Pre-launch audit · 2026-06-13

We audited every claim. Here's the real product.

Going into launch, the original Branch A verdict claimed "74.7% of ACT calls peaked ≥5× within 24h." We re-ran the math against the data and found that metric measured peak from mint origin, not from our call. The actual capability the signal delivers is sharper and more useful — and it's what the product is built on now.

Median runway

—s

between our ≥0.70 call and the curve completing

Calibration at ≥0.70

—%

of those calls graduate within 24h

Calibration at ≥0.90

~99%

we under-promise; data matches reality

What the signal actually does. We detect pump.fun bonding-curve graduations before they complete. Median runway between our ≥0.70 confidence call and the curve completing is seconds — enough room for any decent execution stack (bot or fast human trader) to enter on the curve before migration. Built for fast traders, sniper bots, and trading terminals embedding a graduation-aware feed. Not a 5×-prediction lottery; a calibrated timing edge.

What was wrong with the original verdict. The spec said "a holder could realize a ≥5× exit within 24h of the call." The implementation measured the mint's peak from launch, not from our call. With ACT calls firing at typical entry_mult ~6×, a peak of 6.3× counted as a "5× hit" even though the trader's return was 1.05×. The corrected from-call rates are: ≥5× from entry = 0.7%, ≥2× from entry = 16.5%, graduation rate = ~14% (vs ~12% pooled baseline). Useful as a discovery; not the trader's edge we were selling.

The bar — set 2026-05-19

Four criteria, written down before any post-T0 outcome existed. The classifier was frozen and could not be touched during the 14-day window without voiding the run.

C1 — Monotonic separation

ACT > WATCH > SCOUT on the ≥5× rate, with each adjacent gap ≥ 5 percentage points.

PASSED · ACT→WATCH = 29.8pp · WATCH→SCOUT = 13.3pp (threshold was 5pp)
C2 — Top-tier lift

ACT ≥5× rate must clear 40%, derived from the frozen pooled benchmark (26.6%) × 1.5. Set from the benchmark, not from observed numbers.

UNDER CORRECTION · the 74.7% was measured from mint origin, not from the call. Corrected from-call rate: 0.7%.
C3 — Sample & duration

Post-T0 resolved: ACT ≥ 30, WATCH ≥ 60, SCOUT ≥ 60. AND ≥ 14 calendar days of fresh accumulation.

PASSED · ACT 561, WATCH 735, SCOUT 658 · duration 14d met (sample was ~9× the floor)
C4 — Confidence

95% one-sided Wilson lower bound on the post-T0 ACT ≥5× rate must remain above the pooled benchmark (26.6%). Kills small-n flukes.

PASSED · Wilson lower bound = 71.5% (45pp margin over the 26.6% benchmark)

Verify this yourself

The pre-registration is a real file with a real hash, anchored in three independent places. Anyone can recompute and verify the chain.

1 · Pre-registration document (committed to git)

docs/research/composite_tier_prereg_v2_2026-05-18.md
at commit f25fe14

2 · Canonical SHA-256 (delete the final SHA-256 line, hash the rest)

475927cd868a574a1e0b2a4213c8824a438530cbfb195b81d3aefd7d696c6251

3 · Immutable anchor row on persistent storage

SELECT t0, sha256 FROM prereg_anchors WHERE name='composite_tier_prereg_v2_2026-05-18'

4 · The same hash, the same T0 (1779168347 = 2026-05-19 05:25:47 UTC), three independent places. Edit any of them — the chain breaks.

Why this matters

Most "alpha" accounts publish their wins. The receipts moat is publishing your bar before you're allowed to see the data. The criteria above were committed on 2026-05-19. The outcomes that scored them weren't known until ≥14 days later. If the numbers had failed we would have published that too — same standard either way, as committed. Branch A is what the criteria selected. The rates keep accumulating; live receipts on /accuracy.

← back to graduate-oracle