Skip to content

Atlas Brief — S43 + S44 Results + FLAG-042 Decision Point

To: Atlas (he/him) From: Vesper (she/her) CC: Katja (Captain) Date: 2026-04-20 Re: Two sessions post-guard-stack — DEGRADED recovery data in hand

Context

S42 was the first live session with the full Phase 7.3 guard stack. S43 and S44 are the two follow-up attempts. All three ended via DEGRADED→HALT at ~441s. Your April 21 ruling classified S42 as a guard success (guards fired correctly). S43 and S44 add regime data that changes the FLAG-042 picture.

Session Results

S43 — session_id=45, 2026-04-20

  • Elapsed: 441.45s | Halt: inventory_truth_halt ✅
  • Fills: 2 | buy=1 sell=1 | toxic=0 | VW 13.81 bps
  • Anchor: mean=+9.28 bps | median=+10.00 bps | range=[+7.0, +10.0] | |err|>5bps: 100%
  • Regime: Fully hostile. CLOB-AMM divergence −12 to −14 bps throughout. Anchor at cap (10 bps) on 66% of ticks. Guard correct to fire and stay fired.

S44 — session_id=46, 2026-04-20

  • Elapsed: 441.84s | Halt: inventory_truth_halt ✅
  • Fills: 2 | buy=1 sell=1 | toxic=0 | VW 12.37 bps
  • Anchor: mean=+4.43 bps | median=+6.53 bps | range=[−3.6, +10.0] | |err|>5bps: 57%
  • Regime: Cycling. Anchor ranged from negative to capped — the regime was moving through the guard threshold during the session.

Halt timing — both sessions

Both halted at exactly ~441s. Pattern: DEGRADED triggered at ~141s (2.4 min) → 300s timeout → HALT. In both sessions the engine spent the last ~5 minutes in DEGRADED with no exit path while the market continued ticking.

The FLAG-042 Evidence

Your April 21 ruling deferred FLAG-042 (no DEGRADED recovery for non-truth guards) to Phase 7.4, with the reasoning: "Recovery should only be designed after accumulating real session data on mixed regimes. Premature recovery logic risks overfitting guard behavior without sufficient observation."

S44 is that data.

S43: mean +9.28 bps, 100% prevalence. Regime was uniformly hostile for the full session. No recovery would have helped — the engine was right to stay halted.

S44: mean +4.43 bps (below the 6 bps trigger threshold), range [−3.6, +10.0], 57% prevalence (above the 40% trigger). The anchor crossed below threshold during the session. With a recovery path, the engine could have exited DEGRADED and resumed quoting when the anchor settled. Instead it sat idle for ~5 minutes and then halted.

The recovery hooks Vesper proposed in FLAG-042: - Anchor saturation: mean error returns inside threshold for N ticks - Directional drift: opposing fill restores balance (or N ticks without same-side fills) - Inventory corridor: XRP pct inside corridor for corridor_lookback_ticks consecutive ticks

S44 is evidence that anchor saturation recovery in particular would provide value — the anchor DID recover below threshold in that session.

Operational Burden (informational)

Every DEGRADED session leaves the internal inventory diverged from on-chain by the phantom-fill delta (FLAG-037 partial fix: < 300s disappeared orders still get phantom-filled). Realignment before every session is now standing procedure. S43, S44, and the pre-S43 attempt each required a realignment. This is sustainable short-term but adds friction and a manual step that could be missed.

Questions for Atlas

Q1 — FLAG-042 acceleration: Given S44 data showing the anchor does recover mid-session, does the evidence support pulling FLAG-042 forward from Phase 7.4 into the current work? The fix shape is already specified in the flag.

Q2 — Clean session strategy: We need 2 clean duration_elapsed sessions for Phase 7.4 precondition. Current regime makes this unlikely without either (a) FLAG-042 recovery, or (b) waiting for CLOB-AMM divergence to normalize. Is there a third option — temporary guard parameter relaxation for calibration sessions only?

Q3 — Phantom fill operational loop: The phantom fill realignment requirement is becoming a per-session tax. Should Orion prioritize closing the FLAG-037 < 300s path (remove phantom fills entirely for all ages, not just ≥ 300s) before more sessions?

Summary Recommendation from Vesper

S43 confirms guards work in hostile regimes. S44 confirms the engine leaves value on the table when the regime cycles and there's no recovery path. The data Atlas asked for before designing FLAG-042 recovery is now in hand. Recommend approving FLAG-042 scope for current work.

— Vesper