# PREDRAFT — Arm A2-id — both directions, written before any reconciliation ran
Date: 2026-07-25. Per O7, both directions are drafted now so that neither reads as
the hoped-for outcome. Placeholders `⟨…⟩` are filled only from RESULTS. The final
report may land between these poles or outside them ("cannot answer" and "procedure
defective" remain first-class).

---

## Direction 1 — the published values are substantially recomputable

Under the preregistered join (⟨key, coverage⟩) and policy ⟨P⟩, ⟨k⟩ of 17 addressed
columns reconcile at T0 for ⟨n⟩/34 rows; the permutation null places the observed
match counts at ⟨position⟩ relative to the null's median and 5th percentile. Cell
accounting: ⟨a⟩ reconciled / ⟨b⟩ divergent / ⟨c⟩ no antecedent / ⟨d⟩ structural
(sum 1,292). The boundary remains: reconciliation at published precision under a
declared policy is not a statement that the publisher computed the values this way,
and 21 of 38 columns have no antecedent in the instance table at all — the majority
of the artifact's surface is out of scope of any recomputation claim. High agreement
on the addressed minority must not be read as validation of the artifact as a whole,
nor as exact reconstruction where only tier-N tolerance was achieved.

## Direction 2 — the published values are not recomputable from the instance table
(or only partially, or the question cannot be answered from these bytes)

Under every declared policy, ⟨k⟩ of 17 addressed columns fail T0 for ⟨n⟩ rows
(cell accounting as above). The boundary remains: failure to reconcile under the
declared policy family is not a statement that any published value is wrong, that
any process failed, or that the two artifacts were ever intended to correspond —
the publisher's own pipeline sources the leaderboard numbers from `scores.json`,
not from the instance sidecars, and the sidecars are described as best-effort
extractions from archives. Divergence is compatible with: policies outside the
declared family, upstream data unavailable to this study, or the two artifacts
answering different questions. No cause is attributed (L2). If the join itself
fails coverage, the primary finding is "this corpus cannot answer the question,"
and the measurements are reported as boundary artifacts, not as discrepancies.

## Shared, direction-independent commitments

- The executor-row disclosure (§3.1) appears in both directions, with the same
  mechanical rule, whatever it shows.
- All policies, nulls, floors, ranges, and the 1,292-cell accounting appear in both.
- Neither direction licenses "error", "failure", "fabricated", or any attribution.
- Suitability (O14) is decided from coverage/gates, not from which direction obtains.