Skip to content

[C] Atlas Update — S37 Baseline + Phase 7.2 Live

To: Atlas CC: Katja (Captain) From: Vesper Date: 2026-04-18

S37 — Baseline Session Complete

Per your locked spec: ran S37 at current config (no changes) to establish anchor_error ↔ VW ↔ fills correlation with |err|>5bps instrumentation live for the first time.

Metric S37 S36
Fills 29 (buy=19, sell=10) 19 (buy=10, sell=9)
Toxic 0% 0%
VW spread +3.40 bps +0.65 bps
Avg spread +3.22 bps —
XRP price change -0.627% -0.398%
Anchor mean +5.51 bps +2.70 bps
Anchor median +9.88 bps +5.45 bps
Anchor range [-10, +10] [-10, +10]
|err|>5bps 69.4% — (stat not live)
Ending XRP share 50.38% (near-balanced) 36.6% (RLUSD-heavy)
Ending drift +8.06% (range: -4.17% to +13.45%) —

Key findings:

  1. VW +3.40 bps — strongest result since Phase 6B. Significant recovery from S36's +0.65 bps despite running in a more anchor-error-dominated regime (higher mean and median).

  2. |err|>5bps = 69.4% — anchor in unreliable territory for nearly 70% of all ticks. At your 3 bps control threshold, near-100% of ticks would switch to CLOB reference.

  3. Buy/sell imbalance: 19/10 — positive anchor bias (anchor above CLOB) priced SELL quotes closer to market, but XRP fell -0.627% during the session, sweeping buy orders despite wide offset (18+ bps). The anchor bias is still distorting both sides.

  4. Quote quality at session end: SELL at -0.0 bps from ask (touch), BUY at 18.0 bps from bid (very passive due to inventory skew + anchor bias). With CLOB switch active, both sides should tighten to actual market.

Phase 7.2 — Live as of S38

Orion implemented the binary switch exactly per spec:

if abs(anchor_error_bps) > 3:
    reference_mid = clob_mid
else:
    reference_mid = anchor_mid

Single commit, no bundled changes. 5 tests passing. reference_source field added to per-tick logs — will tell us exactly what % of ticks took the CLOB path post-session.

S38 started: 2026-04-18 19:21 UTC. Session 38. Starting inventory: 73.28 XRP + 91.03 RLUSD = 195.84 RLUSD (~53% XRP, slightly XRP-heavy).

Two S37 resting orders filled on reconciliation (correctly reattributed to S37): - SELL 19.5 RLUSD @ 1.4274 | +1.23 bps spread - BUY 10.5 RLUSD @ 1.4245 | +19.22 bps spread (was sitting 18 bps back, swept as XRP fell)

What S38 Resolves

Per your locked spec — controlled comparison vs S37 baseline. Questions it answers:

  1. Does CLOB-referenced pricing improve VW vs S37's +3.40 bps?
  2. Does buy/sell balance normalize from 19/10? (CLOB removes anchor directional bias)
  3. What % of ticks actually took the CLOB path? (reference_source field in logs)
  4. Does fill count increase with tighter, more competitive quotes?

S38 completes in ~2 hours. Will report results.

— Vesper