Ghost calibration MEASURED on hardware: straggler 1.462x → 1.189x

docs/BUILDS.md

@@ -146,8 +146,28 @@ backend arm's derived baseline is 1.469 (§2, and the value `scripts/gpu-benchma
 MEASURED default-constants straggler is **1.438** (§1). The regression is real on any of the three —
 they are quoted here so a reader is not left reconciling two numbers for one claim. The likely cause is rescaling Ghost's Schnorr constant by the old
 13.77x ECDSA ratio, when the hint removes decompression from *both* curves — taproot key-path spends
-call `lift_x` too — so that ratio should have narrowed. **Ghost needs its own calibration; it does
-not have one.**
+call `lift_x` too — so that ratio should have narrowed.
+
+✅ **Ghost DOES now have its own calibration — measured 2026-09-05, two L40S, real GPU proving on
+block 962,000.** The predicted narrowing is confirmed: the fitted Schnorr:ECDSA ratio is **1.97x**,
+not the old 13.77x.
+
+```
+HAZYNC_COST_EC_OP=85636  HAZYNC_COST_SCHNORR_OP=168542
+HAZYNC_COST_INPUT_BYTE=2 HAZYNC_COST_INPUT_BASE=53162
+```
+
+| arm | straggler | max | mean |
+|---|---|---|---|
+| shipped defaults | 1.462x | 149 s | 102 s |
+| Ghost refit | **1.189x** | 124 s | 104 s |
+
+⚠ These are GHOST's constants and they are NOT Core's (§2). Neither set transfers: Core's refit
+moves Ghost 1.407 → 1.884, about a card. Each channel carries its own fit, and a build that takes
+the wrong one is worse off than one that takes none.
+
+⚠ The card arithmetic was not re-derived — read this as "the ≤1.35 straggler target of §4.2 is
+met", not as a card count.
 
 ✅ **#139 and the field backend are orthogonal and stack**: 10.676x → 11.467x, +7.4%. They accelerate
 different things — #139 replaces `secp256k1_ecmult`; the field backend replaces the representation

docs/GHOST_NEXT_BUILD.md

@@ -53,6 +53,48 @@ The packer balances its own model perfectly; that is what makes a wrong model in
 and block 741000 carries **2 Schnorr verifies out of 723** — so it cannot inform the curve split at
 all, and the curve split is the dimension Ghost most needs right.
 
+## ✅ MEASURED ON HARDWARE 2026-09-05 — the refit is worth 1.462x → 1.189x
+
+Two L40S, block 962,000, 16 chunks, **real GPU proving**. Same guest, same block, one box per arm,
+differing only in the four `HAZYNC_COST_*` values:
+
+| arm | straggler | max | mean |
+|---|---|---|---|
+| shipped defaults | **1.462x** | 149 s | 102 s |
+| refit | **1.189x** | 124 s | 104 s |
+
+That clears the ≤1.35 threshold §4.2 of `BUILDS.md` states for the fourth card. ⚠ I have not
+re-derived the card arithmetic itself, so read this as "the straggler target is met", not as a
+card count.
+
+Refit constants, fitted on block 965,500 (7.7% Schnorr), intercept forced to zero:
+
+```
+HAZYNC_COST_EC_OP=85636  HAZYNC_COST_SCHNORR_OP=168542
+HAZYNC_COST_INPUT_BYTE=2 HAZYNC_COST_INPUT_BASE=53162      Schnorr:ECDSA = 1.97x
+```
+
+### ⛔ How this nearly came out backwards
+
+The run first measured **1.567x — WORSE than doing nothing** — and that was reported. It was one
+chunk. Across 32 chunk timings on two boxes every measurement sits at **2.4–3.0 s/segment**; hz-b
+chunk 1 came in at **4.89 s/segment** (171 s for 35 segments). Re-run twice: **94 s, 94 s, 2.69
+s/segment** both times. The 171 s never reproduced.
+
+Three checks were needed to get there, and two of them refuted a hypothesis rather than confirming it:
+
+1. **The segment straggler said the opposite** — 57 → 49 max segments, i.e. the refit was *better*
+   by the machine-independent metric. That is what said the wall-clock number was wrong.
+2. **"The boxes differ in speed"** — plausible, and *false as an explanation*: an identical
+   33-segment chunk runs 87 s on hz-a and 90 s on hz-b, only 3.4% apart, **and a straggler is a
+   ratio, so box speed cancels out of it entirely.**
+3. **Wall time is proportional to segments** — 2.4–3.0 s/seg across 31 of 32 samples — which
+   isolated the 32nd as a transient rather than a property of the partition.
+
+⇒ A single un-replicated chunk timing was enough to invert the conclusion. Re-run an outlier before
+believing it, and prefer the quantised metric (segments) over the continuous one (cycles): proving
+bills in whole segments.
+
 ## The work, in order
 
 1. **Profile the Ghost arm on a taproot-bearing block.** `HAZYNC_PROFILE_EXEC=1`, 16 chunks, execute