Skip to content

[C] Atlas Briefing — FLAG 053 Orion Findings + S40 Diagnostic Ready

FLAG-053 — Orion's Findings

Full report: 07 Agent Coordination/[C] Orion Investigation Report — FLAG-053 EMA During ANCHOR_IDLE.md

Q1 — EMA deadlock hypothesis: REFUTED

observe() in dual_signal_calculator.py:176-241 runs unconditionally on every valid-snapshot tick. ANCHOR_IDLE does not gate Step 8. The EMA converged normally in S55. This was not the problem.

Q2 — Real deadlock found: exit evaluator window source

The exit evaluator (main_loop.py:2887-2997) routes through _select_anchor_guard_window (main_loop.py:2729-2766). Two possible sources:

  • Residual window — only available after warmup_ticks (50) + residual_hysteresis_lookback_ticks (20) = 70 valid observations. This is the correct signal source in a FLAG-048 world.
  • Legacy capped window — fallback before tick 70. Fed from last_anchor_divergence_bps, which is capped at ±10 bps. In a cap-saturated regime (82% in S55), this window holds values near ±10 bps uniformly.

Exit conditions against the legacy capped window: - |mean(window)| < 4.0 bps → fails (mean ≈ 10 bps) - pct(|x| > 5 bps) < 30% → fails (pct ≈ 100%)

Both tests fail on every tick. Stability counter never increments. ANCHOR_IDLE cannot exit for the first 70 ticks, no matter how long the session runs.

After tick 70, the residual window takes over — but then 30 consecutive stability ticks are still required to exit. Minimum ticks to exit in a hostile regime: ~100 ticks ≈ 8.5 minutes. S55 entered ANCHOR_IDLE at tick ~3, ran ~84 ticks total, and SIGINT'd — it was just barely reaching the window where exit becomes possible.

Q3 — Dashboard sign mismatch: not a bug

Three signals, three sign conventions: - clob_vs_amm_divergence_bps (tick log): (clob_mid − amm) / amm → negative when amm above mid - last_anchor_divergence_bps (dashboard): (quote_anchor − mid) / mid, capped at ±10 → positive when amm above mid - structural_basis_bps (EMA input): (mid − amm) / mid → negative when amm above mid

S55's −11 to −13 bps (tick log) and +9.56 bps (dashboard) describe the same regime — amm ~11–13 bps above clob_mid — from opposite sign perspectives, with the dashboard value additionally capped at +10. No sourcing bug. A real legibility gap, not a calculation error.

Orion requests a ruling on one of:

(i) Reduce pre-residual lockout — earlier residual-window handover, or a benign-regime exit path on the uncapped structural signal during the legacy-source phase
(ii) Standardize sign conventions across the three metric paths
(iii) Both
(iv) Accept current behavior — require sessions long enough to clear the 100-tick minimum (≥ 10 minutes in ANCHOR_IDLE before exit is achievable)

S40-Baseline Diagnostic — Ready to Run

The clone is set up and seeded: - Repo: C:\Users\Katja\Documents\NEO GitHub\neo-2026-s40-baseline - Commit: 8716821 (branch: s40-baseline) - DB: neo_s40_diagnostic.db — seeded with current on-chain balances (79.684 XRP / 83.463 RLUSD) - Codebase state: D2.2 (truth reconciliation, DEGRADED mode, realignment tool) + reconciler anomaly log. NO anchor saturation guard, NO drift guard, NO corridor guard, NO FLAG-042/044/046/048.

This is the last known clean codebase — S40 ran 316 ticks, 84 fills, halt=duration_elapsed on this code.

We are ready to run the 10-minute diagnostic on your go. Awaiting your ruling on: 1. Whether to proceed with the S40 diagnostic run now 2. Which option (i–iv) to pursue for the FLAG-053 exit-evaluator fix

— Vesper