Seed lease map for the exploratory workspace

This is the authoritative allocation for exploratory work in this session. Every range here is development or training data: once read, it can never become confirmation evidence. No protected (0x7d……) or final (0xd7……) seed has been opened.

Why these prefixes

Candidate prefixes were chosen by extracting every eight-hex-digit constant from docs/research/history.md, approaches/, src/, research/ and artifacts/ and selecting prefixes that appear nowhere. Historical gameplay seeds cluster densely in 0x3d000000–0x3ea00000, with further blocks at 0x2d, 0x2e, 0x3f, 0x4d, 0x5d700000, 0x6d, 0x6e and 0xa5700000. The only 0xa5 constants anywhere in the repository are 0xa511e9b3, 0xa54ff53a and 0xa5700000, none of which fall in the ranges below.

Allocation

Training and evaluation ranges are disjoint by construction: evaluation lives under 0xa51d, training under 0xa52. No model trained on SEEDLEASE-A52 has ever seen an evaluation seed.

Incident: lease overlap, 2026-08-20

What happened. The first training-corpus run was launched at 0xa51d9000 for 16,384 games. 0xa51d9000 + 16384 = 0xa51dd000, so it consumed 0xa51d9000–0xa51dcfff and overlapped the scenario-minting sub-lease 0xa51dc000–0xa51dffff by 4,096 seeds (0xa51dc000–0xa51dcfff). The range was sized in games and the end address was not checked against the allocation table before launch.

Why it mattered. Both roles are development data, so no protected or final seed was touched and no qualification claim is affected. The real hazard was contamination of a benchmark: scenario positions minted from those 4,096 seeds could have appeared in the training corpus, so a model trained on that corpus would have been evaluated partly on positions harvested from its own training games. That is precisely the leak design-01-benchmark-suite.md exists to prevent.

Resolution. The contaminated corpus (d2eps08.states, 1,029,206 records) was destroyed, not filtered, and the corpus was regenerated from SEEDLEASE-A52, a range with no overlap with any evaluation or scenario lease. Regeneration cost 103 seconds. The 4,096 overlapped seeds remain validly allocated to scenario minting; they are recorded above as withdrawn from training so that no future run reuses them for that purpose.

Process change. Every block in SEEDLEASE-A52 starts exactly where the previous one ends, the arithmetic is printed and checked before launch, and the generation script records the consumed end address for each block.

Incident 2: sub-leases declared outside this file, 2026-08-21

What happened. Nine sub-leases were issued to delegated agents in their task briefs and recorded only in those agents' own READMEs and finding headers. They were absent from this table and from research/seeds/leases/, so for several hours the authoritative allocation map was incomplete. An independent reconciliation pass found four of them; the remainder were added at the same time.

Why it mattered. No collision resulted — the ranges were issued sequentially from a single coordinator and are disjoint, and a check against the concurrent OpenCode contributor's ranges (0x5da70000–0x5da70fff and 0x5eed0001–0x5eed0008) confirms no overlap in either direction. But the protection was luck plus sequential issuance, not a record. The first incident in this file was caused by exactly that failure mode: arithmetic done at launch time rather than checked against a table.

Resolution. Every range is now listed above. Machine-readable research/seeds/leases/ records remain outstanding for the 0xa5…… family and are owed before any of this work is promoted beyond the exploratory namespace.

Process change. A sub-lease is not issued until it appears in this table.