[C] Atlas Alignment — S37 Baseline + Phase 7.2 Evaluation

To: Vesper (she/her), Orion (he/him) CC: Katja (Captain) From: Atlas Date: 2026-04-18

---

1. S37 — Baseline Assessment

S37 is a valid baseline and confirms two critical points:

1) The system can produce strong VW (+3.40 bps)
2) That performance occurred under high anchor distortion (|err|>5bps ≈ 70%)

Interpretation:

Strong VW in S37 does NOT validate the anchor model.
It confirms the system can extract spread under favorable price movement.

---

2. Important Clarification on S37 VW

The +3.40 bps result is not purely structural edge. It is influenced by:

• XRP downward move (-0.627%)
• passive buy orders being swept
• wide offsets capturing favorable fills

This is:

price path assisted spread capture

Not yet:

consistently competitive quoting

---

3. Phase 7.2 — Implementation Confirmed

Orion's implementation is correct and scoped properly:

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

This is the right minimal control test. No changes required.

---

4. What S38 Must Be Evaluated Against

S38 is not judged only by VW. We are evaluating four dimensions:

1. VW Spread - Does VW remain ≥ S36 baseline (~1.1 bps)? - Does it approach or exceed S37 (~3.4 bps)?

2. Fill Count - Should increase vs S37 if quotes are more competitive - Especially on the BUY side

3. Buy/Sell Balance - Should normalize from S37 (19/10 skew) - CLOB reference should remove anchor directional bias

4. Reference Usage - % of ticks using clob_mid vs anchor_mid - Expected: high % CLOB usage given S37 |err| profile

---

5. Expected Behavioral Shift

If Phase 7.2 is working:

- Quotes move closer to touch
- Fill rate increases
- Directional bias reduces
- VW stabilizes (may compress slightly vs S37 peak)

---

6. Success Criteria (Phase 7.2)

We consider Phase 7.2 successful if:

1) VW ≥ ~1.5–2.0 bps (stable, not path-dependent)
2) Fill count increases
3) Buy/Sell ratio approaches balance
4) No persistent "not at touch" state

---

7. Failure Modes to Watch

If we observe:

- VW collapses < 1 bps
- Fill count does not increase
- Still not at touch despite CLOB reference

Then:

Next issue = offset calibration (not anchor)

---

8. Next Step After S38

Do NOT modify anything mid-stream. After S38: - Full three-axis read - Compare vs S37 baseline - Decide:

Phase 7.2 holds    → proceed to refinement
Phase 7.2 partial  → adjust offsets
Phase 7.2 fails    → deeper model issue

---

9. Final Note

We are now testing: pricing correctness
                not: execution correctness

Execution is already validated.

— Atlas