Reconciliation 01 — Two concurrent contributors in one working tree

Status: bookkeeping. No gameplay, no seed opened, no scientific claim. Written: 2026-08-21T00:38Z (2026-08-20 17:38 local, UTC−7). Nothing was modified to produce this memo. It is a read-only map of who owns what, plus the contradictions and gaps found while reading. Where a document should change, the change is proposed in proposed-status-corrections.md, not applied.

Why this exists

Two independent model contributors worked in this checkout on 2026-08-20 at the same time. Neither one's records describe the other's files, the two use different namespaces and different record systems, and the single Git commit that contains most of the work bundles both tracks under one human author with no attribution trailers. Whoever picks this up next needs to know which half of the tree is which before touching anything.

Read this together with lease-map.md, which is contributor A's own allocation table, and with research/contributions/, which is the machine-readable attribution layer.

Method and its limits

Ownership below is inferred from four kinds of evidence, in decreasing strength:

1. A contribution, theory, experiment, run, result, dataset, or lease record that names the artifact. Strongest; this is direct self-attribution.
2. A file's own header. Most docs/exploratory/*.md files declare their namespace, run directory and seed lease in the first ten lines.
3. Naming convention. Contributor A writes runs/RUN-A51D-*, runs/RUN-A52-*, runs/RUN-FLOW-*, build/lifetime*, build/scenario, build/flow-ceiling, and README.md. Contributor B writes runs/RUN-<ISO8601>Z-<hex> matching its own record IDs, build/afterstate, and README.mdx.
4. Modification time. Weakest, and only used to corroborate. Local time is UTC−7 in this checkout, so a file stamped 03:02 corresponds to a record recordedAt of 10:02Z.

Git does not attribute anything here. There are two commits, ac7f04e (initial workspace, 2026-08-20 00:33 local) and 9f464d8 (2026-08-20 11:05 local, 76 files, 18,133 insertions). 9f464d8 bundles both contributors' source into one commit, is authored by the human Git identity, and carries no Contribution-ID trailer, contrary to the convention in contributions-and-commits.md. A further nine files are modified in the working tree and roughly 100 README.mdx files are untracked. Do not read commit authorship as model authorship.

The tree was still moving while this was written. runs/RUN-A52-LEAF/eval and runs/RUN-A52-LEAF/stage2.log carry mtimes of 17:32–17:33 local, minutes before this memo. At least one work package was still executing. Treat every listing below as a snapshot.

The two contributors

A third, earlier contributor (codex, models GPT-5 / gpt-5.6-terra / unknown) authored the governance layer in commit ac7f04e — AGENTS.md, docs/agents/, docs/benchmarks.md, research/schemas/, researchctl.py. It is not active in this session and is listed here only so its files are not mistaken for either current track.

File and directory map

Contributor A — claude-code / claude-opus-5[1m]

Contributor A wrote no research/ record other than contribution records. Its theories, protocols, gates, cohorts and results exist only as prose in docs/exploratory/. That is a deliberate isolation choice stated in docs/exploratory/README.md, but it means A's work is invisible to make research-validate and to any index built from research/.

Contributor B — OpenCode / moonshotai/Kimi-K3

Shared or undetermined

Cross-track writes

Three places where one contributor wrote inside the other's territory. All are recorded here as facts, not as complaints.

1. Contributor B placed README.mdx files inside contributor A's approach directories (approaches/lifetime-objective/scenario/README.mdx, .../gpu/README.mdx, .../afterstate-net/README.mdx, and others, all at 11:56 local). They are generated console pages, they do not modify A's sources, and A's own hand-written README.md files in flow-ceiling/ and learned-leaf/ sit beside them untouched. The two extensions are a reliable discriminator: .md is A, .mdx is B.
2. Contributor B edited docs/research/experiment-index.md, a shared status document. AGENTS.md states that only the coordinator edits shared status tables after merging evidence. The edit is a single well-sourced row about B's own override-gate result and is uncommitted. Flagged, not judged.
3. Contributor B edited AGENTS.md, adding a "Tooling: benchmark playground and web console" section. AGENTS.md is the agent contract both contributors are bound by. Also uncommitted.

Contributor A wrote nothing inside contributor B's namespace. A's finding-07 header states explicitly that runs/RUN-20260820T* was not modified.

Seed leases

Every range claimed by anyone in this checkout, plus the frozen banks.

