Levandor

Fill-Structuring Space Closed + 15m Pivot Program (PROPOSED)

11 min read

Solution-design session (2026-07-18) that (a) formally closes the 5m τ∈[60,150)s fill-structuring space by consolidating every prior fill-lever NO-GO under one named mechanism (structural last-look), and (b) proposes a W0–W3 program that pivots the search for a capturable edge to a 15m offline gate, built for the first time on an honest fill harness rather than the certain-fill assumption that produced “offline +EV, live −EV.” Everything in the program section is a proposal.

SUPERSEDED (2026-07-21) — the 15m pivot PREMISE was FALSIFIED

A grounding pass on the approved re-site spec falsified this proposal’s primary bet before any build: on true CLOB labels every 15m/slow-clock cushion pocket is net-negative, and the “near-breakeven” read was on-chain label error. The pre-committed exit rule below fired; recommendation is to shelve the trading lane (keep observe recording). See crypto-shortterm-15m-resite-premise-falsified-2026-07-21 for the verdict. The SETTLED part of this note (the 5m fill-structuring space is closed) still stands; the PROPOSED W2/W3/15m-gate program is now closed out as falsified.

STATUS — PROPOSED, PENDING ANDRAS'S DECISION

No decision has been taken. No code, config, or live state was changed. The W0–W3 program, the pre-committed exit rule, and the rejected-alternatives list below are all a recommendation awaiting Andras’s go/no-go. All live lanes remain dry_run (canary [[crypto-shortterm-canary-confirmed-fill-ev-live-2026-07-16|#1210 aftermath]] confirmed closed, sources back to dry). The only settled part of this note is the consolidated conclusion that the 5m fill-structuring space is exhausted — that is a read of prior evidence, not a new action.

For Agents

Two separable claims live here. (1) SETTLED (evidence consolidation): the 5m [60,150)s taker fill-structuring space is closed — do not propose another taker bid-structuring study on this band without new evidence or a materially faster execution path. Mechanism = structural last-look. (2) PROPOSED (pending decision): the W0–W3 program (taker-rebate check → honest-fill harness → 15m offline gate → 5m slow-band falsification) and its pre-committed shelve rule. Treat W0–W3 as not yet adopted. Numbers under “New verified facts” are from a read-only verification pass on 2026-07-18 and are load-bearing for the plan.

The Settled Conclusion — the 5m fill-structuring space is formally closed

Every taker-side lever that could plausibly recover the fill adverse-selection loss on the validated 5m τ∈[60,150)s band has now been examined with rigor and come back negative or, at best, shave-to-breakeven. This note consolidates them into a single closed set:

LeverVerdictSource
A0 first-touch avoidanceNO-GO — telemetry/regime-confounded (A0 data only exists on the 3 drawdown days)crypto-shortterm-a0-fill-policy-offline-scope-nogo-2026-07-12
Exit / loss-cut (×2 independent runs)NO-GO — exiting is significantly −EV vs holding (collapsing bid + weak loser-lead)crypto-shortterm-exit-loss-cut-study-2026-07-01
Chase / raise the limit (edge-band ceiling)NO-GO — chasing the reprice up is −EV; break-even limit correctly refusescrypto-shortterm-disagreement-ceiling-nogo-2026-07-03
Market / FOK / resting GTC-maker order typeNO-GO — buys the above-BE −EV pocket / hands a free option to sub-second botscrypto-shortterm-order-type-fak-verdict-2026-07-09
Co-location / further producer-side latency spendShelved OOS — does not clear the bar; ask-run is front-loadedcrypto-shortterm-phase1-execution-decision, crypto-shortterm-discriminator-catchability-ask-drift-2026-07-13
Depth-aware sizing (size UP)NO-GO — a size-up, not a de-risk; no per-share edgecrypto-shortterm-depth-aware-sizing-nogo-2026-07-10
Size-down / depth-selectivity (size DOWN)NO-GO — it’s exogenous book depth, not our size; selectivity only shaves 3/6 majors to breakevencrypto-shortterm-size-depth-study-liquidity-not-size-nogo-2026-07-17
Honest recalibrationNO-GO for EV — signal healthy, the gate already absorbs the miscalibration at the ~0.72 askcrypto-shortterm-honest-recalibration-edge-gate-verdict-2026-07-11
EMIT_MAX_ASK tighteningNot supported by the fill data as an EV recoverycrypto-shortterm-discriminator-catchability-ask-drift-2026-07-13
Time-of-day overlayNO-GO — in-sample great, OOS sign-reversedcrypto-shortterm-time-of-day-filter-nogo-2026-07-06

