[C] Atlas Alignment — S36 Verdict + Phase 7 Architecture

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

S36 — What It Proves¶

vs S35: - S35 → structural failure (0 fills, anchor pinned) - S36 → fills returned (19 fills)

What didn't recover: - VW only +0.65 bps — still below known baseline (~1.1 bps)

Conclusion:

S36 confirms:
→ system works mechanically
→ but pricing quality is degraded

This is not instability. This is a consistent pricing inefficiency under current model.

The Key Signal¶

From S36 anchor stats: mean=+2.70 bps | median=+5.45 bps | range=[-10, +10]

You are oscillating between:
- mildly wrong pricing
- severely wrong pricing (when cap hits)

Even when NOT in full structural mode, anchor is still frequently wrong enough to degrade VW.

Q1 — S36 Verdict¶

Yes — Case B confirmed.

More precise wording:

This is not a temporary inefficiency.
This is a structural pricing limitation of the anchor model.

Q2 — Path A vs Path B¶

❌ Path A (cap 10 → 12):

Yes, it would likely increase fills and slightly improve VW. But:

It does NOT solve the problem.
It only shifts the boundary where the problem appears.

✅ Recommendation: Path B — skip to Phase 7 now.

Why: - cap=10 → failure in S35 - mixed regime → degraded VW in S36

This proves:

The problem is not the cap value.
The problem is the concept of a capped anchor itself.

Q3 — Phase 7 Architecture¶

Should we remove capped_amm?

No — demote it.

Do NOT delete it. Instead:

capped_amm becomes a reference signal,
not the primary price driver.

New structure:

Instead of:

AMM → capped → anchor → quotes

Move to:

CLOB mid (primary)
+ AMM deviation (context)
→ executable_mid
→ quotes

Clean rule:

CLOB defines execution reality.
AMM defines context / opportunity.

Q4 — Regime Model¶

Use anchor_error (not just divergence) as the regime signal.

Recommended thresholds:

Regime	Condition	Meaning
ALIGNED	≤ 3 bps	safe
DIVERGENT	3–8 bps	degraded
STRESS	> 8 bps	unreliable

Behavior per regime:

ALIGNED: - Current strategy - Normal offsets - Full participation

DIVERGENT: - Shift pricing toward CLOB - Slightly widen - Reduce reliance on anchor

STRESS: - Clamp to CLOB - Reduce one side or both - Protect capital

Q5 — Next Session¶

DO NOT change config yet.

Run same config with anchor_error_bps live. You need:

ground truth correlation:
anchor_error → VW → fills

Without that, Phase 7 is guessing.

Orion — Phase 7 Tasking¶

Phase 7.1 (immediate): - anchor_error_bps logging ✅ (already merged Apr 18) - Confirm stats live: % time > 5 bps, mean error per session ✅

Phase 7.2:

clob_mid = (best_bid + best_ask) / 2

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

That is the first executable model. Simple. Testable. Safe.

Important Clarification¶

Katja's instinct is exactly right — widening the cap is a band-aid. Precise version:

The anchor is not wrong because it is capped.
It is wrong because it is not the execution venue.

Decision Summary¶

S36 = Case B confirmed — system functioning, pricing degraded
Not a tuning issue — pricing model limitation
No incremental cap widening — skip Path A entirely
Move to Phase 7

Next session: Current config unchanged. Collect anchor_error_bps baseline. Establish correlation vs VW and fills.

Architecture: CLOB = primary price reference. capped_amm = secondary signal. Anchor no longer controls the midpoint.

Regime model: ALIGNED / DIVERGENT / STRESS per thresholds above.

Goal:

Move from anchor-based quoting
→ executable market-aware quoting

— Atlas