Skip to content

[C] Atlas Alignment — S36 Results + Phase 7 Direction

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

Context

Your last ruling (Apr 17): run one diagnostic session at current settings (anchor cap 10, size 15), then apply the decision tree. That session is complete. Results below.

Cleanup also complete: FLAG-034 ✅ | max_inventory_usd retired ✅ | FLAG-033 ✅ | FLAG-028 ✅ | anchor_error_bps instrumentation live ✅

S36 — Diagnostic Session Results

Config: bid=10 bps | ask=14 bps | size=15 RLUSD | anchor_max_divergence_bps=10.0 | 2-hour session

Metric Value
Fills 19 total (buy=10, sell=9)
Toxic fills 0 / 19 (0.0%)
VW realized spread +0.6525 bps
Avg realized spread +1.5798 bps
XRP price change -0.398%
Anchor mean +2.70 bps
Anchor median +5.45 bps
Anchor range [-10.0, +10.0]
Anchor bias positive
Buy suppressions 72
Sell suppressions 51
Session min dist to ask 1.3 bps (SELL came very close)
Ending XRP share 36.6% (RLUSD-heavy)
Ending inventory value ~146.5 RLUSD

Decision Tree Application

Per your Case A/B/C framework:

Case A (zero/near-zero fills): DOES NOT APPLY. 19 fills in 2 hours. The anchor-cap displacement regime from S35 (CLOB-AMM divergence 20-23 bps) did not persist. S36 ran in a more moderate divergence regime.

Case B (fills return but weak): CLOSEST. VW +0.65 bps is below Phase 6A baseline (~1.1 bps threshold). Fills returned but spread quality is degraded.

Case C (fills normal): DOES NOT APPLY. VW is below baseline.

Anchor Analysis

The anchor was NOT in structural mode during S36. Range is [-10, +10] — cap engaged in both directions at different points. Positive bias dominant (mean +2.70, median +5.45), meaning CLOB was often below AMM during this session — anchor was pulled down from AMM but stopped above CLOB.

This is a different regime from S35 (CLOB above AMM by 20-23 bps, anchor pinned at -10 all session). S36 saw mixed divergence, moderate cap engagement.

RLUSD-heavy outcome: Engine ended 63.4% RLUSD / 36.6% XRP. Net SELL-side bias. Consistent with positive anchor bias — when anchor is above CLOB, SELL orders are placed closer to the real market than BUY orders.

New Instrumentation — anchor_error_bps

Per Katja's operating principle (Apr 18): if |anchor_error_bps| > 5 bps, that tick is in unreliable territory. The new stat is now live in the session summary:

Anchor: mean=+2.70bps | median=+5.45bps | range=[-10.0, +10.0] | bias=positive | |err|>5bps: X.X%

S36 did not have this instrumentation active (branch merged post-session). Next session will be the first with |err|>5bps in the summary. This gives us the reliability read Katja asked for.

Katja's Phase 7 Signal

Katja has flagged (Apr 18 morning brief) that the anchor cap is a model-level limitation, not a tuning problem. Her framing:

The capped_amm anchor is a simplification. When CLOB-AMM divergence is large and persistent, the anchor systematically lags the real market. Widening the cap is a band-aid — it delays the symptom without solving the cause.

Her proposed direction for Phase 7: - Executable price-aware quoting — quotes reference actual CLOB prices, not an anchored mid - Regime model — classify each session tick as Aligned (divergence < 5 bps), Divergent (5-15 bps), or Stress (>15 bps), and adapt quoting per regime - anchor_error_bps as primary diagnostic — not a secondary flag, but the leading indicator for quote quality - 24-hour sessions — move from 2-hour runs to continuous operation once the model is stable

What We Need From You

Immediate question (blocking next session):

Per Case B, your decision tree says "partial widening 10 → 12-13 bps." That's the mechanical answer. But given Katja's Phase 7 signal, two paths are available:

Path A — Incremental: Widen anchor_max_divergence_bps 10 → 12 bps. Run one session. Compare VW improvement. Stay within current architecture. Buys time for Phase 7 design.

Path B — Skip to Phase 7: Do not tune the cap further. Treat S36 as the last capped_amm baseline. Begin Phase 7 architecture: what does executable price-aware quoting look like, and what is the experiment sequence to get there?

Katja is signaling Path B. We want your architecture read before committing.

Specific questions:

  1. S36 verdict: Does VW +0.65 bps at 19 fills confirm Case B? Any other interpretation of the data?

  2. Path A vs Path B: Do you agree widening the cap further is a band-aid, or is there diagnostic value in a cap=12 session before Phase 7?

  3. Phase 7 architecture: If we move to executable price-aware quoting, what does the anchor model look like? Do we remove capped_amm entirely, or run it in parallel as a reference?

  4. Regime model design: Aligned / Divergent / Stress — what are the right thresholds, and what changes per regime (offset, size, both)?

  5. Next session: What exactly should it be? The new anchor_error_bps stat will be live. Should we run at current config to get a clean baseline with the new instrumentation, or change something?

We are ready to run. Waiting on your read before starting the next session.

— Vesper