Orion Investigation Scope — FLAG-048: Anchor Dual-Signal Calibration¶

Six investigation questions, pre-code. Results delivered as pre-code findings memo before any branch is cut.

Question 1 — Replay Feasibility Scout¶

Do the S48/S49/S50 tick tables carry clob_mid + amm_price separately, or only the collapsed anchor_error_bps?

If collapsed: replay may need to reconstruct from circuit_breaker events or extend the tick schema backward. This answer determines everything downstream.

Starting here.

Question 2 — Signal Pipeline Trace¶

Exact insertion point for dual-signal computation in the current _tick flow. Which existing ±10 bps cap call-sites get retired vs. relocated?

Question 3 — EMA Window Choice Backed by Data¶

Not intuition. Benchmark 100 / 300 / 600 / 900 ticks on real replay and pick the smallest that: - (a) tracks time-of-day drift - (b) drives residual near zero during stable windows - (c) still responds to genuine distortion

Vesper suggested 300 as starting point. Will defend whatever is proposed with numbers.

Question 4 — Cross-Session Persistence Design¶

Open question. Leaning toward persisting rolling_basis_baseline_bps + an effective sample count in engine_state at close, so the next session resumes the EMA at the correct weight rather than flat-starting.

Will evaluate before recommending: - 48h-idle staleness - Fresh-DB behavior - Schema-change fail modes

Question 5 — Exit Reachability Proof¶

Atlas's critical requirement. Find at least one window in S49/S50 replay where residual would drop below exit threshold.

If no such window exists in the replay data, something is wrong before code starts.

Question 6 — Rail-Lock Counterfactual¶

Show S49/S50 residual does NOT saturate under the new model. The 12–20 bps structural basis should be absorbed into the rolling baseline, leaving residual near zero during stable afternoon windows.

Question 7 — Operator-Visible Surface Map¶

Every place today's anchor_error_bps surfaces — dashboard, session summary, circuit_breaker rows, tick telemetry — paired with a dual-field replacement plan.

No hidden substitution: operator must be able to see both structural_basis_bps and residual_distortion_bps separately.

Deliverable Order¶

Investigation results (all 7 questions)
Pre-code findings memo: 03 Branches/feat-anchor-dual-signal-calibration/[C] Orion Pre-Code Findings — FLAG-048.md
Vesper review
Branch cut
10–15 tests including the three Atlas-mandated validation tests (S48/S49/S50 replay, rail-lock proof, exit reachability proof)
Standard patch bundle + delivery memo

Scope Discipline¶

In scope: anchor guard state, signal computation, config parameters, logging, dashboard fields, session summary fields, tests.

Out of scope: drift guard, corridor guard, truth reconciliation, DEGRADED/HALT paths. This is a scoped calibration layer only.

— Orion (Director of Engineering) 2026-04-22