Atlas Ruling — Anchor Calibration: Structural, Not Regime¶

In response to: NEO Desk/handoffs/TO_ATLAS_anchor_calibration_question.md Ruling covers: S43–S50 anchor saturation pattern, ±10 bps cap appropriateness, anchor reference price methodology, ANCHOR-CALIBRATION workstream authorization

Ruling Summary¶

The anchor is miscalibrated. The current implementation measures the persistent CLOB-AMM basis (structural) plus short-term price movement (signal). These are two fundamentally different things and should not be combined into a single metric. The ±10 bps cap is inappropriate when the structural divergence routinely runs 12–20 bps. The result is that the anchor is always at the rail during afternoon ET, making ANCHOR_IDLE the steady state rather than the exception. This is a calibration problem, not a regime problem.

The answer to Katja's question is (B).

Architectural Finding¶

What the anchor is currently measuring¶

The anchor error is computed as:

anchor_error = AMM_price - CLOB_mid

This picks up two signals simultaneously:

Structural pair basis — the persistent discount at which CLOB mid trades relative to the AMM price on this pair. For XRP/RLUSD during afternoon ET, this is approximately −12 to −20 bps and is stationary (it does not meaningfully recover within a session).
Short-term regime distortion — transient deviations from the pair's normal basis that represent genuine market hostility: flash divergence, flow pressure, temporary liquidity gaps.

Only signal (2) should be triggering ANCHOR_IDLE. Signal (1) is background noise for this pair, not a tradeable regime indicator.

Why the cap makes it worse¶

With ±10 bps cap and a structural floor of ~12 bps minimum raw divergence during afternoon ET:

Raw anchor error: 12–20 bps (every tick)
Capped to: +10 bps (every tick)
Exit threshold: <6 bps for 30 consecutive ticks
Exit condition: mathematically unreachable

This is confirmed by S50: min raw divergence 11.9 bps, mean 10.0 bps at cap, 184/184 ticks cap-locked. The engine cannot exit ANCHOR_IDLE during afternoon ET regardless of actual market conditions.

Why S48 behaved differently¶

S48 ran at ~07:00Z (overnight ET). The CLOB-AMM structural gap narrows during overnight hours — likely 3–6 bps rather than 12–20 bps. With the anchor error closer to structural-only and within the cap range, the signal was not saturated. This is the same pair with the same methodology but a different underlying basis — further evidence that the current anchor is measuring a time-varying structural feature, not a regime distortion.

What the Anchor Should Measure¶

The correct anchor formulation:

anchor_signal = (clob_mid - amm_price) - rolling_pair_basis

Where rolling_pair_basis is a slowly-updating estimate of the structural CLOB-AMM discount for this pair at this time of day. The anchor signal then measures deviation from that baseline — i.e., the residual distortion relative to normal.

Under this formulation: - When the CLOB-AMM gap is exactly the structural baseline → anchor error ≈ 0 → ACTIVE - When the gap widens beyond baseline → positive anchor error → potential ANCHOR_IDLE - When the gap narrows within normal range → anchor error ≈ 0 → ACTIVE or exit

This correctly separates the structural pair basis from the time-varying regime signal.

Answers to the Four Questions¶

Q1: Is the anchor reference price correctly constructed for the XRP/RLUSD CLOB?

No. It is susceptible to permanent bias from persistent CLOB-AMM divergence on this pair. The structural basis is embedded directly in the anchor error with no normalization.

Q2: Is the ±10 bps cap appropriate?

No. With structural divergence running 12–20 bps, the cap clips all signal before it can be evaluated. A wider cap alone does not fix the problem — the root issue is the formulation, not the cap size. However, the cap will need recalibration once the residual-basis formulation is implemented.

Q3: What should the anchor actually be measuring?

Residual distortion relative to the pair's structural baseline. See above. The anchor should be zero-centered around the normal CLOB-AMM relationship for this pair, not around zero absolute divergence.

Q4: Is this a calibration problem rather than a regime problem?

Yes. FLAG-046 (ANCHOR_IDLE) was the correct structural fix for the episode budget issue. But it does not fix the root: the engine is idling during conditions that are actually tradeable. ANCHOR_IDLE is the right behavior when the market is genuinely hostile. It is the wrong behavior when the anchor is miscalibrated and the market is active.

Authorization¶

New workstream: ANCHOR-CALIBRATION is hereby authorized. This workstream owns the anchor reference price methodology, cap calibration, and any threshold adjustments.

Constraints (mandatory)¶

Do NOT just widen the cap. Raising ±10 bps to ±25 bps without fixing the formulation moves the rail but does not fix the underlying bias. The engine would still be measuring structural basis + short-term signal.
Do NOT loosen thresholds or lower prevalence requirements without a corrected formulation in place.
Do NOT proceed to Phase 7.4 clean session preconditions until the anchor model is corrected. Sessions under the current calibration cannot be evaluated for Phase 7.4 gate — the anchor behavior is not interpretable as a regime signal.

Interim operating guidance¶

While ANCHOR-CALIBRATION is in progress, treat anchor-triggered idling as inconclusive, not authoritative. The engine is pausing for an unknown reason (calibration error + regime, combined unresolved). Do not count sessions against Phase 7.4 clean session precondition in either direction.

Required Deliverable Before Implementation¶

I want a short calibration memo from Vesper covering three candidate anchor formulations. For each formulation, explicitly state:

What it blocks
What it allows
How it would have behaved in S48 vs S49/S50

Memo format: NEO Desk/handoffs/TO_ATLAS_vesper_calibration_memo.md

That memo comes before any code changes. Do not begin implementation until I have reviewed the formulations and ruled on one.

Three Candidate Formulations (to evaluate in memo)¶

Option 1 — Residual Basis Anchor¶

anchor_signal = (clob_mid - amm_price) - rolling_pair_basis

Rolling basis updates slowly (e.g., EMA over last N minutes). Anchor error is relative to historical norm.

Option 2 — CLOB-First Anchor¶

CLOB mid is the primary price reference. AMM price is diagnostic context only — used to detect unusual conditions but does not drive the anchor error calculation directly.

Option 3 — Dual Signal¶

Separate the structural basis metric (slow-moving, time-of-day aware) from the regime distortion metric (fast-moving, tick-level). ANCHOR_IDLE fires only on the distortion metric, not on the basis metric.

New Flag¶

This ruling generates FLAG-048 — ANCHOR-CALIBRATION: anchor reference price methodology requires correction. Vesper to file.

— Atlas (CSO) Filed by Vesper (COO) on behalf of Atlas 2026-04-22