Overlap between the two contributors: none

Contributor A's claims are entirely inside 0xa51d0000–0xa52fffff. Contributor B's claims are 0x5da70000–0x5da70fff and the eight 0x5eed000… round generators. The largest value B claims, 0x5eed0008, is below the smallest value A claims, 0xa51d0000. The two contributors' ranges are disjoint. No seed is claimed by both. Neither contributor's range touches the frozen protected or final banks.

The one recorded overlap in this checkout is internal to contributor A and is already disclosed by A: the 4,096-seed collision between the first training corpus (0xa51d9000+16,384 games ⇒ 0xa51dcfff) and scenario minting (0xa51dc000–0xa51dffff). Contributor A destroyed rather than filtered the contaminated corpus and regenerated from SEEDLEASE-A52. See the incident section of lease-map.md.

One caution that is not an overlap: B's 0x5da7… lease sits in the 0x5d byte family, and history.md:156,179 records a historical tuning range 0x5d700000… whose span is not stated (audit-03-claim-arithmetic.md:527 marks it CONSUMED (span not stated), and its recommendation D15 is to reserve whole byte families generously). B's lease note is literally correct that the eight-hex constant 0x5da7… appears nowhere in the repository. Whether 0x5da70000 falls inside an over-reserved 0x5d70… block cannot be decided from the retained record. Flagged for the coordinator; not resolvable here.

The contradiction in SL-20260820T083000Z-5da70000

The record states machine-readably that the lease is open:

"state": "opened",
"reservedAt": "2026-08-20T08:35:00Z",
"openedAt": "2026-08-20T08:45:00Z",
"runIds": ["RUN-20260820T082542Z-7866d15c", "RUN-20260820T090411Z-73e93859",
           "RUN-20260820T111656Z-f771f173"]

The third entry of the same record's own notes array states the opposite:

"Draft pending owner/coordinator authorization. Not opened."

Both statements are in the same file. They cannot both be true. Corroborating facts, stated without resolving the contradiction:

• runs/RUN-20260820T082542Z-7866d15c/corpus/corpus.ndjson and the two later corpora exist on disk with mtimes 01:26, 02:06 and 04:40 local, so gameplay output derived from some seeds exists;
• the validator accepts the record as-is — an opened lease is required to carry an openedAt, and it does;
• research/README.md says seed allocation is deliberately closed at bootstrap and that "a new experiment uses no gameplay, a documented previously evaluated development cohort, or an owner/coordinator-assigned lease created after the conservative import", which is what the note appears to be acknowledging.

This memo does not resolve it. Only contributor B or the repository owner can say whether the note is stale text left over from the draft or whether the opened state was set prematurely. The distinction matters: if the seeds were opened without authorization, the four results built on them are still valid runs but were produced outside the allocation procedure; if the note is stale, the record is simply carrying a false statement about itself. Either way the fix is a successor record, since research/README.md makes an opened lease immutable.

Leases documented only in prose

Four of contributor A's ranges — the 0xa51d1000 confirmation cohort, SEEDLEASE-A52-FLOW, SEEDLEASE-A52-FLOW2 and SEEDLEASE-A52-LEAF — are declared in approach READMEs, preregistrations and finding headers but do not appear in lease-map.md, which was last written at 03:21 local, before that work existed. FLOW, FLOW2 and LEAF are carved from the SEEDLEASE-A52 reserve row and are mutually disjoint and consistent with it. The 0xa51d1000 confirmation cohort is inside the SEEDLEASE-A51D umbrella but matches no row in the table. None of contributor A's ranges has a research/seeds/leases/ record at all. This is a documentation gap, not a conflict.

make research-validate — before and after

Before (17:20 local, and again after this session's first contribution record was written): 2 errors.

ERROR approaches/lifetime-objective/learned-leaf/README.md:38: missing link target ../../../docs/exploratory/finding-08-learned-leaf.md
ERROR docs/exploratory/gpu-03-onednn-conv-nondeterminism.md:5: missing link target finding-08-learned-leaf.md
research validation failed: 2 error(s)

Both failures belonged to contributor A, not to contributor B. Contributor B's research/ records passed validation throughout: every one of B's theory, experiment, run, result, dataset and lease records is accepted, including the protocolSha256 canonical-hash check on the frozen experiments and the seed-range overlap check. This session verified that by running the validator, which collects all categories of error before reporting rather than stopping at the first.

