Cross-Version Reconciliation
The deliberate barrier that covers a structural blind spot of fan-out: a primitive labeled new in one version that actually shipped earlier. A pass over all 396 primitive rows at once, and the two re-claims it caught.
Pass: deliberate cross-version barrier. Scope: every new-primitives table across 63 versions, 396 primitive rows.
Why this pass exists
The per-version battle-test agents fan out in isolation: each one only sees its own release’s changelog slice. That fan-out has a structural blind spot it cannot close from inside a single version: a primitive labeled new in version X that actually first shipped in an earlier version. No per-version agent has the other versions in context, so the mislabel is invisible to it by construction. This reconciliation pass is the deliberate barrier that covers exactly that blind spot. It reads across all versions at once and matches primitives by their identifier tokens (backtick identifiers, slash-commands, flags, env vars), not just by prose, so a re-claim survives even when the later version rephrased the row title.
Method note
Naive exact-name matching across versions caught only workspace.git_worktree and would have missed the known /effort xhigh case, because 2.1.154 retitled that row versus 2.1.111’s xhigh. The detector was switched to identifier-token matching and re-validated: it now flags the 2.1.111 plus 2.1.154 xhigh pair, confirming the blind-spot class is actually covered. Every multi-version token was then hand-reviewed to separate legitimate increments (the same command or flag family extended to a new surface) from genuine re-claims (the same primitive re-listed as net-new).
Findings
| Primitive | First shipped | Re-claimed in | Correction |
|---|---|---|---|
workspace.git_worktree (status-line field) | 2.1.97 | 2.1.98 | 2.1.97 already added the field. 2.1.98 re-lists the identical primitive with a near-identical line and no back-reference. Demoted to “already shipped in 2.1.97.” |
/effort xhigh | 2.1.111 | 2.1.154 (as “extended to Opus 4.8”) | Already corrected: 2.1.154 annotates the row “not new, introduced in 2.1.111” and reclassifies it as an extension. |
Beyond the already-corrected xhigh case, the pass found exactly one uncorrected wrong re-claim (workspace.git_worktree), now fixed. Every other cross-version token is a clean per-version increment.
The lesson
Independent parallel verification catches local errors (hallucinations, mis-quotes, demoted primitives) and is blind by construction to global errors (this new thing was introduced three releases ago). The fan-out’s strength, each agent isolated to its own context, is exactly the source of its blind spot. The fix is a deliberate barrier after the blind parallel stage: one agent, or one human, holding all the records at once. That is the single seat where the expensive model earns its rate.