vyndr/specs/collapsed-sequence-edge.md

The collapsed sequence edge — two proven links whose product is too small to use

Both components are real. Their product is 0.37pp, and detecting it would take 52 seasons. This is the most instructive negative in the programme so far, because nothing in it failed: link-by-link proof did not produce a usable edge.

---

One correction to the framing

The order states Link 2 proved you can't predict the reliever. Half true, and the other half matters: the individual grain failed (17.2% accuracy), but the quality grain PROVED — predicted pen quality separates 2.70pp of realized hit rate. Pen-season-quality here is a measured predictor, not a fallback after a failure. That strengthened the plan going in.

Two of the three specified inputs could not be used honestly

specified	status
pen quality	PROVEN (Link 2 coarse grain) — used
pen archetype	did NOT prove (0.5669 vs 0.5309 modal baseline, corrected interval spanning zero) — excluded, building it in would chain on an unproven link
hitter approach identity ("fastball-hunter", "finesse-vulnerable")	does not exist in this registry. MLB batter archetypes are BOMBER / GHOST / TORCH / BRUSH / DRIVER / FLEX / ALPHA / HYBRID / CATALYST. Inventing an identity to condition on is the fabrication the gate exists to catch

A hitter power/contact split derived from the sequence data itself was tested as a separate gated addition rather than assumed into the main effect. Neither half proved (power −0.0002, contact −0.0001, both intervals spanning zero).

---

The gate — two-part, on the concentrated subset, 114 cumulative tests

subset	n	games	mean shift	Brier Δ	CI	verdict
concentrated (early-exit × WEAK pen)	1,931	141	0.0193	−0.0001	[−0.0014, +0.0010]	NOT_PROVEN
mirror (early-exit × STRONG pen)	2,574	189	0.0186	0.0000	[−0.0011, +0.0010]	THEATER
all early-exit later ABs	6,869	451	0.0140	−0.0001	[−0.0007, +0.0005]	NOT_PROVEN
pooled all later ABs	17,891	803	0.0141	0.0000	[−0.0004, +0.0003]	THEATER

Not pooled-diluted: the concentrated subset was gated on its own and is no better. Two subsets are THEATER by the gate's own definition — the adjustment moves the number ~1.9pp and improves accuracy by essentially nothing.

---

Why: the mechanical ceiling

The descriptive pass found the direction the order predicted (early-exit + weak pen +0.74pp, early-exit + strong pen −0.79pp vs a deep-starter baseline). The signs are right. The magnitude is the problem, and it is structural:

P(faces pen | early-exit flagged)  0.8075
P(faces pen | starter goes deep)   0.7149
  exposure the flag actually buys  0.0925      <- NOT a switch to the pen

hit-rate swing across pen quality  0.0394      (weak 0.2491 vs strong 0.2038)

MAX JUSTIFIABLE ADJUSTMENT = 0.0925 x 0.0394 = 0.00365   (0.37pp)
adjustment actually applied (mean |shift|)   = 0.01930   (1.93pp)
OVER-MOVEMENT FACTOR                          = 5.3x

A hitter's third or fourth plate appearance is ALREADY against the bullpen 71% of the time even when the starter is projected to go deep. Link 1 lifts that to 81%. It buys nine points of extra pen exposure, not a change of opponent — so any adjustment riding on it is capped at about a tenth of the pen-quality swing.

The 5.3× over-movement is exactly why the mirror subset reads as THEATER rather than as a small true effect: the adjustment asserts five times more than the mechanism can support.

And a correctly-scaled version is not detectable either

concentrated subset n = 1,931   SE of hit rate = 0.00985
max justifiable effect 0.00365 = 0.37 SE
to detect at the corrected bar (z~3.46 for 114 tests): n = 168,488
shortfall 87x  ->  ~52 seasons of concentrated-subset accrual

This line is structurally closed, not sample-blocked. Waiting does not fix it.

---

Not wired, and the self-check deliberately not wired either

The adjustment does not prove, so it feeds nothing. The order also asks for a self-check flagging where our sequence read diverges from the line's starter-script, as an opportunity signal. That is not wired, because flagging divergence on an adjustment measured as absent would advertise an edge we have just shown does not exist — the same failure as fabricated reasoning, one layer up.

The lesson worth keeping

Link 1 proved (MAE 3.22 → 2.80 batters faced). Link 2's quality grain proved (2.70pp of realized separation). Both are real, both are point-in-time, both survived cumulative correction. Their product is still too small to use.

Link-by-link validation guarantees each link is real. It does not guarantee the chain transmits anything. The multiplicative structure has to be sized BEFORE building — one exposure term of 0.09 is enough to reduce a genuine 3.94pp signal to noise, and no amount of downstream care recovers it.

Parallel track — total_bases per-archetype (logged, not run)

Unchanged: total_bases settled n=948 pooled, BOMBER × TB 340, short by 160. Sample-readiness only, not a verdict. The specs/per-archetype-grade-bands.md blocker still stands — the grade does not yet separate within any archetype.

Link 3 confirmed SKIPPED. Counter and frozen clusters byte-identical.