Exploratory workspace (separate namespace)

This directory holds exploratory, second-opinion work that is deliberately kept out of the main research record until it has earned promotion. Nothing here modifies docs/research/, artifacts/, frozen protocols, or any existing approach source. It is a parallel namespace, not a replacement.

Why this exists

The repository owner asked for two things at once:

1. progress toward the frozen million-point qualification standard, and
2. an independent second set of eyes on work that has already been attempted, preferring a re-examination of an existing approach over a new direction.

Audit output and new exploratory evidence are therefore written here, with explicit provenance, so that the existing record stays intact and any correction to it is a deliberate, reviewable act rather than a side effect.

Namespace and isolation rules used by this work

Concern	This workspace uses	Never touches
Documentation	docs/exploratory/*.md	docs/research/*, docs/*.md
Source	approaches/lifetime-objective/*	any existing approach directory
Build output	build/lifetime/	build/native-suite, build/fair-depth4
Run output	runs/RUN-A51D-*/	artifacts/
Seeds	lease SEEDLEASE-A51D = 0xa51d0000–0xa51dffff	every historical range

Seed lease SEEDLEASE-A51D

Role: exploratory development diagnostic. Once read, these seeds are development data forever and can never become confirmation evidence.

The range was chosen by extracting every 8-hex-digit constant appearing in docs/research/history.md, approaches/, src/, research/, and artifacts/, and selecting a prefix that appears nowhere. Historical gameplay seeds cluster densely in 0x3d000000–0x3ea00000; the 0xa51d prefix is unused. This is a conservative choice, not an authorization to allocate STANDARD, QUALIFY, PROTECTED, or FINAL cohorts — none of those are claimed here, and protected and final seeds remain unopened.

Contents

File	What it is
audit-01-engine-fidelity.md	Independent audit of the native and TypeScript rules engines
audit-02-fair-d4.md	Independent audit of the fair-D4 reference policy
audit-03-claim-arithmetic.md	Recomputation of every numeric claim in the historical ledger
audit-04-blind-spots.md	Which rejections were clean and which were confounded
audit-05-optimistic-curriculum.md	Whether optimistic-teacher curriculum ordering was ever ablated
finding-01-score-is-survival.md	Measured decomposition of where Drop7 Hardcore score actually comes from
finding-04-terminal-utility-saturated.md	Valid negative: D4's death penalty is at its stop and is not a lever
finding-05-chance-strata.md	Positive, replicated: exact chance enumeration unlocks the fourth ply
lease-map.md	Seed allocation table and the lease-overlap incident
finding-06-flow-ceiling.md	Flow balance is achievable clairvoyantly; the occupancy equilibrium at ~20 cells
finding-07-fair-planning-ceiling.md	Splitting the clairvoyant advantage into planning vs hidden information
finding-08-learned-leaf.md	Learned survival leaf inside the search; a preregistered prediction refuted
finding-09-reveal-sampling.md	Chance-node decorrelation is exchangeable with search depth
finding-10-suite-validation.md	The scenario suite fails its own validation gate; leaf weights point the wrong way
finding-11-planner-distillation.md	Distilling a legal planner; negative, and the 8-tape cohort correction
finding-12-fair-planner-ceiling-extended.md	Pushing the legal planner's ceiling, and re-baselining it
finding-13-fast-engine.md	3.08x bit-exact engine speedup, and a 10.4x error in the cost model
finding-14-leaf-reweight.md (in progress)	Refitting the frozen leaf's weights toward achievable clear rate
finding-15-depth5-exact-estimator.md	Two withdrawn readings and a detection-floor analysis: the depth-5 contrast was never measurable; chance exactness is what pays (see its §10)
finding-16-factored-reveal-sampling.md	Depth and chance resolution substitute rather than compound; the reveal axis saturates at M≈6; the frontier is flat on top near fair D4
gpu-02-openblas-sgemm-race.md	numpy float32 matmul is silently corrupt on this host
gpu-03-onednn-conv-nondeterminism.md	torch Conv2d is nondeterministic on this host's CPU
reconciliation-01.md	Which contributor owns what, and every seed lease
proposed-status-corrections.md	Proposed edits to the main record, split safe-now vs needs-replication
design-01-benchmark-suite.md	Design of the fully-specified scenario benchmark and its validity gates
gpu-01-rocm-enablement.md (in progress)	Whether the gfx1151 iGPU is usable for training, with benchmarks
finding-02-scenario-benchmark.md	Scenario engine parity, exact solver, and board-clear reachability
finding-03-rollout-veto-17k.md	Retest of the 25-move rollout veto under corrected scoring

Status of everything in this directory

Exploratory. Evidence tier is at most pilot or development. No result here is a qualification claim, and no result here upgrades the historical ledger. Promotion into docs/research/ requires the coordinator to review the underlying run records first.