Study AV · Whole-system baselines
Which matching policy is right? All gates pass, the default flips to the thousandfold-cheaper judge
Whole-system baselines
Study Overview
The Question AU Could Not Answer
Study AU established that Jev is not the same judge as fable at the PDF pipeline's match stage: a one-directional disagreement, reuse rates of 45.1% against 16.7%, and a gate that declined to fund the deeper experiment on a failed interchangeability check. Agreement cannot say which policy is RIGHT. Study AV, the fourth whole-system study, registered and run by the stem-bench analyst from the hand-off proposal, scores both policies where it matters: two full AS-corpus pipeline passes differing only in the match engine, graded on downstream structure. If generic-heading reuse is actually fine, fable's strictness over-builds at the pipeline's most expensive stage on every run; if it is not, the fidelity bill should show it. Fence $70, spend about $37.
A Pinned Instance and Two Stop Rules
Unlike AS through AU, which ran on the shared dev service, AV pinned a dedicated worktree instance at v3.293.3 so both arms share one commit and the engine is set per process rather than by mutating a shared service; there is no version split of the AU kind anywhere in the pairing. Two registered stop rules ran before anything scored. The harness smoke revalidated the endpoint across replicator's SvelteKit 3 migration. The engine-plumbing check, which demands the two arms actually decide differently on the same items, was settled decisively by the match stage's latency fingerprint: fable deliberates for 4.2 to 16.8 seconds per page, Jev answers in 231 to 670 milliseconds. The fable baseline was bought fresh for about $20 rather than reusing frozen records, applying the record-reuse lesson stem-bench Study 5 taught the week before. One operational note is disclosed: external kills interrupted the jev arm twice mid-item; no scored record was affected, the items re-ran from scratch, and the stray staging folders were swept at teardown.
All Gates Pass, and the Default Flips
AV-G1: 12 of 12 items complete in each arm. AV-G2: zero fallback widgets in 140 builds. AV-G3, the gate: median paired structure delta (jev minus fable) of -0.014 over 16 pages, inside the registered -0.03 no-harm bound. Per the registered decision rule the recommendation is to switch the match stage's default engine to Jev, shipped as a normal versioned change with the env switch retained as rollback, exactly the apparatus AU left dormant. The hand-off to the replicator agent is committed alongside the results with the care items the texture suggests.
The Price of Permissiveness, Measured
AU's offline policies reproduced end to end: fable reused 17.0% of 106 regions and built 88 widgets; Jev reused 50.9% and built 52, which is 40.9% fewer trips through the pipeline's most expensive stage. Match wall-clock fell from 161 seconds to 12 corpus-wide, total pipeline time from 42.7 to 32.3 minutes, and the stage's model cost from about $1.50 to about $0.003 per pass. The downstream price of all that permissiveness is the gate number itself: a median 0.014 of structure, measured, inside the bound.
The Texture, and a Wrong Prediction
The registered prediction said divergence would concentrate in T2, the composition tier whose heading-class regions drove AU's disagreement anatomy. It did not, and the inversion is the study's most instructive line: T2 absorbed stretched matches best (median delta +0.008, and t2-agenda is the corpus's one positive divergence, +0.068 with 9 of 10 regions reused), while the simple T1 pages paid most (tier median -0.042, with t1-memo at -0.111 the one page clearly past the bound on its own, and t3-case-study at -0.078 behind it). A loose match hurts exactly where a page has few regions to hide it. The scorecard: prediction 1 right, 2 half-right (the build reduction in band, but the match speedup is 13 times end to end, not the 45 extrapolated from AU's offline timing), 3 wrong as described, 4 right (the dark page paired normally at the shared commit).
What Stays Unfunded, and a Program Milestone
The conditional routed arm stays unfunded per its own registration: with the gate green, routing is moot, though t1-memo's -0.111 is the exact cell a future region-count or confidence threshold would target, and the hand-off records that as a care item rather than a requirement. One housekeeping line for the program ledger: with AV's 32 scored pages, the research program's scored-run count crosses forty thousand.