family combines lossless, cache-stability, causal pruning, and certification — but the door is open.

Faithful baselines

Best-form, not strawmen

The baselines are faithful reference implementations of the real technique families — sliding-window truncation, extractive importance pruning (the LLMLingua / Selective-Context lineage), abstractive summarization, naive minification — each in its best reasonable form. They genuinely remove tokens. They just can't prove they kept the decision.

Raw ≠ real

Two ways to "win" that don't count

A method can post a big raw token cut yet flip decisions (truncate-tail: 27% saved, only 25% of decisions preserved). Or it can shave tokens yet bust the prompt cache and cost more in real dollars. The benchmark prices both, so neither illusion survives.

Verify it yourself

Plug in a real tool

Don't trust our reference baselines? Register any installed compressor through the external seam and it's measured on the identical axes. The claim "Distil leads on certified savings" is reproducible — and falsifiable. That's the point.

Reproduce it

# the full standings, offline, zero API key
distil benchmark

# render the comparison as a shareable page
distil benchmark --html standings.html

# verify against a REAL external compressor (list[str] -> list[str] over block texts)
distil benchmark --external mypkg.compressor:compress:MyTool

# grade with the live model instead of the deterministic runner
distil benchmark --runner anthropic --tokenizer anthropic
Honest scope. These numbers are on Distil's bundled 7-domain corpus with the deterministic (structural) decision-equivalence runner — reproducible by anyone, no key required. For task-accuracy on a public benchmark (τ-bench, SWE-bench, GSM8K), ingest its traces into a corpus and run with --runner anthropic: the same comparison, graded by the live model. The harness is the deliverable; the corpus is swappable.