Harness / v1.0.0
Eight ways
to disagree.
The skill is a procedure. The harness is the part that can say no. It renders the page in headless Chrome, measures the frame, audits the scene graph, drives the loop, inspects the shell and returns one verdict.
01The loop
See. Measure.
Change one thing.
The harness exists because an agent that cannot see its own output cannot art-direct. The renderer gives it eyes; the metrics give it a direction; the gates stop it from lying to itself.
01 / RENDER
Deterministic frame.
Headless Chrome, a fixed camera, a fixed seed and a fixed viewport. A small CDP client with no npm dependency. The page sets __READY after a few frames.
02 / MEASURE
Multi-scale numbers.
Exposure, clipping, palette, edge density, colourfulness, band energy at fine/mid/coarse scales and spatial p10 over an 8×5 grid. Global means hide flat regions.
image-stats.py · score.py03 / DIAGNOSE
Three ranked fixes.
The lowest ratio is mapped to a specific change. One row, one change. The model does not have to decide what “looks worse” means.
diagnose.py04 / VERIFY
Gates, not opinions.
Shaders, census, collision, lights, playtest, product, mobile and look. The look gate is last because a beautiful frame does not rescue a missing game.
verify.mjsrender the fixed cameras score each; keep the worst diagnose the worst; take the top item change only that item re-render; keep it only if the worst camera improved run every gate; revert on any regression stop at the budget or two rounds without a gain
02The instruments
Small tools.
One verdict.
Each tool does one job and prints something a model can act on. The acceptance command composes them; you can also run any of them alone.
| Tool | What it does | Output |
|---|---|---|
render.mjs | serve the page, launch headless Chrome, wait for __READY, screenshot, collect console errors and page stats | PNG + JSON |
image-stats.py | exposure, clipping, midtones, palette, edge density, multi-scale band energy, spatial tiles | JSON |
score.py | ratio a frame against a profile; add binary tone checks | ENERGY + tone |
calibrate.py | median a profile from 2–3 target frames | refs/profile.json |
diagnose.py | rank the three lowest metrics and map each to a fix | ranked list |
objectbox.py | measure a crop: block detail, specular peaks, contact proxy | per-box rows |
census.mjs | enumerate every rendered instance; check segments, maps, identity | PASS / failures |
collide.mjs | collider parity, penetration, bounds, bare edges, walk-through | PASS / lists |
lightaudit.mjs | every light has a fixture within 0.4 m | PASS / distances |
playtest.mjs | drive selftest or a declarative plan: win, loss, cause, restart | PASS / report |
product.mjs | shell, settings, save, touch, content breadth, HUD, audio, budgets, cameras | PASS / findings |
mobile.mjs | 390×844 layout, overflow, title reachability, touch | PASS / findings |
verify.mjs | run all eight, print one verdict, exit non-zero on failure | PASS/FAIL |
qa.mjs | fixed-camera table from __PRODUCT.cameras | table + JSON |
structcheck.mjs | duplicate top-level symbols and repeated blocks | report |
03Measured
Black Ice,
eight green.
The stock harness on the shipped build, with no project profile: 8/8 in about 66 seconds. The look gate ran in fallback mode and said so.
shadersNo compile/link errorPASScensus167 meshes, 0 failuresPASScollide102 colliders, 0 penetrations, 0 walk-throughPASSlights12 lights, all anchoredPASSplaytest3/3 progress, named cause, restartPASSproductShell, content, save, touch, budgetsPASSmobile390×844, touch, no overflowPASSlookTone + emptiness fallbackPASS$ node harness/verify.mjs \
--dir ../black-ice --page index.html
shaders PASS
census PASS
collide PASS
lights PASS
playtest PASS
product PASS
mobile PASS
look PASS
look mode: FALLBACK
VERDICT: PASS (8/8)
real 65.76sMeasured 2026-11-14. The project's own gate reports 163 draw calls, 127k triangles and 12 lights on the title frame; the numbers move during play.
04Design rules
A gate that cannot fail
is not a gate.
The harness is intentionally small, and intentionally harsh. These are the rules that keep it honest.
01
No project facts in the gate.
No coordinates, mesh names, counters or item ids. Those live in the page's __TEST plan or selftest. A gate with a hard-coded coordinate only works on the run that produced it.
02
The gate wins over the model.
If the gate disagrees with the model's opinion, the gate wins. If the gate is wrong, fix it with a counter-example, never bypass it. A bypass is a future silent regression.
rule governance03
Failures are ordered by class.
Behaviour before craft before visual before polish. A collision bug invalidates a beautiful frame; a missing settings screen invalidates a high look score. The order is the priority.
failure class04
The look gate is a compass.
It measures similarity to one reference, not finished-ness. A stylized game can score low by construction. Report the worst camera, the tone checks and the mode; never let the score outrank the product gate.
calibration05What it does not do
No engine.
No taste.
No promises.
The harness is a measuring instrument, not a creator. It will not make a bad subject good, choose an art direction, or replace the human eye at the end.
NOT AN ENGINE
It drives your page.
There is no scene format, no renderer and no asset pipeline. The page is the deliverable; the harness only inspects it. Engine builds can be adapted through the same contract.
NOT A TASTE MODEL
The eye veto still exists.
A noise layer can raise a metric and look like television static. If a change raises the number and looks worse, the number is wrong. A human or a visual check can veto a frame.
NOT A GUARANTEE
Green means the floor held.
The gates check what they can see. They do not prove the game is fun, the writing is good, or the art direction is worth the effort. They prove the floor did not collapse.