[C] Atlas Alignment — Phase 7 Locked Spec
To: Vesper (she/her), Orion (he/him) CC: Katja (Captain) From: Atlas Date: 2026-04-18
1. Phase 7.1 — Confirmed¶
Phase 7.1 instrumentation is complete. The |err|>5bps stat is exactly what we need. No further changes before the next run.
2. Threshold Clarification (Phase 7.2 blocker resolved)¶
Decision: Phase 7.2 threshold = 3 bps
Why 3 bps: - ≤ 3 bps → pricing still "close enough" to be considered aligned - 3 bps → degradation begins (even if not catastrophic yet) - 5 bps → reliability breaks (Katja's rule)
These are complementary, not conflicting.
3. Phase 7.2 — Minimal Implementation (LOCKED SPEC)¶
Orion implements exactly this:
clob_mid = (best_bid + best_ask) / 2
if abs(anchor_error_bps) > 3:
reference_mid = clob_mid
else:
reference_mid = anchor_mid
Constraints: - No blending yet - No weighting yet - No additional regime logic
This is: binary switch → clean experiment → high signal
4. Architecture Confirmation¶
Fully aligned:
5. Next Session — APPROVED¶
Run next session NOW at current config (no changes).
Purpose: Establish baseline — anchor_error ↔ VW ↔ fills
Without this, Phase 7.2 lacks calibration context.
6. What the Next Session Validates¶
- What % of time is |anchor_error| > 5 bps?
- How does VW behave inside vs outside that band?
- Are fills occurring primarily when error is low?
This gives us causal mapping, not correlation guessing.
7. Orion — Next Steps (queued, not executed yet)¶
After baseline session: - Implement Phase 7.2 switch (threshold = 3 bps) - No other changes in same commit - Run controlled comparison vs baseline
8. Progression¶
We are at: Observe → about to enter Control
Summary — Locked Decisions¶
| Item | Decision |
|---|---|
| Phase 7.1 | ✅ Complete — |
| Phase 7.2 threshold | 3 bps (locked) |
| Phase 7.2 spec | Binary CLOB/anchor switch |
| Next session | Current config unchanged |
| Cap widening (Path A) | ❌ Rejected |
| Regime model | ALIGNED ≤3 / DIVERGENT 3-8 / STRESS >8 |
— Atlas