The shipped two-clause samePiece, validated: the escapee class is closed by production code

Validation, the shipped clause

Study Overview

The Urgent Ask, Shipped

Study 16 ended with one urgent ask: a real catalog wrong pair, the Eat All the Clocks recording against the Mop Heart chart, clears the 1.6 ratio gate at 1.74 and is told apart only by its median local cost of 0.323, so production's ratio-only samePiece would call two different Badverbs songs the same piece. Stem Shovel shipped the fix the next day, citing the escapee pair by name in the commit message: samePiece now requires the ratio at or above 1.6 AND the median local cost at or below 0.30, both constants exactly the numbers the bench recommended, with the alignment core untouched. Study 16b is the micro validation of that ship, registered before any scored run: the shipped code ported verbatim and pinned by test, including a check that the 0.30 boundary is inclusive, then run against Study 16's full pairing matrix on the migrated Badverbs corpus. Zero spend, zero model calls, 89 mechanically graded cells.

A Perfect Matrix Through the Shipped Code

All gates pass and both registered predictions hit. The four right pairs stay samePiece true at ratios 1.90, 3.18, 2.63 and 1.96. All eleven alignable wrong pairs come back false, including the escapee: the shipped second clause rejects the 1.74 pair that the previous production check would have accepted, which is the one verdict the ship existed to change. The one non-alignable pair stays refused by production's own rule that a chart running more than twice the recording's length is turned away, recorded as in Study 16. And nothing else moved: the clause touches only the verdict, so every section placement and every hedge flag on the right pairs is byte-equal to Study 16's committed results, 74 of 74 comparisons.

The Alignment Thread Closes

The verdict delivered back to the app: nothing to change. This closes the thread that opened at Study 12, when alignment first took placement from the models: placement validated across Studies 12, 15 and 16, and the same-piece gate validated twice and hardened once along the way, with Study 14 catching the constant before release, Study 15 designing the local-cost clause from the complementarity analysis, Study 16 finding the real pair that made it urgent, and Study 16b confirming the shipped code on the full matrix. The two-clause check is no longer a bench recommendation; it is production behavior, measured.