The Single-Source Oracle Was Exact. The Protocol Wasn't.

An exact answer from a single frozen source does not make an experiment exact.

E029 was designed to measure whether Luna could use two unranked moves from a single frozen Syzygy tablebase source while retaining the same unrestricted action tool in both arms. The advice was exact relative to the frozen Syzygy corpus and adapter, not independently replicated truth.

The intended comparison was about using advice, not following a software- constrained menu. The suggestions change information while the action space stays fixed. The full terminal design and result belong to essay 29 (In-page hash only); this note preserves only the pre-terminal chronology of why no interim effect was publishable.

The source and execution chronology had to survive two failures without rewriting either one into success.

A corpus can be legal and still be infeasible

The preregistration fixed six five-piece material families, 360 generated candidates, 60 final positions, two arms, two fresh replicates per arm, and a balanced ten-way schedule. The first source selector crossed side to move, material family, and root outcome too aggressively. It required every material-by-color cell to contribute outcome classes that some endings simply did not contain.

The immutable v1 acquisition exposed the contradiction. One registered cell had no eligible decisive-loss position. The selector stopped with an exact shortfall instead of silently drawing more boards, changing the seed, adding a material family, or borrowing a nearby outcome.

That is a source NO-GO, not a disappointing scientific result. No Luna call had been made. No model outcome existed. The v1 candidate manifest, raw tablebase journal, adapter, and failure record remain preserved as the version that did not work.

Source Selection Amendment 1 changed selection, not the question

The repair was frozen after tablebase feasibility was known but before any model response. That timing matters. It makes the amendment post-label but pre-outcome.

The amended selector retained the registered scientific margins that were jointly feasible:

• 60 positions and 240 planned calls;
• 10 positions in each side-to-move by root-outcome cell;
• 10 positions from each of six material families;
• five positions in each material orientation within every family;
• 30 ABBA and 30 BAAB blocks; and
• balanced first-versus-second display order for the two suggestions.

Selection became a deterministic integral min-cost flow over the already frozen 360-position pool. It generated no new position and performed no new tablebase lookup. The v2 corpus reached the exact registered margins, retained complete labels for every legal move, and passed a second provider-free parse.

This repair does not make v1 successful. It creates an amended v2 source whose limitations include the post-label feasibility observation.

Then the interface failed after the answer returned

The first execution frontier reached the provider and returned successfully, but the event boundary rejected the first action before any scored event was written. The prompt asked for UCI notation while also displaying suggestions in UCI and SAN. The unrestricted tool accepted a generic string. The inherited event parser accepted only UCI.

A legal SAN answer was therefore transport-complete and semantically recoverable, yet incompatible with the registered event schema.

This was not a provider failure. It was not a bad chess move. It was not missingness. And it did not authorize another chat. The durable journal already contained the paid intent, returned exchange, and authoritative identity metadata for each protected attempt in the first frontier.

The preflight-rejected Execution Recovery Amendment 1 draft had a replay guard that treated a repeated request-body hash as evidence of duplication, but two prospectively registered replicates in the same arm can legitimately have byte-identical requests. No provider call occurred under that rejected draft. Its rejection is preserved rather than promoted as a usable repair.

Recovery Amendment 2 separates syntax from identity

The accepted recovery uses the frozen attempt identity—not request-body uniqueness—as its no-replay boundary. For each returned action, it applies one uniform interpretation rule across both arms:

1. accept a syntactically valid legal UCI move;
2. otherwise accept only an unambiguous legal SAN move against the frozen position;
3. store every accepted move in canonical UCI; and
4. retain illegal or unparseable strings as observed malformed actions with utility zero.

The raw response evidence remains unchanged. Prompts, labels, advice order, schedule, model identity, budget, and estimand remain unchanged. The protected first-frontier attempts must prove that they need no new chat before their events are reconstructed. Later registered replicates remain eligible even when their request bytes match an earlier replicate.

This is a narrow interface amendment. It cannot rescue an unfavorable move, replace a provider failure, or change an experimental arm.

Why the result gate stayed closed in the pre-terminal chronology

An executing experiment is not a public result. Neither returned actions nor a partial arm contrast belonged on the site at the pre-terminal point recorded here. E029's public protocol capsule therefore contained only design counts, frozen source dispositions, amendment chronology, claim limits, and content identities. It excluded positions, prompts, move strings, responses, request identifiers, provider metadata, live progress, tokens, costs, account state, and every interim outcome.

Only a terminal, content-addressed analysis may open the result gate. That analysis must reconcile all 240 registered cells, distinguish provider missingness from malformed observed actions, bind the amended corpus and recovery evidence, and report the registered paired utility estimand with its uncertainty and unresolved-cell bounds.

A post-run publication audit found one more boundary that the terminal record must disclose: the execution lock bound the preregistration, corpus, runner, and runtime schemas, but did not bind the analysis implementation hash promised by the preregistration. The estimands, inference seeds, and decision rule were specified in prose before execution; the exact analyzer bytes were not pre-hash-locked. A later deterministic implementation can calculate those registered quantities, but it must be labeled through an explicit analysis provenance amendment rather than described as a fully pre-locked analysis implementation.

For this pre-terminal chronology, the strongest accurate statement is procedural: E029 had a frozen single-source exact-tablebase design, an immutable source NO-GO, a disclosed post-label source amendment, a transport-complete interface incident, and a no-replay recovery contract. The later terminal result is reported separately in essay 29 (In-page hash only).

What the pre-terminal chronology could—and could not—say

At the time covered by this pre-terminal chronology, a later result could have answered whether two unranked single-source exact-policy suggestions changed Luna's exact 50-move-rule utility on this synthetic five-piece benchmark relative to board-only Luna.

It cannot establish full-game transfer, Elo improvement, base-model learning, human-like reasoning, stronger vision or search, benefits outside tablebase states, or unassisted Luna skill. Success would measure the system's ability to use a tiny single-source exact oracle. Failure would be equally local to this interface, model snapshot, population, and scoring policy.