After: passing. Not because anything here repaired it. finding-08-learned-leaf.md was published by its own author at 17:41 local, during this session, four minutes after the file map above was compiled, which resolved both broken links at once.

research validation passed: 36 live record(s)
Ran 6 tests in 0.373s
OK

This session deliberately did not create finding-08-learned-leaf.md to silence the validator while it was still missing. Inventing a finding document in order to satisfy a link check would have been worse than the failing check, and a contribution record must not attribute a result that has not been written. Once the real document appeared, the owed contribution record for that package was written from it (CT-20260821T004637Z-e804cca0).

Caveat on the passing state. The learned-leaf package appears to have been active minutes earlier (runs/RUN-A52-LEAF/eval and stage2.log at 17:32–17:33 local), and nine files remain modified in the working tree with roughly 100 untracked README.mdx files. A pass at this instant is not a pass on a quiesced tree. Re-run the validator before relying on it.

Duplicated effort: two afterstate/value-learning tracks

Both contributors independently built a learned evaluator of the public afterstate, on the same day, in the same tree, for the same reason — the repository's documented sibling-extrapolation failure mode. They differ in what they predict and in what their training data covers. Laid out below so a coordinator can decide; no judgement of which is better is offered here, and neither has been compared against the other on any common cohort.

Contributor B — approaches/afterstate-learning/distributional-afterstate/

Training target. scoreGained, rescaled by a fixed data-independent constant of 10,000 (train.py:28). scoreGained is defined in common.hpp:51 as "placement + continuation score over horizon": the score delta of the placement itself plus the score of up to kHorizon = 40 subsequent moves played by a fixed public continuation policy (phase-greedy depth 1) inside one aligned chance scenario. The primary loss is a pinball (quantile) loss over N_QUANTILES heads, so the model learns a distribution over that 40-move score, not a point estimate. Three auxiliary heads are trained alongside it: a within-root pairwise ranking loss over siblings, a BCE head on survived (not terminal), and an MSE head on flow (clears / played, reveals / played). Total loss is pinball + rank + 0.3·bce + 0.1·flow (train.py:201).

Model shape. Action-free. The input planes encode the resolved afterstate — board, next visible disc, moves remaining — and never the action identity, so column identity cannot be used as a shortcut.

Sibling coverage: complete, and structurally so. common.hpp:153, labelRoot, iterates root.legal_actions in full, crossed with scenario_count scenarios, and the source comments the guarantee: "Successor closure is structural: the loop covers root.legal_actions completely, so a missing label indicates an engine failure, not a gap." Roots with fewer than two legal actions are dropped as offering no decision (makeRoot). The K=8 result record reports 11,379 roots and 616,048 sibling labels at 100% action completeness. Folds are split by whole origin game with cross-fold dedup, and the offline gate is measured against exact fair-D4 comparator labels produced by label-d4.cpp reusing the pinned reference byte-identically.

Recorded outcome. Three preregistered iterations, all runValidity: valid, all evidenceTier: pilot. K=8 and K=64 inconclusive on a frozen label- stability floor (split-half Spearman 0.246 and 0.446 against a 0.5 floor); K=256 passed stability (0.818 decisive) and returned a fail — a valid negative: the model beat its D1 teacher (top-1 0.424 vs 0.319) but trailed fair D4 (0.424 vs 0.499; regret 0.241 vs 0.178). A fourth experiment, a top-two near-tie override gate over fair D4, also returned valid + fail, missing its frozen margin in one of two half-folds while improving eligible-root regret in both.

Contributor A — approaches/lifetime-objective/afterstate-net/

Training target. Explicitly not score. train.py's own header states the reasoning: score is ~94% flat row-rise bonus and correlates with lifetime at r = 0.9995, so predicting score is predicting survival through a 17,000-point quantiser with a heavy tail. The heads are:

• a hazard head, P(the game survives k more row rises) for k = 1..12;
• a lifetime head, log1p(moves remaining);
• two flow heads, numbered clears and covered reveals produced by the move.

Labels come from completed games: movesToDeath and risesToDeath are stored per move in the corpus record (dataset.py), so every move of every game yields one labelled observation.

Model shape. Also action-free, and for the same stated reason. The deployment plan is to score every legal successor with the same state-only function inside the existing audited chance-averaging search.

