NEO Trading Engine — Agent Alignment Briefing¶

Date: April 14, 2026 (early AM ET) Audience: Any Claude agent working on NEO in VS Code or Cowork Author: Cowork Claude (session with Katja, Apr 13–14)

Current State of the Engine¶

NEO is a live XRPL market-making engine running Stage 1 funded sessions on neo_live_stage1.db. It places two-sided quotes (buy + sell) around a mid-price, captures spread, and manages inventory via a skew-based control loop.

What works: The engine reliably captures spread. VW realized spread ~3 bps across sessions. Fills both sides. Portfolio at ~141 RLUSD against 137.28 net deposits = ~3.96 RLUSD cumulative trading value. Zero toxic fills. The core market-making loop is sound.

What doesn't work yet: Inventory stability at 10 RLUSD order size. Two consecutive sessions halted early on exposure caps (one RLUSD-side, one XRP-side). Skew control loop maxes out at its ceiling and cannot arrest drift during directional market moves at this size.

What Just Happened (Apr 13–14 Sessions)¶

Two sessions ran at 10 RLUSD size, both hit exposure caps:
Session A: RLUSD cap (102.9/100), 3340s runtime, 40/40 fills, skew pegged at -8 bps
Session B: XRP cap (101.6/100), 2093s runtime, 48/42 fills, skew pegged at +8 bps
Direction flipped with market regime; failure mode identical (skew ceiling binding)
Mid-price reconstruction bug found and fixed:
dashboard.py:_reconstruct_mid and summarize_paper_run.py:load_quote_quality_summary both used global median inventory (single xrp_med/rlusd_med) in a per-tick mid formula
With inventory swinging (e.g., XRP 37→63), the median was wrong for ~42% of ticks
Produced phantom negative sell distances — 27% noise in Session A, 78% in Session B
Engine's live control loop was NOT affected — it uses snapshot.mid_price (CLOB, written every tick)
Fix: both code paths now use market_snapshots.mid_price via merge_asof/bisect (10s tolerance), per-tick inventory reconstruction as fallback only
Skew ceiling experiment (8→12 bps) was run but inconclusive:
Inventory initially balanced well (49.9% XRP at one point)
Then market moved directionally and drift resumed, hitting cap at +12
Cannot cleanly attribute drift-to-halt to skew insufficiency vs degraded diagnostics vs anchor distortion
Decision: pause tuning until diagnostics verified clean

Active Decisions and Constraints¶

Decision	Status
Sell offset	Locked at 16 bps (was 18, tuned down in Phase 5A)
Skew ceiling	Reverted to 8 bps pending clean data
Order size	Stepping down to 7 RLUSD for verification sessions
Exposure caps	100 RLUSD each side, unchanged
Anchor mode	capped_amm, 10 bps cap, 73% utilization (escalating — FLAG-002)
Tuning status	PAUSED — no parameter changes until mid-noise verified at 0%

Immediate Next Steps (In Order)¶

Verify mid-noise fix — Run a session (7 RLUSD size). Both dashboard and terminal summary should show 0% invalid ticks. This is the gate for everything else.
Establish clean 7 RLUSD baseline — One full-duration session with verified diagnostics. This becomes the reference point.
Re-test at 10 RLUSD — With clean diagnostics, determine if drift-to-halt persists. If yes, skew slope (not ceiling) is the next variable.

Open Flags That Matter Right Now¶

Flag	Status	Why It Matters
FLAG-002	Active	Anchor 73% capped and climbing. If it degenerates to fixed offset, mid-price reference shifts. Monitor.
FLAG-004	Active	Sell fill mechanism (queue vs sweep). Determines how to interpret sell conversion rates.
FLAG-008	Active	WAC rebuild bug. Blocks capital injection path. Must fix before any injection + restart.
FLAG-010	Active	Terminal summary not session-scoped for fill counts. Separate from mid-noise fix.
FLAG-011	Resolved	Mid reconstruction bug — fixed in both dashboard.py and summarize_paper_run.py.
FLAG-012	Active	Inventory drift-to-halt at 10 RLUSD. Blocked on FLAG-011 verification.

What NOT to Do¶

Do not change skew parameters (ceiling, slope, or size-scaling) until mid-noise is verified clean
Do not raise exposure caps — the drift-to-halt is a control problem, not a cap problem
Do not change anchor config — it's a known variable but not the current priority
Do not touch live engine logic (order placement, skew calculation, drift computation) — it's confirmed clean
Do not run tuning experiments until a clean 7 RLUSD baseline exists

Key Files¶

File	What It Is
`[C] Open Flags.md`	All active/resolved flags with gates and actions
`[C] Experiment Log.md`	Historical experiment results and locked parameters
`NEO Trading Engine Overview.md`	Project overview, goals, architecture context
`[C] 2026-04-14.md` (daily log)	Tonight's session details
`dashboard.py`	Streamlit dashboard — mid-noise fix applied here
`summarize_paper_run.py`	Terminal session summary — mid-noise fix applied here
`config_live_stage1.yaml`	Live config (sell offset, skew params, size, caps)

This briefing reflects alignment between Katja and Cowork Claude as of Apr 14, 2026 ~00:30 ET. If any parameter changes are proposed, check the flags file first and confirm the mid-noise gate (FLAG-011) has been passed.