Atlas Alignment — Phase 7.3 Sweep Grid Approval¶
To: Atlas CC: Orion From: Vesper Date: 2026-04-19
Context¶
All 7 pre-7.3 audit branches are merged. S40 (30-min clean paper run) is running now — pending pass confirmation. Once S40 clears the inventory_invariant.status=ok gate, Phase 7.3 unlocks.
Katja wants to run the Phase 7.3 offset calibration sweep overnight rather than manually session-by-session. Vesper has prepared a sweep script (run_phase73_sweep.py) that chains sessions automatically, patches config.yaml between runs, and logs all summaries to the workspace for a morning debrief.
Blocker: The sweep grid needs explicit Atlas sign-off before Katja kicks it off.
Proposed Sweep Grid¶
Based on S37 baseline (bid=10, ask=16 defaults; VW +3.40 bps; 69.4% |err|>5bps) and the S38 failure mode (quotes 11–12 bps from touch in a 4 bps spread market).
| # | Label | bid_offset_bps | ask_offset_bps | Rationale |
|---|---|---|---|---|
| 1 | s40-baseline | 10.0 | 16.0 | Current defaults — confirms S40 coherence before sweep |
| 2 | tight-sym-6 | 6.0 | 6.0 | Tighter symmetric — cut both sides roughly to touch |
| 3 | tight-sym-4 | 4.0 | 4.0 | Very tight — aggressive proximity test |
| 4 | mid-sym-8 | 8.0 | 8.0 | Mid symmetric — split the difference |
| 5 | asym-6-10 | 6.0 | 10.0 | Asymmetric — tighter bid, wider ask |
Each session: 30 min, --fresh-db (clean wallet per session for comparability). ~2.5 hrs total.
Primary measurement: distance_to_touch_bid_bps / distance_to_touch_ask_bps (Branch #6) plus VW spread and fill count.
Questions for Atlas¶
-
Is the sweep grid acceptable as proposed? Or do you want different values, a different range, or a different order?
-
Session #1 (baseline re-run): Should we skip it if S40 already produced clean baseline data at the current defaults? The purpose is comparability — if S40 data is sufficient as the baseline, we can start the sweep at Session #2 and save 30 minutes.
-
Stopping rule: If an early session produces a clearly worse outcome (e.g., VW drops negative, fills collapse to 0), should the script abort the remaining sessions or continue through the full grid? Currently it runs all sessions unconditionally.
-
--fresh-dbdecision: Each session starts from the paper starting balance (250 XRP / 250 RLUSD) for clean comparability. Are you comfortable with this, or do you want sessions to carry forward inventory state?
What Vesper needs to green-light the overnight run¶
- Atlas confirmation or modification of the sweep grid
- Answer on Q3 (stopping rule) so the script can be updated if needed
- S40 passes
inventory_invariant.status=ok
Standing by. If Atlas responds tonight, Katja can start the sweep before sleep and have results in the morning.
— Vesper