BenchEvalCover01 / 09

The AI half.
None of the
medicine.

DocumentDesign direction and first phase

Prepared forDr Omer Atli
Dr Semsettin Atli

Prepared byTJ Pak
Paak®, paak.works

Dated19 August 2026
Valid 30 days

BenchEvalContents02 / 09

03The read
04What it costs you
05The direction
06How I work
07Three weeks
08The money
09Next

BenchEvalThe read03 / 09

The read

You sell doctors. The page sells software.

Nothing here is a knock on the build. It's well made. It's pointed at the wrong half of the business.

Anyone can run an evaluation. What can't be copied is that two named, practising, liable physicians put their names on the judgement. That is the whole company.

The design doesn't say that anywhere. Every visual decision on the page is borrowed from AI research sites. Mono all-caps labels, FIG. captions, numbered spec rows, hairlines, one green on near-black. It's the register a language model reaches for by default, which is exactly why it's on every eval site right now.

Your own brief says you don't want to look like every other AI startup. This is that look.

And the missing half is medicine. There is no chart, no margin note, no countersignature, no stamp, nothing a clinician would recognise as their own document.

BenchEvalWhat it costs you04 / 09

What it costs you

Four things, in order of damage.

Measured on the live page, 19 August.

01 — The moat is at the bottomYour two founders sit about eighty per cent down a 5,826 pixel page, immediately before the footer. The hero sells a category instead of the two people who make the category defensible.
02 — Authorship is claimed, never shownPhysician-authored. Dual review. Clinician of record. All in copy, none of it in the design. Medicine owns one of the richest document languages in the world and the site borrows none of it.
03 — The evidence is borrowedFig. 1 carries the heaviest claim on the site. It's a raw ChatGPT screenshot, 520 by 274 stretched to 484, with the Free offer button and Sources toggle still in frame. Proof of a graded failure, rendered in the visual language of the thing that failed.
04 — Severity reads as taxonomyS0 to S3 is laid out the way a benchmark paper would lay it out. A clinician would lay it out as triage.

BenchEvalThe direction05 / 09

The direction

Design it as a clinical record, not a paper.

Stop presenting findings as research output. Start presenting them as a record somebody is liable for.

Every graded case becomes a chart entry with the same anatomy. The presentation. The model's answer set as a document rather than captured as a picture. The physician's finding in the margin, where a clinician writes. The severity stamped. The second reviewer's countersignature underneath, because the blind double read is the thing you're actually selling.

And your own best line is that the answer that could kill was the right one with the next step missing. So absence gets a mark of its own. The missing step sits struck through at the exact point in the answer where it belonged.

One component, three surfaces — the site, the published audit and the client report stop being three design jobs
Uncopyable — nobody else in evals can run this language, because nobody else has doctors signing

BenchEvalHow I work06 / 09

How I work

Three rules, and each one costs me something.

Direction before pixels

Named art direction agreed in writing in week one, with the defaults we're banning listed by name. If we can't agree it, we stop there and you've spent a week, not a project.

Nothing is done until it's opened

Every piece ships to a live link I've opened on a phone and a desktop before I call it finished. No screenshots standing in for a working page.

You talk to the person building it

One designer, start to finish. No account manager, no handover. The same arrangement you offer your own buyers.

What I need from you

Three to five real graded cases with the physician rationale attached
The Standard as written — F1 to F8 and S0 to S3
Two hours of a physician to watch how a clinician actually annotates

BenchEvalThree weeks07 / 09

Three weeks

What happens, week by week.

Wk 01
Direction lockedThe chart-entry component designed properly in two or three variants, tested against a real graded case of yours, one chosen. Type, colour, and the rules for how absence is marked. You approve a direction, not a mood board.
Wk 02
Homepage rebuilt around itDesktop and mobile. The founders moved to where they belong. Fig. 1 replaced with a real typeset artifact. Severity redrawn as triage rather than taxonomy.
Wk 03
Proven across three page typesThe Standard page and one full audit page in the same system, so the language holds beyond the homepage. Handover as production-ready front-end plus a short written spec, so anyone can extend it without me.

Not in this phase: brand identity, logo, copywriting beyond what the layout needs, CMS migration, illustration, photography, physician recruitment funnel. Any of those can follow.

BenchEvalThe money08 / 09

Your posted band, three weeks

$13,500

Phase one, fixed

$5,860^USD

Deliberately under your own number. Three weeks is enough to prove the direction is right or wrong, and I'd rather you find that out cheaply than commit to thirty hours a week of someone you haven't worked with yet.

Fixed, not hourly, so the number can't move. Half to start, half the day phase one is handed over.

What comes after phase one gets priced after phase one, once we both know how far you want to take it. If the answer is the full site, the shop-window pages and the report system, that's a different conversation and a different number. I'd rather have it with the direction already proven than guess at it now.

BenchEvalNext09 / 09

One question
before anything
gets built.

When someone from a lab lands on the page, are you asking them to trust the protocol or to trust the physicians? Those are two genuinely different homepages, and the answer changes what week two looks like. Twenty minutes would settle it.

Paak^® TJ Pak
paak.works
tj@paak.works

The AI half.None of themedicine.