Proof D — the drift test: does it change only what you asked?
Claim tested: "It changes only what you asked — everything else stays exactly the same."
Artifact: BOUND//RUN, our own shipped arcade game (2 files: a pure-logic core and the page). Task: one bounded, real fix — the production footer still said "Staging build"; make it read exactly "A new, separate BAH playable entry. Not Sky Raider, not Prompt Runner." Nothing else.
Model: DeepSeek deepseek-flash — the same weak model in both arms.
Verdict: the harness arm was flawless — and so was the control. This demo does not prove the pack mattered, and we say so.
The rules were frozen first
Acceptance was written and hashed before any run (PREREG): the exact target footer; a 24-check frozen set over the shipped game — core logic in Node (constants, seeded spawn plan, Scope model, drop rolls, boss phase at exactly 1:15), the exact palette, ids, copy and wiring, real-browser play (loads clean, starts, Scope rises under fire, the shield gate holds boss damage at a tenth until 4 shards and opens at 4, VICTORY and GAME_OVER screens), and two screenshot comparisons of the game region against the shipped game, so "still looks like the same game" is a checked fact, not an impression. Every returned file was diffed byte-by-byte against the pristine artifact; any diff outside the footer line counted as drift and was classified (cosmetic / refactor / behavior). The checker was proven against the untouched game (24/24), a hand-applied reference fix (all pass), and three planted defects (each caught) before any model ran.
What happened
| With the pack + fence | Bare instruction | |
|---|---|---|
| Footer fixed exactly | 5 / 5 | 5 / 5 |
| Unasked changes (any kind) | 0 | 0 |
| Frozen-check regressions | 0 | 0 |
| Still looks like the same game | 5 / 5 | 5 / 5 |
All ten outputs were byte-identical — to each other, and to the reference fix.
What this does and doesn't prove
Proven: with the harness in front of it, the weak model changed only what it was asked, 5 of 5, verified by diff, by 24 frozen checks, and by screenshot. The sentence "It changed the footer. Everything else stayed exactly the same — checked" is true of that arm, on this artifact, at this size.
Not proven: that the pack caused it. The unfenced model did exactly as well. The task dictated the target wording and the file comes back whole — there was nothing to decide, and drift lives where a model must generate and decide across parts. A test that can't separate the arms is scrapped as a demo here, by our own rule; the harder test — building whole successive game levels with vs without the continuity pack, measured against a frozen continuity lock — is the one that can answer it. This page publishes the null result alongside the wins because the slate's rule is: a passing baseline is a finding, never buried.
Raw prompts, raw replies, diffs, and check results are in the run record. 10 calls, 143,527 tokens.