Levandor

Maker-Sim Gate — Spec Grounding Round (v2, pending review)

4 min read

Grounding round for the maker-sim gate design spec (docs/superpowers/specs/2026-07-22-maker-sim-gate-design.md, v2) — the proposed offline study asking whether a maker (resting-order) fill posture could recover the edge that taker fills lose to last-look adverse selection. Spec v2 is written and pending Andras review; nothing is built. This note captures the durable data facts and spec corrections the grounding pass produced, so a future study does not re-derive them.

For Agents

This is spec-grounding, not a green light — the lane is still shelved, and maker-sim is a candidate direction under review. The load-bearing outputs here are (a) four verified 5m data facts you must respect in any 5m study, and (b) five corrections made visible in spec v2. Follows the Andras spec-grounding standard: primary-source verification with provenance, retract errors visibly.

Verified data facts (respect these in any 5m study)

1. 5m near-tie rate is 53% → on-chain resolved_up is unusable as a 5m scoring label

  • 53.25% of btc 5m windows settle within ±5bps of strike; 78.39% within ±10bps (n = 6,043 windows, 21 days). Compare 15m: only 22–31% near-tie.
  • At 5m the settlement is a coin-flip so often that the on-chain resolved_up label carries near-tie noise that swamps any edge signal. CLOB tokens[].winner backfill is mandatory for ALL scored windows in any 5m study — do not score 5m windows on resolved_up.
  • This sharpens the standing on-chain-label-overstates-win-rate caveat with a measured 5m magnitude.

2. Gamma slug/condition_id query is BROKEN for 5m updown markets

gamma /markets?slug= and ?condition_ids= both return [] for 5m updown

Works for 15m markets, silently empty for 5m. Do not build a 5m study on gamma slug resolution.

  • Resolution path that works: get condition_id from emitted_signals.condition_id for fired windows → CLOB /markets/{condition_id} returns the market and confirms market_slug = btc-updown-5m-{window_id}.

3. data-api /trades?market={conditionId} verified as the 5m tape source

  • Returns the full trade tape for resolved markets. Paging limit/offset ≈ 2k rows/page; timestamps are unix seconds; taker-side only (no maker-role flag on prints).
  • A sampled 5m window had 4,786 prints. Tape backfill budget for ~7.5k windows ≈ 15–35M rows.
  • The taker-only limitation matters: a maker-sim must infer resting-fill outcomes from the taker tape + book snaps, not from a maker flag that does not exist in the data.

4. pm_book_snaps 5m cadence ≈ 1.3s

  • Measured ≈1.3s between snaps (~228 snaps/window), sizes 100% populated. Btc 5m recording spans 06-20 → present, 8,259 windows, complete.
  • 1.3s is the finest real observation grain — it bounds any sim latency model (see correction 5).

Spec corrections made visible in v2

Per the grounding standard, v1 errors were retracted in-place in v2, not silently patched:

  1. 7-tick / +0.07 reprice figure — the sanity anchor for how fast the ask runs away lives in the A0 scope doc (07-13 follow-up / crypto-shortterm-discriminator-catchability-ask-drift-2026-07-13), NOT in latency_budget.rs. v1 mis-cited the source.
  2. Maker-rewards citation was wrong — the competitor’s rebates were TAKER rebates (W0 audit confirms our own maker-rebate = $0, pure takers). Maker-rewards applicability to 5m markets is unverified and must not be assumed as a maker-sim income term.
  3. Sim latency tiers changed to cadence multiples {1.3, 2.6, 3.9}s — v1’s sub-second tiers were unsupported: the 1.3s snap cadence (fact 4) is the observation floor, so no sub-second tier is measurable from this data.

Babylon coordination state (as of 2026-07-22)

  • PM answered #1189: sub-500ms fill latency is feasible after fee-param caching; sub-300ms needs a WS book feed.
  • PM built the fee fix fix/venue-cash-fee-coefficient @ eee211e: books the venue fee at 0.07, removes 2 critical-path fee RTTs, and closes a zero-fee fail-soft. Based on PROD ca07ba5, not main.
  • crypto asked PM to independently verify 0.07 on their own cash before rolling — babylon #1239.
  • The 5,322-row historical fee backfill is gated on Andras (re-basing pmv2’s over-booked fees, per the W0 fee-booking bug, #1233).

Provenance

  • 5m near-tie stats, book-snap cadence, and window coverage: read-only queries against the crypto_shortterm schema / pm_book_snaps.
  • Market-resolution path and tape: emitted_signals.condition_id → CLOB /markets/{condition_id} and data-api.polymarket.com /trades?market={conditionId}.
  • Spec: docs/superpowers/specs/2026-07-22-maker-sim-gate-design.md (v2).

Graph View

Backlinks