Skip to content

STATE.md

Latest commit

 

History

History
72 lines (66 loc) · 5.21 KB

STATE.md

File metadata and controls

72 lines (66 loc) · 5.21 KB
Raw

STATE — limit-orders agent (BATCH-LMSR with limit prices)

Branch: exp/limit-orders in ~/simulations-limit. Merge to origin/main when tested (pre-authorized). Venv: mechanism-design/.venv-batch/bin/python (python3.11, numpy scipy pytest, -e galanis-market -e batch-amm). Never touch ~/simulations or other worktrees.

Task (owner, 2026-07-04)

Extend BATCH-LMSR (mechanism-design/batch-amm) with limit prices; re-run headline comparisons vs the market-order baselines (CRN). Orders become (direction, qty, limit); buy fills iff uniform clearing price pi <= limit, sell iff pi >= limit; clearing = fixed point of pi = avg-exec-price(eligible net flow), standard uniform-price treatment with marginal pro-rata fills. Honest: limit = own posterior (tight) + loose variant (posterior +/- slack); qty = competitive sizing as now. Manipulator: bounty-B, best-responds, extreme limits OK. Price-only disclosure, no attribution, seeds fixed. Measure: (1) manip cost per unit distortion — does the 2.4-4.2x batch discount close? (2) honest execution quality — does the 2.2-4.1x per-trader advantage shrink? (3) aggregation speed/accuracy — extra rounds R needed; accuracy at fixed R. (4) jamming/denial-of-aggregation (BATCH.md §9) — does limit protection change it? Deliverable: BATCH.md new section "12. Limit orders" + tables + plain verdicts on all 4 and on whether the two headline numbers survive/shrink/reverse. Commit, merge to origin/main, push.

Design decisions

  • New module src/batch_amm/clearing.py: vectorized fixed-point clearing. D(pi) = eligible net demand (step, nonincreasing); phi(X) = LMSR avg exec price (increasing); bisect g(pi)=phi(D(pi))-pi. Continuity fixed point -> full fills; jump at a limit value -> pi* = that limit EXACTLY, X* solves phi(X*)=pi*, marginal orders (|limit-pi*|<1e-9) filled pro-rata (fill light marginal side fully first if both sides marginal). Price-cap: X clipped to bounds, pro-rata alpha as base engine.
  • engine.py: mech "batch_lmsr_limit"; Config.limit_slack (0=tight, np.inf==market-order, reproduces batch_lmsr BIT-EXACTLY — final computation uses identical arithmetic on x*f with f=1); Config.manip_scale (gamma, best-response grid over order-size multiplier); manip limit = +/-inf.
  • Disclosure (price-only): Gaussian — engine converts executed net flow to t_total_obs = X_exec/(sb) + nlogit(p), feeds existing reveal_batch_anon (self-consistent: in the mean-field model limits never bind). Galanis — new reveal_batch_anon_limit(p_open, x_obs, ...): exact consistency on executed flow; per omega predict honest orders+limits, clear, keep omegas with X_pred == X_obs (1e-9); empty mask -> no update + state["jams"] += 1 (jam instrumentation, also added to reveal_batch_anon).
  • Manipulator hypothesis to test for Q4: honest tight limits PIN pi at an honest posterior (marginal pro-rata) -> manipulated outcome pools with honest outcomes -> strict-consistency update stays nonempty -> jam partially defeated.
  • Baselines re-run in-script with same env instances (SEED=20260704, SEED+n per N) for CRN pairing.

Plan

docs/superpowers/plans/2026-07-04-limit-orders.md (tasks 1-6).

Progress

  • Oriented: BATCH.md, engine.py, envs.py, metrics.py, tests, run_sweeps.py, run_disclosure.py read
  • STATE.md + plan written
  • venv healthy, 36 pre-existing tests green
  • Task 1: clearing.py + tests — committed 0790ecc
  • Task 2: engine.py batch_lmsr_limit branch + tests (bit-exact at slack=inf incl. manip; single-trader posterior bound; seat invariance; welfare identity) — committed 1559cac
  • Task 3: envs.py reveal_batch_anon_limit + jams counter (also in reveal_batch_anon); run_market returns "jams"; 61 tests green — committed
  • Task 4: run_limits.py + full M=20000 run (100s) -> results/limits_{core,manip,jam}.json. Found + fixed latent base-engine bug: alpha = X_exec/X blew up on exact-cancellation nets (|X|<1e-14 guard in engine.py AND clearing.py, identical arithmetic keeps bit-exactness; changes published Galanis R=3 manip PnL by <=9%, noted in §12; regression test added)
  • Task 5: BATCH.md §12 written, all quoted numbers verified against JSONs; verdict pointer added to top paragraph; Reproduction block updated (62 tests, run_limits.py line)
  • Task 6: scoped suite grew to 106 tests, all green (44 galanis-market + 62 batch-amm); committed, merged to origin/main, and pushed

Results summary (for §12; full numbers in results/limits_*.json)

  • Q1: 2.4-4.2x manip discount SURVIVES (gamma=1: c/dp 0.067 vs seq 0.165-0.284, every slack); gamma best-response (3-5x) doubles distortion at c/dp still < seq last seat; N=10 profitable.
  • Q2: 2.2-4.1x execution advantage SURVIVES, ticks up to 2.3-4.3x; fill 97.6-99%; welfare cost 0.0004-0.0006/market (~1-2% of profits) via slightly worse R=1 price.
  • Q3: no extra rounds (R=2 heals at every slack); fixed-R accuracy does not degrade (R>=2 tight limits even shave the anonymity plateau, N=25 err 0.158->0.107); Galanis exact identical.
  • Q4: jam unchanged for a funded manipulator (fills at inf limit; same jams/acc at all slacks); structural narrowing only: unfillable inconsistent orders can no longer jam (no-fill leaves X_exec honest/explainable — honest runs jam 0).