The $30 proof plan. Local-first. Hard cap USD 30 total, USD 5 per day. No auto-spend. Source: docs/z15/Z15_30_DOLLAR_PROOF_PLAN.md.
| Phase | Experiment | Local? | Cost | What it proves | What it does not prove |
|---|---|---|---|---|---|
| A | Literature map (arXiv + OpenAlex + S2) | yes | $0 | What prior art covers; where the field is structurally thin | Nothing about Z-15 implementation |
| B | Corpus compression on a small public corpus | yes | $0–1 | Pipeline runs end-to-end; token + entropy numerics on a tiny set | Frontier scale; toy only |
| C | Tiny model perplexity / val-loss delta | partial | $1–2 on RTX 4090 | Treatment beats raw text on val loss at sub-1M params | Scaling-law extrapolation |
| D | Agent correctness vs raw-prompt baseline | yes | $1–5 | Hallucination rate, unsupported-claim count, latency, token cost | Absolute correctness |
| E | Proof packet assembly + SHA signing | yes | $0 | Five-artefact bundle renders; BUNDLE_SHA256 verifies | External review |
git clone <this repo> python3 -m z15_engine.cli scout # dry-run by default; for a live arXiv/OpenAlex/S2 fetch: ops/scripts/z15-literature-scout.sh --live --max-results 20 ops/scripts/z15-literature-scout.sh --emit-map ops/scripts/z15-literature-scout.sh --emit-gap-map
python3 -m z15_engine.cli build-corpus --source z15/data --output /tmp/corpus.txt python3 -m z15_engine.cli compress --raw /tmp/corpus.txt --substrate /tmp/substrate.txt python3 -m z15_engine.cli baseline --raw /tmp/corpus.txt
See ops/vast/Z15_VAST_RUNBOOK.md. The runner refuses without Z15_ALLOW_GPU=1. Budget gate ops/scripts/z15-gpu-budget-check.sh refuses any rental whose ETA would breach either cap.
Every run writes to z15_engine/memory/traces.jsonl. Recompute bundle_sha256 against the artefact set and compare to the file's recorded SHA. If they differ, the bundle is invalid.