Cases / evidence

Three runs.
Two verdicts.
One reference.

Same model, same single-file constraint, different harness. These are not controlled model benchmarks; they are controlled harness comparisons. The question is always: what did the harness change, and what did it catch?

DeepSeek V4.1 Flash 6 runs · 2 genres Evidence from session logs, not README claims

01The evidence table

Every run,
including the weak ones.

A case study without the failures is marketing. Ironsight is here even though it was not published: it is the run that produced the product gate.

RunGenreHarnessSizeGatesVerdict
Sundown CircuitCity kartnone890 KB—Strong core, thinner shell
Sunset DistrictCity kartskill + gates1.58 MB8/8Use it
Neon BreachShooternone304 KB—Strong content, stylized
IronsightShooterskill v11.50 MB6/6Not enough
Black IceShooterskill v2318 KB8/8Use it, with a budget
SaltlineShooter—1.42 MB—Reference

Gates are the project's declared acceptance command at the time of the run.Not a controlled model benchmark

02How to read a run

What a controlled run
does not prove.

The model is the same across the A/B runs. The harness is the variable. That isolates the harness, not the model. It does not prove that the skill is better than a human, or that 318 KB is better than 1.4 MB.

IT PROVES

The gates can change a build.

The product gate moved the second shooter from a frame with one enemy to a game with four roles, a boss, settings and a save. The run changed because the acceptance command changed.

harness effect

IT DOES NOT PROVE

That the model has taste.

The skill delegates judgement to checks. It still needs a human eye for the art direction, the subject and the final veto. A green gate is a floor, not a ceiling.

human veto

IT NEEDS

The same reference, or none.

A profile captured from another subject ranks style, not quality. Without a project profile the look gate explicitly falls back to tone and emptiness.

calibration

Build one

Bring your own subject.

A container yard, a workshop, a stairwell, a market. The grammar is constant; the mechanics are not. If you build one, open a case study with the failures left in.