A week ago we wrote about giving our compression engine a feedback loop — a way to know when a compression hurt an answer, not just how many tokens it removed. This post is the follow-up: what that loop found in its first week, what we shipped because of it, and the honest before-and-after numbers.
Seven days ago, compression on Nyquest was good at exactly one thing: shrinking prose. Everything around that had gaps we could not see:
It caught and helped fix its own first bug. The quality loop flagged a cluster of degraded answers, attributed them to one specific cleanup rule, and a targeted experiment confirmed it: the rule was mangling file paths, which are exactly the kind of fact a user asks to get back verbatim. The category of protected content — code, numbers, currency, URLs, dates, and now paths — is preserved byte-identical, enforced by the engine and checked by automated tests every day. Finding-to-fix took one day, and the same experiment cleared a second suspected rule — the loop tells us what not to change, too.
Web-grounded requests now compress. The live-web context a grounded answer depends on is preserved exactly, while everything around it — and the boilerplate inside scraped pages — gets compressed. A third of traffic went from 0% savings to normal coverage, with citation accuracy verified by the same judge loop.
Conversation condensation got ~5× faster and starts earlier. We benchmarked condensation backends head-to-head on fact retention and latency, swapped to the winner, and used the speed budget to bring condensation to shorter conversations. The result is a savings curve that compounds with conversation length:
And on top of compression: provider-side prompt caching. Conversations resend their history every turn. We verified that on two of the four major providers we route to directly, that resent context is now served from provider caches at 50–90% lower cost — stacking with compression rather than replacing it.
| A week ago | Today | |
|---|---|---|
| Web-grounded traffic | 0% savings (bypassed) | compressed, citations verified |
| Fragile content (code, numbers, URLs, dates, paths) | occasionally corrupted | byte-identical, enforced + tested daily |
| Conversation condensation | large chats only, slow | from ~7 messages, ~5× faster |
| Long conversations | up to 75% | 75%, reached earlier in the session |
| Live-traffic average | unmeasured | 33% (30-day window) |
| Quality measurement | none | sampled shadow evals, multi-judge, per-rule attribution |
| Resent-context cost | full price | 50–90% off via provider caching (2 of 4 vendors) |
Every change above followed the same sequence: measure, then move. Two proposed changes were rejected outright because the benchmarks said no — including one that would have made condensation both slower and less accurate. The wins shipped because the same instruments said yes. Seven engine releases in seven days, each one gated by the test suite (now 109 automated production checks, run daily) and each one reversible.
Try Nyquest → No account needed to start — or bring your own API key.