Boundary Map

Every experiment on this site ends with one question: whose failure was that? This page is the running answer. It is the ledger the whole program exists to build.

Why attribute at all

"The AI could not do it" is almost never a finding. It is a place where three very different causes have been left unseparated, and the difference decides what you do next.

Class	What it means	What you do about it
Model-bound	The capability is not there. No prompt, contract, test or process recovers it.	Wait for a better model, or change the goal.
Harness-liftable	It fails under a vague ask and passes once the environment states the contract, exposes the state, and tests it.	Fix the environment. The model was never the problem.
Process	The work itself was managed badly: the wrong thing was asked for, or the wrong thing was accepted.	Fix the method. This one is about the manager.

Four rungs in, the honest headline is this: the model-bound column remains empty. The observed failures so far have either changed under a new contract or remained outside the scope of the available gates.

The ledger

Rung	Complexity jump	What the harness lifted	What stayed below the line	Attribution
01 Pong null control	None. Two paddles, a ball, a score. The floor of the ladder, where the task has nowhere to hide.	The scope lottery. One vague sentence produced three different games; an explicit spec collapsed the lottery. Acceptance tests did not appear until the spec demanded them.	Nothing about how the game felt to play. And the deterministic falsifier itself failed two correct builds before it was hardened.	Harness-liftable, plus process. At this altitude the failures were specification and management failures, not model capability.
02 Breakout state coupling	A persistent brick field. One brick now has to stay coherent across rendering, collision, score, the remaining count, the win condition, and replay.	Mechanical verifiability, completely. All three engineered builds exposed the test API and passed the driver: deterministic replay, paddle bounce, brick collision, final-brick victory, repeated-miss loss.	The player. None of the three passing builds offered a way to restart after game over. Prompts stayed stuck on screen, difficulty drifted, and the build a human actually preferred was an uninstrumented overlay the driver could not test at all.	Verifiability: harness-liftable. Human-facing quality: unattributed. It did not rise under the same contract that lifted correctness, and that is not the same as proving it cannot.
03 Space Invaders active opposition	The passive target field becomes a synchronized enemy formation with hostile fire, destructible shields, speed-up, terminal states, and waves.	One primitive Haiku artifact missed the required mechanical coupling. That miss did not recur in a new Haiku artifact after the prompt, API contract, and driver conditions changed together; causal contribution remains unseparated. The final driver passed all four engineered cells.	The Pro graft exposed a shield-probe false negative and required multi-lane driver hardening. Current gates do not judge pressure, fairness, atmosphere, mechanic magnitude, or replay desire; those findings came from one operator's human-terminal review.	Observed outcome: the four engineered artifacts passed the named mechanical checks, while UX and mechanic magnitude received separate process and human verdicts. No model-inherent boundary was established.
04 Pac-Man independent agents Wave A only	The single formation splits into four enemies that navigate a maze graph and decide at junctions on their own. Watching one no longer tells you what the others do.	Observability, not autonomy. An explicit contract added deterministic stepping, scenario placement and a shared keyboard/API path to the chosen seed without changing how the game plays.	Distinctness. The black-box driver reached at most A2: three ghosts produced identical signatures across all 24 scenarios and the fourth differed in 5 of 24, so the four-policy gate failed. Source declares four target rules in three of four artifacts; no blind reviewer felt them.	Practical completeness fell from 3/4 to 1/4 while the generic floor stayed at 3/4 - the first sharp drop on the ladder. The false FAIL from junction geometry is harness-bound; the gap between declared rules and felt character is unattributed. Three Claude cells are deferred, so no model-bound claim is available here either.

Rung 01 is the Pong experiment; its autopsies are Three Pongs From One Line, My Lie Detector Lied and Fun Is Not a Spec. Rung 02 is the Breakout experiment, read through All Three Breakouts Passed. None Could Restart. and The Untestable Game Won. Rung 03 is the Space Invaders experiment, where prompt-floor, mechanical, and human-terminal verdicts stay separate. Rung 04 is the Pac-Man experiment, the first rung where practical completeness fell sharply while the technical floor did not move.

What this map does not establish

A map that only shows the discovered territory is a sales brochure. The blank areas are part of the finding.

This is not a model leaderboard. One artifact per cell. Model tier did not predict fidelity to the source, and that observation is not a ranking.
Fun has never been lifted by a contract. It has also never been shown to be unliftable. It sits unattributed, which is the most interesting cell on the page.
The ladder is four rungs tall, and the fourth is half-run. Every claim here is still a claim near the floor. Pac-Man adds independent agents, but three of its seven cells are deferred, so its own matrix is not closed.
The falsifier is part of the system under test. It lied twice at rung 01. A verdict from an uninspected verifier is not evidence.

The map grows one row per experiment. When a failure finally survives an explicit contract, a hardened falsifier and an honest process, it will be the first real entry in the model-bound column, and it will be worth the whole climb.