VERDICT — no further taker bid-structuring studies on the 5m [60,150)s band

The set above is complete for taker fill-structuring at current execution. The live canary confirmed the gap is real (signal would-win 83.8% vs fill-win 70.2% = 13.6pt adverse selection), and the ask-drift study showed the missed winners are not physically catchable (the ~7-tick ask lift completes within the emit→first-touch latency window and plateaus). The remaining candidates are all breakeven-ceilinged or overfit. Stop searching for a taker bid-structure fix on this band.

Mechanism named — structural last-look

The unifying explanation for why every lever caps at breakeven: our signal is public-information, and the venue’s protective taker delay lets sub-second makers reprice it before our order matches.

  • Our fair-value signal is derived from public feeds (Binance trades + on-chain Chainlink). Any maker watching the same feeds sees what we see.
  • Polymarket imposes a mandated ~250ms taker protective delay (last-look window). Within that window, makers pull/reprice the ask above our limit — this is the front-loaded ask-run measured directly (99.9% of no-match rows had the real ask repriced above our limit, median +0.07).
  • The decisive number: inside the clean 07-09→11 window, the A1+ reprice fills sit at ≈ −0.5¢ (breakeven) (A0 scope study). That is the price we get after the last-look reprice. Even infinite producer-side speed only gets us to that repriced ask — so the ceiling is breakeven regardless of how fast we fire.

Why this closes the space rather than just adding a NO-GO

“Structural last-look” makes the breakeven ceiling a property of trading public info as a taker inside a protected book, not a property of any one lever. That is why speed (co-lo), order type (FOK/GTC), and clip shape (size up/down) all independently bottom out at the same place. Escaping it requires changing one of the structural inputs: a non-public / faster-than-last-look signal, a different microstructure (a book where we are not the picked-off taker), or a different timeframe where the reprice is slower relative to our edge. The program below bets on the third.

New verified facts (read-only verification pass, 2026-07-18)

Four facts established this session by a read-only verification agent; each is load-bearing for the proposed program.

1. An offline 15m gate is runnable NOW

btc15/eth15 observe recording has been running since 2026-07-02 19:45Z (correcting the prior “since 07-09” belief) — materially more history than assumed:

  • 1,490 windows per asset, ~96/day, complete.
  • Outcomes 100% settled (1490/1490 on both assets).
  • Both token books ≥94% populated.

→ an offline 15m gate can be run today. Caveat: thin — only 2 assets and roughly 2.7-day purged folds, so the deflated bar and per-asset discipline matter.

2. The 5m slow-clock band τ∈[150,300)s is population-sufficient but ask-expensive

  • Population is fine: ~3,311 windows over 14 days pooled.
  • But the median first-qualifying ask is 0.90–0.93 — squarely the documented −EV pocket. Only the thin [210,300) sub-band reaches 0.86–0.90.
  • Conviction and ask co-move on this band: by the time the signal is confident, the market has already priced it. → this leg is downgraded from “candidate edge” to cheap falsification (run it because it’s a one-line filter, expect it to fail).

3. The honest-fill harness was never built — this is the root cause of “offline +EV, live −EV”

The 2026-07-03 backtest-fill-harness spec was never implemented. The current walkforward gate assumes:

Those three assumptions are exactly the gap between the offline gate passing and the live lane losing. Harness dependencies (refitlib/exitlib) already sit on branch feat/calibration-refit; the fill ground truth (pm_fak_attempts, pmv2_autotrade_orders) requires a cross-schema read.

4. Emit’s 15m block is a one-line compiled guard

emit_supported (crates/observe/src/bin/emit.rs:86-93) is a compiled allowlist that refuses any non-300s asset. Arming 15m is a one-line change — but it must be gated on a validated 15m calibration + regime, never before (the same discipline that kept bnb observe-only by construction).

PROPOSED program — W0–W3 + pre-committed exit rule

The whole of this section is PROPOSED, pending Andras's decision.

