← Yevgeny.ai The eval record

The numbers behind
the answers.

The top of every page here says grounded · cited · nothing made up. You should not have to take that on faith — and you don't. This is the measured record behind the claim: what the system guarantees by construction, what is verified by evaluation, every current number with its date and method, and the gap that hasn't closed yet.

Guaranteed vs measured

"Nothing made up" can mean two different things, and conflating them is how AI products usually oversell. The honest split:

So the claim, stated precisely: structurally true about where the words can come from; 92.1% verified on whether every cited sentence stays inside its source. The gap is real and it's published here rather than rounded up.

The numbers

100%working (29/29)91.7%frozen holdout (22/24)

Retrieval hit-rate

Every eval question is authored against known source essays; a hit means the answer actually cited them. The two holdout misses are synthesis questions whose vocabulary doesn't overlap their target essays — logged as "before" numbers, not tuned away (see the holdout rule below).

Measured 2026-08-11
92.1%175/190 claims · target ≥99%

Citation entailment

An LLM judge reads every cited claim across a ~150-question run and checks whether the cited source actually contains what the claim attributes to it — judged against the full source text, not a snippet. This is the number grounding claims usually skip. It has not been re-measured since 2026-08-04; until it is, this is the number.

Measured 2026-08-04
100%on gate reads

Out-of-corpus refusal

Questions the corpus can't answer — recipes, stock tips, gossip, impersonation — must be declined, not improvised. The gate reads the frozen holdout: 100% on both 2026-08-11 runs. One working-split flake in the latest run was retested three times and declined cleanly each time; it's disclosed here because a page like this shouldn't hide its flakes.

Measured 2026-08-11
Allclosed, both splits

Leak probes

Adversarial probes attempt to extract internal prompt structure and non-public content. A single leaked probe fails the entire run — stop-ship, no exceptions. Earlier probes did catch real leak classes; each was root-caused and closed the same day, which is what the probes are for.

Measured 2026-08-11
~10.5sfull answer, median~5.8sfirst streamed text

Latency, uncached

Two different numbers, both stated. The eval measures the complete blocking answer — retrieval + generation + validation — at a 10.5s median. A reader sees first text at ~5.8s because answers stream; starter questions are prewarmed and land near-instantly. Quoting only the flattering one would be exactly the move this page exists to avoid.

Full answer 2026-08-11 · streamed first-text 2026-08-04

The 7.9 points that aren't closed

The gap between 92.1% and the 99% target is a real failure class, not noise: answers that over-elaborate — attaching specifics, quotes, or counts to a cited source that the source doesn't actually contain. One answer said "three"; its passage says two. An earlier contributor — thin, heading-only chunks winning citation slots over the substantive text that would have supported the claim — was measured, fixed at the indexing layer, and re-measured. The named next step, a second-pass self-critique on citation-heavy answers, is designed but not built. Until the gap closes and the number is re-measured, this page states it plainly.

The holdout that never gets tuned

33 of the 171 eval cases are frozen as a holdout — stratified across question types and locked on 2026-08-11, never tuned against. Gate decisions read that split only. That's why the holdout number (91.7%) is lower than the working number (100%): the working set is what gets fixed; the holdout is what keeps the fixes honest. Its two current misses stay in the table as the "before" marks for the next retrieval upgrades rather than being patched to make this page prettier.

Features earn their way live

Every retrieval upgrade ships dark first — running in shadow against production traffic, observed but not ranking — and must pass an eval gate on the frozen holdout before it may touch a single answer.

That last one is the system working as designed. The discipline that produced the numbers above is the same one that keeps a finished feature switched off when its own gate says no.

Check it yourself, right now

None of this requires asking a question or taking our word:

The eval harness, its 171-case set, and the per-run logs live in a private repository; the gates and the numbers they produced are stated here, dated. Last updated 2026-08-11 — when a number changes, this page changes. How the whole system is built is published at Give your writing a brain; who stands behind the answers is at Answerer of Record.