Levandor

15m Re-Site Premise FALSIFIED — Exit Rule Fired, Shelve Recommended

7 min read

Grounding-pass verdict (2026-07-20/21): Andras asked for a grounding pass on the approved slow-clock 15m re-site spec (“make sure it will actually produce reliable data”). Three read-only passes falsified the program’s core premise before any code was built — every 15m/slow-clock cushion pocket is net-negative on true CLOB labels, the near-breakeven signal that motivated the pivot was a label-error artifact, and the pre-committed exit rule fired. Recommendation: shelve the trading lane (keep observe recording as a data asset), pending Andras’s decision.

VERDICT — 15m re-site premise FALSIFIED; exit rule FIRED

The 2026-07-18 W0–W3 pivot proposal’s primary bet (a capturable 15m offline edge) does not survive contact with true CLOB labels. This note SUPERSEDES the conclusion of that proposal: W2 (the 15m offline gate) and the slow-clock band are both dead pockets on our own books; the “near-breakeven” illusion was on-chain label error. The pre-committed shelve rule is invoked. No live state changed; observe (5m+15m) keeps recording.

For Agents

SETTLED: the 15m re-site / slow-clock trading premise is falsified — do NOT propose another 15m or τ∈[150,300)s taker gate on the current venue/execution without a structural reopener (below). The honest-fill harness was NOT built (no consumer survived the falsification). Method findings are durable and load-bearing: (1) the 15m label trap — near-ties are 22–31% of late-window populations, the outcomes store has no condition_id, so a CLOB backfill is mandatory for any 15m scoring; (2) the 07-09 fill-regime break — pm_fak_attempts lives in pmv2’s schema (the 07-03 harness spec’s crypto_shortterm qualifier is wrong). All numbers below are from read-only SQL (via pgq.sh) + Polymarket gamma/CLOB pulls, 2026-07-20/21.

The three grounding passes

Andras’s request was narrow — verify the approved re-site would produce reliable data — but each pass came back falsifying the premise itself, in escalating severity.

Pass A — the conviction gate saturates 15m into a dead ask zone

Our 5m conviction gate |p_fair − 0.5| ≥ 0.30 ∧ τ ≤ 300s saturates on 15m: ~98% of the 1,733 windows/asset qualify (vs the selective ~a-few-percent it picks on 5m). Worse, the first-qualifying asks are 0.96–0.97 — a dead zone even more expensive than the documented 5m slow-clock τ∈[150,300) trap (0.90–0.93, see [[crypto-shortterm-fill-space-closed-15m-pivot-proposal-2026-07-18|fact #2]]). By the time the signal is confident on 15m, the market has fully priced it. The gate as-designed cannot be inherited onto 15m (which the proposal already warned — “re-derive the band, don’t copy 5m”).

Pass B — the competitor’s cushion band IS priced at his-era levels on our books (the only non-dead pocket)

The one pocket that was NOT instantly dead: the competitor’s cushion band (cushion ∈ [5,10)bps, τ ≤ 300s). On our books it prices at roughly his-era levels — btc15 median ask 0.875 (n=878), eth15 median ask 0.800 (n=788) (his bucket sat ~0.820). This was the only survivable-looking entry pocket found in the entire W0–W3 exploration, and motivated the decisive Pass C.

Pass C (decisive) — every cushion bucket is net-negative on TRUE CLOB labels

A full true-CLOB label backfill of all 3,131 first-trigger windows (slug {asset}-updown-15m-{window_id} → conditionId → tokens[].winner; 0.03% miss) killed the pocket:

  • EVERY cushion bucket is net-negative on both assets.
  • Pooled, day-clustered own-ask EV = −1.14¢, 95% CI [−3.0, +0.5].
  • The calibration segment reads +1.6¢ in-sample but flips to −2.2¢ out-of-sample (n=1,209) — on both assets independently (not a single-lane artifact).

The near-breakeven illusion was LABEL ERROR, not a real edge

The “near-breakeven / almost-tradeable” read came from on-chain resolved_up disagreeing with the true CLOB winner on 18–23% of near-ties. True near-tie win-rate is 0.63–0.69 (~a coinflip), NOT the 0.77 the on-chain label implied. This is the same on-chain-optimistic label mask that has bitten every prior 5m study — here it manufactured a fake tradeable pocket on 15m. The competitor’s +3.8¢ tape prior was a selection ceiling; on our own books his pockets price to fee-negative.

Durable method findings (keep these regardless of the shelve)

1. The 15m label trap — CLOB backfill is mandatory for any 15m scoring

  • Near-ties are 22–31% of late-window populations on 15m — a much larger near-tie mass than 5m, so label error dominates any naive score.
  • The outcomes store records no condition_id → you cannot join it to the CLOB winner directly; a CLOB backfill (slug → conditionId → tokens[].winner) is mandatory for any 15m EV measurement.
  • Off the near-ties, CLOB vs sign(settle − k) agree 99.1% — the disagreement is entirely concentrated in the near-tie band, which is exactly where 15m fills land.

2. The 07-09 fill-regime break + schema correction

  • Fill rate broke upward in stages: ~4–7% → 13.6% → 27–35%. The 07-09 shift is a regime break, not a smooth trend — do not pool fill telemetry across it.
  • pm_fak_attempts spans 07-02 → 07-16, with best_ask_at_snap 100%-populated from 07-03, and it lives in pmv2’s schema — the [[crypto-shortterm-backtesting|2026-07-03 harness spec’s crypto_shortterm qualifier is WRONG]] (the fill ground truth requires a cross-schema read into pmv2, consistent with the proposal’s cross-schema note).

Corrections retracted visibly (in the rewritten spec)

Three claims from the earlier proposal/spec were wrong and are retracted in place:

  • “p_cal replicates his side skill” — the 94% figure is agreement with the mechanical side, not his discretionary skill. His +3pp side/timing skill is NOT reproducible by p_cal. (This was W2’s whole substitution bet — it fails.)
  • “eth15 as co-primary” — the competitor’s eth-15m book lost $22k; eth15 is not a co-primary existence proof.
  • “fill telemetry from 07-09” — the telemetry actually starts 07-02; 07-09 is a regime break, not the data start.

Outcome & recommendation

Recommendation (pending Andras) — SHELVE the trading lane

  • Honest-fill harness NOT built — the falsification removed its only consumer (there is no 15m gate to feed).
  • Exit rule fired per the pre-committed W2+W3 shelve condition. Recommendation: shelve the trading lane.
  • Keep observe (5m + 15m) recording as a data asset (costs nothing, banks the history for any future reopener).
  • Run W0 — the taker-rebate check (free). The rebate program is confirmed real (the competitor harvested +$14.4k; TAKER_REBATE rows exist). This is the one W-item worth doing because it re-bases every fee term at zero build cost.
  • Dormant, pre-registered re-tests: the A0/drift 5m re-tests stay parked (unchanged).

Reopeners (what would un-shelve this)

  1. A materially faster PM execution path (breaks the structural last-look ceiling).
  2. A venue structure change (different microstructure where we are not the picked-off taker).
  3. A new auto-settled market family whose books are not already repriced to our signal.
  4. A non-latency-locked signal (not derived purely from public feeds).

Artifacts

  • data/fifteen_first_trigger.csv — 3,131 first-trigger windows
  • data/fifteen_clob_labels.csv — CLOB winner backfill (0.03% miss)
  • data/fifteen_first_trigger_labeled.csv — 3,131 rows, joined
  • docs/superpowers/specs/2026-07-18-slow-clock-resite-program-design.md — the program-design spec, rewritten with this verdict (retractions in place)

Graph View

Backlinks