W0 — taker-rebate check (free, do first). Confirm whether our own wallet accrues Polymarket taker rebates (~16% claw-back → 1.72% effective fee). Named twice before (fee study, API gotchas) and never run. Zero build, potential free EV, and it re-bases every fee term in the honest harness.

W1 — build the honest-fill harness per the 2026-07-03 spec. A fill model calibrated on our ~1,400 recorded live fills, self-validated on held-out fills, using won_up (real CLOB) labels only — never resolved_up. Latency parameters come from the positionmanager exec-latency decomposition (babylon #1189, below). This harness is the shared measurement instrument for W2 and W3; without it, any 15m gate inherits the same certain-fill lie that broke the 5m lane.

W2 — PRIMARY: the 15m offline gate.

  • Fresh fit-then-frozen 15m calibration on days 1–5 of the btc15/eth15 history.
  • Its own regime band and ask cap — do NOT inherit the 5m τ∈[60,150)s / EMIT_MAX_ASK values. The 15m microstructure is different (the early 15m NO-GO found the ask ≈ win-rate across the whole τ curve; the band must be re-derived, not copied).
  • ONE confirmatory test on the remaining days (minimal-trials discipline — the deflated bar punishes a grid search on 2.7-day folds).
  • Honest fills at measured latency, per-asset, deflated bar.
  • Existence proof it must beat: the competitor’s btc-15m late band booked +3.8¢/share, CI [+2.2, +5.4], at HIS fills — and he is a ~1s taker like us (competitor dissection). So a capturable 15m edge at our speed class is not physically impossible.
  • Refutation it must overcome: the same study found the mechanical trigger on the population was −1.1¢ — his edge was side/timing skill, not the raw band. Our candidate replacement for that skill is p_cal (which agreed with his side 94% of the time on 5m). W2 is the test of whether p_cal substitutes for his discretionary skill well enough to clear the bar under honest fills.

W3 — 5m τ∈[150,300)s falsification through the same harness (a one-line filter on the existing pipeline). Expectation is low given fact #2 (ask co-moves with conviction), but it is nearly free to run and closes the slow-clock question cleanly.

Pre-committed exit rule (part of the proposal)

If W2 and W3 both fail the bar under honest fills → shelve the trading lane. The observe containers keep running as a data asset (they cost nothing and bank the history). The two pre-registered 5m re-tests (green-regime A0 telemetry; genuine-erosion frozen-gated fill-win <73% over ≥5 days) stay dormant unless PM ships a materially faster execution path or the venue structure changes. Committing to the shelve condition before running W2/W3 is what keeps this from becoming another post-hoc shave-stack.

Alternatives considered and rejected

  • Canary #2 now — a second bounded −EV live canary has breakeven-bounded upside on the closed 5m band; it teaches little the first canary didn’t. It can instead ride along a future 15m canary once W2 passes, at no extra standalone cost.
  • Maker / market-maker flip — we have no speed and no information edge vs the incumbent makers (our signal is public info). Rejected already on 2026-07-09 and 2026-07-09 expansion scan; as a maker on this band we would BE the stale quote the last-look bots pick off.
  • Further latency spend — the ask-run is front-loaded and the ceiling is last-look, so speed is non-binding (proven by the ask-drift study). More co-lo money buys a breakeven fill faster.
  • Shave-stack re-arm — re-arming 5m with a stack of the surviving marginal filters (depth-selectivity + EMIT_MAX_ASK tighten + overnight) is a breakeven ceiling built from post-hoc filters = overfit (the “market prices every overlay” lesson).
  • Longshot-fade — no free lunch here: the documented near-settlement bias is favorite-UNDERpricing, not longshot overpricing, so a fade trades the wrong way.
  • HYPE / a 7th asset now — adding an asset multiplies the broken fill economics rather than fixing them; HYPE stays on the passive watchlist (expansion scan) until the fill problem is solved on the existing lanes.

Babylon coordination

  • Open task #1189 (positionmanager exec-latency decomposition, READY @ 45fa0eb) is noted as the latency-parameter source for W1’s honest-fill harness. Reply is deferred until direction is set (no point wiring latency params into a harness that may not be built).
  • Canary #1210 aftermath confirmed closed — kill confirmed, sources back to dry_run. No live exposure.

Graph View

Backlinks