the player's wiki. Everything worth remembering, in one file that works offline forever.

Right now

Latest build — Castle Siege, 2026-08-01 to 08-02. Two repos: V1 is the playable castle builder with wave defense; V2 forked from it, passed a mechanics bar in three rounds and spent six more failing to reach the reference image. Five new method articles came out of it, below.
Live — Darwin's Island II at games.antlab.cloud/darwin2. Grid-based, click-to-move, 55 creatures, four rivals, a map editor. What the gauntlet build actually taught.
Live — Darwin's Island 3D at games.antlab.cloud/darwin3d. V3's simulation with the camera dropped to ground level. Six rounds of measurements, and an honest account of why it stopped.
Tomorrow's build — Darwin's Island, the Warcraft III survival map, rebuilt in Three.js. Architecture, build order, and the five things to decide first.
The question being answered — where Three.js actually breaks, and how to go through each wall.
Your own engine — ChronoBlock is hardened, reusable, already paid for. Copy it; never regenerate it.

Pushing 3D in a browser

Where Three.js actually breaks — the walls in the order you hit them
Draw calls are the budget — the single most important idea
A million grass blades in one draw call — the technique, step by step
The painterly world demos — real parameters from two public codebases
Adding lights flattens a scene; one key separates it — a four-source rig built to create midtones scored worse than the two-source rig it replaced
Shader compilation is the hitch you can't find — the 1.2-second frame in a game reporting 94fps
The mechanism defending your frame rate is your worst hitch — the quality autoscaler cost 446.6ms per action and defended a bar under the spec
Generating textures, meshes and sound from code — shipping with zero art assets
Single-file HTML vs a real build — measured on your own projects
When an effect supplies 18% of your key — how to tell a highlight bloom from an accidental exposure control, and why the fix is never one slider
Where you reset the counter decides what you measured — an AO prepass re-rendered the whole scene and never showed up in a single reported number

Method — how to actually finish a big build

Fan out for content, single-thread the code — two projects reached this separately; one disproved its own founding prompt
Verify with pixels, not assertions — turning taste into an exit code
Determinism is a prerequisite — why a seeded RNG is load-bearing
Contracts before creativity — fix the vocabulary before anything generates against it
Content is data selecting from a vocabulary — 12 creatures became 55 in one round by deleting 12 functions
Agents cannot judge feel — three rounds of critics missed unplayable controls; build mechanisms they can test
Scope an agent to every file it must touch — a withheld file makes it route around you, and the workaround ships
Review was the bottleneck, not building — 112 minutes a round, almost all of it an agent watching a clock
Your measuring instrument rots too — two rounds lost to a wrong statistic and a stale constant
An instrument can certify a claim it never measured — a test reported "submerged 67% of the player" about a quarter-unit puddle with his head in the air; it had never looked at the water surface
Trust the working tree, not the agent's report — two lanes returned ok:true for files that did not exist, and a third narrated another lane's edits as its own
Three consecutive collision fixes each relocated the defect — roof teleports, then 29% of falls through solid stone, then a 1.6u sideways shove; sweep a parameter, don't spot-check a value
A comparison bar is not a checklist — "judged blind against this image" let six rounds of polish raise the score while the castle stayed a square ring; count the nouns in the exemplar first
Measurement plateaus before taste does — the geometry reached 0.62% against the reference's 0.59% and the frame still read as a greybox; three rounds netting zero is the stopping signal
Calibrate the instrument against a known load before you measure anything — the two controls that disqualify a clock in minutes; one of them found a "60Hz" source that got faster as the work went up
A test that configures the feature has not tested the product — five checks that passed on a build that did not have the feature, and the one technique that found them
Mask the subject, or you are measuring the background — five correct ablations all read as null because the animal was averaged with the ground
Name the renderer or the fps is noise — 1.3 and 157 on the same build, same minute
Spawn to the deficit, not the schedule — a correct spawn mix still inverted the living world
One predicate cannot answer three questions — one toggle flipped, 40% of the animals ended up in the ocean
A boolean cannot express a progression — solid: true made a knee-high stone and a 3.4-cell pine the same wall; 30% of the world's blocked cells were things you should have stepped over. The sibling defect to the one above, and a different fix
Most "missing features" are unbound keys and unflipped defaults — grep the request's nouns before you estimate it; three of five complaints were already-written code
A player who cannot name the mechanism still points at the right file — four reports landed on a boolean, a camera feedback loop and a Poisson process; the fifth was true and blamed an innocent subsystem
Write a differentiation contract — name what the next project is forbidden from reusing
The frozen-build rule — a shipped build is never edited again, ever
Give the game a test surface on purpose — and write down the traps
A shared namespace needs one owner and a priority table — one key stole combat in three consecutive builds and no single file was wrong
Every fix ships with a switch that brings the bug back — a check you have never seen go red is decoration; three arms that prove otherwise
A distance threshold between centres gets tighter as the bodies grow — a 1.19-cell reach against bodies summing 2.87; a whole round spent on two hypotheses that were both wrong about an arithmetic error
A feature can be correct, tunable, and read by nothing — eight dead settings and three whole mechanics that reached no pixel; a test of the feature passes, so ask what reads the field instead
Three instruments went red and not one of them had found a bug — a committed design change timed a turning circle and made three correct test files wrong; derive every window from the live setting instead
A verifier that cannot return "partly" is a rubber stamp — four true measurements on a fix that did not work, and the two zero readings that meant opposite things
Adjacency is what makes a fence, so damp the neighbour and nothing else — trees walled the player in because independent rolls are a Poisson process; the smallest spacing radius beat every wider one on both metrics and kept the most forest

Case study

Claude of Duty — 65k lines from a six-sentence prompt, and an unusually honest post-mortem
Darwin's Island 3D — six rounds, round by round with the numbers, and the ending: the substrate decomposition was right, the camera came down in round one, and every commit after round 5 went to how the animal moves
Castle Siege — 2026-08-01/02. Mechanics bar passed in three rounds and held; visual bar never cleared in six more. 12.2M subagent tokens, and an honest split between the part that worked and the part that didn't

How this wiki works

Articles are markdown in articles/. Links like [[some-slug]] resolve by filename and display the target's real title. A red link means the article doesn't exist yet — that's the to-do list, not an error.

See How this wiki works for the rules.