Levandor

Discriminator Re-Verification + Ask-Drift Catchability

4 min read

Two-part 2026-07-13 follow-up to the A0 offline-scope NO-GO: the same per-asset/per-regime scrutiny that dissolved the A0 story was re-run on the underlying unfilled-would-win discriminator — and this time it HOLDS (opposite outcome to A0). The missed winners are genuinely +EV at the quoted ask, BUT they are largely not physically catchable: the ask runs ~7 ticks near-instantly after emit, so no cheap fill-capture lever adds EV. Source: docs/superpowers/a0-fill-policy-offline-scope-2026-07-12.md (Follow-up 2026-07-13 section).

Bottom line — the prize is real but not catchable

The +EV lives in the winners we fail to fill, and at their quoted ask those winners clear breakeven handsomely (+17.6pt at ask 0.5, +10.4pt at 0.7). But the winning-side ask completes a ~7-tick lift during the emit→first-touch latency window and plateaus — the catchable fills are the stale-ask losers (the A0 set), while the +EV fills require filling at the emit ask before a near-instant reprice, i.e. a sub-latency race. Chasing the +7 ticks ≈ breakeven at a 0.72 ask, negative above 0.78 — which is exactly why the break-even limit correctly refuses. No cheap fill-capture lever adds EV. This independently re-confirms co-lo on EV grounds.

Part 1 — the discriminator is VALIDATED (unlike A0)

Data: data/signal_fill_decomposition_2026-07-10.csv, 5,555 fires, 07-04→07-10, real_win = CLOB truth (validated 100% vs the winner CSVs). Overall: unfilled fires win 80.6% vs filled 73.4% — a ~7.2pt gap (pooled ≈ 6 SE on 3,484 unfilled vs 2,066 filled).

Where A0 flipped sign across days and concentrated in thin alts, this gap survives every cut:

  • Per-asset — positive in ALL 6: btc 9.6, doge 9.3, eth 8.9, bnb 8.9, xrp 8.2, sol 4.0.
  • Per-day (regime) — positive on ALL 7 days, green and red, 5.7–11.3pt, never near zero. A0 went +0.4 → −7.2 across days; this does not. Structural, not a drawdown artifact.
  • Conditioned on conviction — survives: 7.5pt gap within the dominant p_cal stratum (n=5,418). So it is adverse selection on the unobservable, not merely “unfilled fires were higher-conviction.”

This is the rigorous version of the standing FILL adverse-selection verdict: the +EV genuinely lives in the unfilled tail.

The missed winners are +EV-capturable AT THE QUOTED ASK

At each ask bin the missed winners clear their own breakeven — the prize is real and biggest at cheap asks (<0.75):

emit askmissed-winner edge over breakeven
0.5+17.6pt
0.6+14.2pt
0.7+10.4pt
0.8+5.0pt
0.9−4.6pt (the >0.90 pocket EMIT_MAX_ASK already blocks)

Part 2 — but ask-drift kills the speed lever

Turning to whether those winners are catchable: pm_fak_attempts.best_ask_at_snap (the real ask; note the attempt-0 site was still the req.limit proxy until the 45fa0eb/2f0ec6ad fix half-landed mid-07-11, but attempts 1/2 logged the real ask since 07-09). On the real-ask no_match rows (n≈1,810 reprice attempts, 07-09→11):

  • 99.9% of unfilled reprice attempts had the real ask ABOVE our limit — not phantom liquidity, not a lost race: the price genuinely repriced up.
  • The real ask sat median +0.03–0.04 above our laddered limit, median +0.07 above the emit ask (p90 +0.10–0.11). would_win 83% — these are the missed winners.
  • The run is FRONT-LOADED: drift-vs-emit is identical at attempt 1 and attempt 2 (p50 0.07 both) — the ask completes its ~7-tick lift by the first reprice and plateaus. Only ~20% still within 2 ticks of emit. The lift happens during the emit→first-touch latency window, not during the FAK ladder.

Consequence — no cheap fill-capture lever adds EV

  • The catchable fills = stale-ask losers (the A0 set — the fills we already get).
  • The +EV fills (missed winners) require filling at the emit ask before a near-instant ~7-tick lift — a sub-latency race (residual RTT 117–260ms; co-lo shelved as not clearing the OOS bar).
  • A tighter FAK ladder can’t help — the run is already done by attempt 1.
  • Chasing the +7 ticks ≈ breakeven at a 0.72 ask, negative above 0.78 — why the break-even limit correctly refuses.

Honest fill levers are quality-not-quantity, not speed:

  • EMIT_MAX_ASK 0.85→0.78–0.80 (trim the thin near-breakeven tail off the >0.90 side).
  • Refuse the thin-book toxic pocket (doge/xrp).

This re-confirms PM shelving latency/co-lo on EV grounds — independently, from the producer side.

Caveats

  • Clean attempt-0 first-touch drift needs post-2f0ec6ad real best_ask_at_snap (accrue forward). The current inference rests on the reprice-plateau (attempts 1/2), not a clean attempt-0 measurement.
  • Real-ask data is red-days-only (07-09+), though drift is microstructural and less regime-dependent than P&L.
  • The decomposition CSV ends 07-10, so 07-11 attempt-0 rows don’t join to labels.

Graph View

Backlinks