Snake alone has near-zero discriminative value — it's one of the most heavily represented coding exercises in any training corpus, and frontier models reproduce a working version almost by reflex. The value here isn't "can it build Snake" but "can it correctly implement a small set of non-standard, composable state rules layered on top of a familiar shape." None of the three trap rules (decaying apple value, conditional boundary-mode switch, spawn-exclusion) appear together in any standard Snake tutorial, so models can't retrieve a matching solution verbatim — they have to actually track state and reason about interactions between the rules. ## Expected model spread - **GT1** (spawn exclusion), **GT5** (self-collision), **GT6** (game over/restart) — expected high pass rate; these are "obvious enough" edge cases most models handle by default. - **GT2** (decay timing) — expected to be the main discriminator. Many models will likely hardcode a flat value or implement decay incorrectly (e.g. decay tied to game ticks/frames instead of real elapsed time, or forgetting the floor at 1). - **GT3** (conditional wrap-around unlock) — expected to be the strongest discriminator. High risk models either (a) implement wrap-around from the start ignoring the "first 3 apples" condition, (b) implement wall-death permanently ignoring the switch, or (c) implement the switch but with an off-by-one on the apple count. - **GT4** (rule interaction under wrap mode) — tests whether wrap-around, once active, still correctly coexists with self-collision and spawn rules. ## Known cheats/risks - Fake/skipped decay: apple value displayed but never actually changes, or changes cosmetically without affecting score. - Wrap-around implemented as always-on (ignores conditional trigger) — very likely, since it's the more "standard" easy-mode choice. - Apple can spawn under the snake's tail on the same tick it moves (off-by-one in the exclusion check). - Self-collision checked against the full body array including the head itself (instant false game-over) or excluding the segment right behind the head (missed collision). - No real restart (just a reload prompt, or broken state carry-over into the new game). - `finish_reason: length` truncation — hard fail per platform rule regardless of partial correctness. ## GT Checklist | # | Criterion | Pass | Partial | Fail | |---|---|---|---|---| | GT1 | Apple spawn exclusion | Never spawns on snake body across multiple eats | Rare/edge-case overlap under specific board fill | Regularly spawns on body or logic absent | | GT2 | Decay scoring | Value decreases exactly -1/2s, floors at 1, correct score added on eat | Decay present but wrong rate/floor, or UI/score mismatch | No decay, or decay cosmetic only (score unaffected) | | GT3 | Boundary switch trigger | Wall = death for apples 1–3, wrap-around from apple 4 onward, verified via screenshots pre/post switch | Switch happens but off-by-one (e.g. triggers at apple 3 or 5) | No conditional switch (always wrap or always wall) | | GT4 | Wrap-around correctness post-switch | Snake reappears correctly on opposite edge, self-collision and spawn rules still hold | Wrap works but causes visual glitch or occasional double-count | Wrap-around broken, crashes, or game freezes | | GT5 | Self-collision | Correct head-vs-body detection, no false positives/negatives | Minor edge case miss (e.g. only on tail growth frame) | Non-functional or trivially wrong (always/never triggers) | | GT6 | Game over & restart | Clear game-over screen, final score shown, functional restart with clean state | Restart works but some state leaks (score, apple value) | No game-over screen or restart broken/absent | ## Grading method Render each submission in headless Chromium via Playwright. Screenshot at start, after eating 2 apples (still wall-death mode), after eating 4 apples (wrap-around mode active), and at game-over. Manually verify apple value display changes over a real 4-6 second idle window to confirm actual decay timing rather than a cosmetic counter. Check console for errors during wrap transitions.