{"data":{"kind":"file","path":"README.md","version_id":"rd9g2h3035vt8jn2n29uj86i","entry":{"name":"README.md","path":"README.md","is_directory":false,"size":1948,"modified_at":"2026-07-14T00:48:00.484000","content_hash":"2c1ddd424f47c841df3695d83d1dbbabea59f3556206411216d6155fb1630b97"},"entries":[],"content":"# glyph-pipelines\n\n**Program synthesis by selection.** The model emits one-line programs in a\ncontent-addressed composition language — glyphs are 16-hex sha256 prefixes of\nfunction ASTs — instead of writing code. Reward is machine-checked by\nexecuting the program on held-out I/O examples in a closed runtime\n(whitelisted primitives only, no eval path), so the reward cannot be hacked\nby plausible-looking output.\n\n## Task\n\n```\nGiven a list of integers: keep the odd elements, for each, multiply by 3, then take the sum.\nExamples (input -> output): [9, 3, 6, 10] -> 36; ...\n```\n→ expected completion (one line, nothing else):\n```\n#filter_ (#a76dc3328b817d25) | #map_ (#ecd119db259287c8 3) | #2248f2928e93f83e\n```\n\nTasks are procedurally generated — type-correct by construction, deterministic\nper seed, infinite supply — across a curriculum keyed to language features:\n\n| tier | requires |\n|---|---|\n| 1 | single glyph |\n| 2 | composition |\n| 3 | higher-order (map/filter with sub-pipelines) |\n| 4 | let-bindings (`>name` / `$name` — two live values at once) |\n\nMeasured frontier-model baseline (one-shot, July 2026): tier 2 ≈ 100%,\ntier 4 ≈ 25% — the let-binding tier is where the learning signal lives.\n\n## Reward\n\n`1.0·passes + 0.15·shortness + 0.10·reuse` — `passes` (all held-out and\nshown examples reproduce) is a hard gate; shortness (MDL) and reuse\n(fraction of refs hitting learned/derived glyphs) only score on correct\nprograms. Each component is also exposed as a metric.\n\n## Usage\n\n```python\nimport verifiers as vf\nenv = vf.load_environment(\"glyph-pipelines\",\n                          num_train=400, num_eval=80,\n                          tiers=(1, 2, 3, 4), seed=0)\n```\n\n`glyphcore` (the language runtime) is vendored — the package is\nself-contained. Local dev install: `pip install --no-deps -e .` if your\nplatform can't build `prime-sandboxes` (unused here).\n\nPublish: `prime env push` from this directory.\n","encoding":"utf-8","truncated":false,"total_bytes":1948},"status":null}