Documents
The documents this research is bound to: how a game is scored, how a claim is tested, what has been recorded, and what an agent working on the program must follow.
Reader documents
The rules of the program: how a game is scored, how a claim is tested, where the evidence stands, and what comes next.
- Research Statusdocs/research/status.md
- Research Methodologydocs/methodology.md
- Standard benchmark contractdocs/benchmarks.md
- Drop7 strategy catalogdocs/strategies.md
- Research roadmap toward a million-point policydocs/research/roadmap.md
- Reproducibility Guidedocs/reproducibility.md
- Provenance and Evidence Boundariesdocs/provenance.md
- Drop7 Policy Protocol (D7P)docs/d7p-protocol.md
Ledger
Every configuration that was run and what came of it, as recorded.
- Experiment indexdocs/research/experiment-index.md
- Detailed Experiment Ledgerdocs/research/history.md
- evidencedocs/research/status/evidence.md
Findings and audits
Exploratory findings and independent audits, each labelled with its own evidence tier.
- Exploratory workspace (separate namespace)docs/exploratory/README.md
- Audit 01 — Simulator and engine fidelitydocs/exploratory/audit-01-engine-fidelity.md
- Audit 02 — the fair depth-4 reference policydocs/exploratory/audit-02-fair-d4.md
- Audit 03 — arithmetic and claim integrity of the historical evidence ledgerdocs/exploratory/audit-03-claim-arithmetic.md
- Audit 04 — Blind spots in the existing Drop7 corpusdocs/exploratory/audit-04-blind-spots.md
- Audit 05 — Optimistic-teacher pretraining and curriculum ordering for an NNUE evaluatordocs/exploratory/audit-05-optimistic-curriculum.md
- Audit 06 — Engine efficiency: where the fast engine can still go, and where it cannotdocs/exploratory/audit-06-engine-efficiency.md
- Design 01 — A fully-specified scenario benchmark for Drop7 policiesdocs/exploratory/design-01-benchmark-suite.md
- Finding 01 — Hardcore score is survival time, and survival is a flow-balance problemdocs/exploratory/finding-01-score-is-survival.md
- Finding 02 — A fully specified Drop7 scenario, its exact optimum, and what perfect knowledge is actually worthdocs/exploratory/finding-02-scenario-benchmark.md
- Finding 03 — The 25-move rollout veto under corrected 17,000-point scoringdocs/exploratory/finding-03-rollout-veto-17k.md
- Finding 04 — The fair-D4 risk constant is saturated; it is not a leverdocs/exploratory/finding-04-terminal-utility-saturated.md
- Finding 05 — Search depth pays only when the chance estimator is unbiaseddocs/exploratory/finding-05-chance-strata.md
- Finding 06 — Flow balance is achievable: Drop7 Hardcore is a control problem, not a losing battledocs/exploratory/finding-06-flow-ceiling.md
- Finding 07 — Separating planning from information: at most 41% of the clairvoyant advantage is hidden knowledgedocs/exploratory/finding-07-fair-planning-ceiling.md
- Finding 08 — A learned leaf helps, and the chance-bias cap hypothesis is wrongdocs/exploratory/finding-08-learned-leaf.md
- Finding 09 — The chance node's two random quantities are welded together, and unwelding them is worth a whole plydocs/exploratory/finding-09-reveal-sampling.md
- Finding 10 — The scenario suite does not rank fair policies, and the reason is the horizondocs/exploratory/finding-10-suite-validation.md
- Finding 11 — Distilling the *fair* planner: the corpus was right, the teacher was notdocs/exploratory/finding-11-planner-distillation.md
- Finding 12 — The eight-tape cohort was unrepresentative: a re-baselined fair-planner ceilingdocs/exploratory/finding-12-fair-planner-ceiling-extended.md
- Finding 13 — a semantics-preserving fast engine for the fair expectimax searchdocs/exploratory/finding-13-fast-engine.md
- Finding 14 — reweighting the fair leaf toward achievable clearsdocs/exploratory/finding-14-leaf-reweight.md
- Finding 15 — depth 5 with an exact chance estimatordocs/exploratory/finding-15-depth5-exact-estimator.md
- Finding 16 — The two axes do not compound: the arms finding-09 left opendocs/exploratory/finding-16-factored-reveal-sampling.md
- GPU enablement: PyTorch on ROCm / gfx1151 (Radeon 8060S, Strix Halo)docs/exploratory/gpu-01-rocm-enablement.md
- Upstream bug: multithreaded OpenBLAS SGEMM returns wrong results on Zen 5docs/exploratory/gpu-02-openblas-sgemm-race.md
- GPU/CPU note 03 — `torch.nn.Conv2d` on the CPU is nondeterministic on this hostdocs/exploratory/gpu-03-onednn-conv-nondeterminism.md
- Seed lease map for the exploratory workspacedocs/exploratory/lease-map.md
- Proposed status corrections — a proposal only, nothing applieddocs/exploratory/proposed-status-corrections.md
- Reconciliation 01 — Two concurrent contributors in one working treedocs/exploratory/reconciliation-01.md
- Record consistency notesdocs/exploratory/record-consistency-notes.md
Hardware
Profiles of the machines the runs were timed on.
- AMD Ryzen Halo research plandocs/hardware/amd-ryzen-halo.md
Agent contracts
The contracts an agent follows when it works on this program. Agent-facing.
- Agent research handbookdocs/agents/README.md
- Writing strategy and concept pages for the consoledocs/agents/approach-page-template.md
- Contributions, validity, and commitsdocs/agents/contributions-and-commits.md
- Autonomous research orchestrationdocs/agents/orchestration.md
- What this project is — a note for agents and automated reviewersdocs/agents/project-nature.md