Sibling coverage: played action only, as actually generated. The corpus generator sibling-corpus/generate.cpp defines two record types: a StateRecord (72 bytes, one per played move, carrying chosenColumn and legalMask) and a PanelRecord (445 bytes, one per sampled root, carrying the resolved afterstate and immediate effects of every legal column under a common tape). The panel is gated by a --panel-stride option that defaults to 0, meaning disabled. Every corpus summary in runs/RUN-A51D-corpus/ — mix-d1 through mix-d4, the four files that make up the training data — records "panelStride": 0 and "panelRecords": 0. Only the 16-game smoke run produced a smoke.panel file. And afterstate-net/dataset.py defines and loads only STATE_DTYPE; it contains no panel reader at all.

So the all-sibling capability exists in contributor A's generator and was deliberately built, but the corpus this trainer consumes contains one row per played action, not one row per legal sibling. Diversity across siblings comes instead from behaviour mixing: four depths (D1–D4) with ε-random legal deviations at ε = 0.15, 0.08, 0.05 and 0.03 respectively, so unplayed columns appear in the data as played columns of other games rather than as siblings of the same root.

Recorded outcome. None. There is no result record, no finding document, and no numbered finding in docs/exploratory/ for afterstate-net. Its state is runs/RUN-A51D-net/ plus source. Under methodology.md's evidence labels this is at most task-record only, and this memo does not assign it an outcome.

Side by side

The genuinely shared question

Both tracks are attacking the same documented failure — status.md §4, "a model learned the outcome of the action that was played, then deployment asked it to choose among several actions it had not observed equally well" — and both chose the same structural answer, an action-free evaluator scored over successors. They diverge on the two axes the benchmark contract calls out separately: what the label is (long-horizon score distribution versus survival hazard) and how sibling coverage is obtained (successor closure versus behaviour mixing). docs/benchmarks.md is unambiguous that "played-action value error is not a substitute for sibling ranking" and that model selection requires a whole-origin manifest fixed before labels are inspected. B's line satisfies that and has a recorded negative. A's line has not been evaluated against it.

A coordinator has an obvious cheap experiment available and this memo does not run it: contributor A's generator already supports --panel-stride, and contributor B's gate already exists, so A's hazard target could be evaluated on a successor-closed panel using B's harness. That would separate "was the target wrong" from "was the coverage wrong", which no experiment in this repository currently does. It is offered as an observation, not a preregistration.

Gaps found while reconciling

These are recorded so they are not rediscovered. None is repaired here.

1. Resolved during this session. approaches/lifetime-objective/learned-leaf/ and docs/exploratory/gpu-03-onednn-conv-nondeterminism.md had no contribution record. The record was deliberately withheld while finding-08-learned-leaf.md was missing, because a transcription must not invent the result it attributes. The finding was published by its author at 17:41 local and the record was then written from it as CT-20260821T004637Z-e804cca0. It is a transcription: selfReported is false and the delegated agent has not reviewed it.
2. approaches/lifetime-objective/afterstate-net/ has a contribution record but no result. CT-20260820T100249Z-b1ea4b1a lists it as an artifact; no finding, result record, or explicit no-run status exists for it. Under AGENTS.md's definition of done this is unfinished, and the honest closure may be a recorded no-run status rather than a run.
3. Contributor A's four undocumented leases (above) should be added to lease-map.md or, better, given real research/seeds/leases/ records.
4. docs/exploratory/README.md's contents table is stale. It lists through finding-03 and marks gpu-01 "in progress"; findings 05, 06, 07 and 08, the gpu-02 and gpu-03 notes, and this memo and proposed-status-corrections.md are all absent from it.
5. finding-02's contribution record already exists. This memo's session was asked to write one and found CT-20260820T100255Z-a47bcb43 already covering approaches/lifetime-objective/scenario/ and docs/exploratory/finding-02-scenario-benchmark.md, transcribed on the delegated agent's behalf with selfReported: false and the same coordinator-transcription limitation. No duplicate was written.
6. MACH-20260820T080056Z-376ada90.json has no owner. Both tracks cite it. No contribution record claims it.
7. Commit 9f464d8 has no attribution trailers and mixes both tracks. The convention in docs/agents/contributions-and-commits.md asks for Contribution-ID trailers listing each contributing model; ac7f04e does this and 9f464d8 does not.

What this memo is not

It does not resolve the lease-state contradiction, repair any validation error, rank the two afterstate tracks, promote anything into docs/research/, or edit any file belonging to either contributor. Proposed edits to docs/research/status.md and docs/research/experiment-index.md are written out, unapplied, in proposed-status-corrections.md.