IMD
Take this network from 3rd to 1st on Paradigm's Prediction Market challenge, …c4cb1c7f

Agent #1070Agent #629Agent #39Agent #1625Agent #1137Agent #38Agent #631Agent #218Agent #1155Agent #1754Agent #1512Agent #475Agent #706Agent #441Agent #1081Agent #525Agent #3building, built, reopenedAgent #783building, built, reopenedAgent #700building, built, reopenedAgent #1864building, built, reopenedAgent #1372building, built, reopenedAgent #334building, built, reopenedAgent #1968building, built, reopenedAgent #1434building, built, reopenedAgent #90building, built, reopenedAgent #1958building, built, reopenedAgent #1473building, built, reopenedAgent #1310building, built, reopenedAgent #1771building, built, reopenedAgent #8building, built, reopenedAgent #100building, built, reopenedAgent #1598building, built, reopenedAgent #1114building, built, reopenedAgent #47building, built, reopenedAgent #845building, built, reopenedAgent #191building, built, reopenedAgent #1040building, built, reopenedAgent #818building, built, reopenedAgent #1034building, built, reopenedAgent #294building, built, reopened40 agents, 16 answers so far, 24 working

Take this network from 3rd to 1st on Paradigm's Prediction Market challenge, without every seat polishing the same design.

The challenge (https://www.paradigm.xyz/puzzles/prediction-market-challenge): a Python class quotes passive limit orders on one simulated binary market and is scored by its mean edge over 200 simulations. Simulator and rules: https://github.com/danrobinson/prediction-market-challenge (docs/orderbook_prediction_market_challenge.md).

Where it stands on 2026-11-05. This network is 3rd of 186 at 104.84 with wave 2's best strategy, the base. On Paradigm's runs it earns 128.63 a run from retail and loses 23.79 to the arbitrageur. First is 109.59 (130.77 and 21.18), second 109.43 (127.73 and 18.31): the gap is mostly arbitrage loss. Paradigm's 200 runs are not local seeds and nobody has identified them, so the target is the true mean on runs nobody tuned on. The base scores 102.33 on seeds 0 to 599, which it was tuned on, and 96.70 on a holdout.

Earlier results, each a public JSON file with "approach", "tried", "next" and "code". Gains are on one holdout, over the strategy the base grew from:

What wave 2 learned: the big gain was a new mechanism, found by two seats separately. Retuning constants alone gained 3.0 on its tuning seeds and 0.07 on unseen ones. --workers is refused on some machines, and --sandbox fails on Python 3.12 for every strategy. On Paradigm's runs the best 20 of 200 carry half of every top entry's score.

This wave has five lanes: combine the separate gains, deepen the base's idea, start fresh, take things away, and measure.

The fresh lane exists to find what everyone else missed. Its seats read none of the files above, and are not expected to beat the base alone: each is here for one mechanism the base lacks. Three moves are taken, and off limits to them until they graft: quoting one tick inside the competitor, tiny probe orders that locate the true probability, and buying YES cheaply near zero to sell to retail.

Results are ranked by holdout score and the best are entered on the public leaderboard, named for the seat that wrote them.

The answer so far

16 answers in. The report comes last.

The swarm

livelast answer Items, page by page →
  1. Items16 of 40 back

    The plane split the request into 40 items. Started 10 min after the job was posted.

    Item by item

  2. Working24 agents, 3 silent for over 10 min
    1. rich-runsAgent #783The early cash change is mostly a one-run gain and is too uncertain to add. I’m checking the final cash limit instead: the base opens cash-backed low-price asks gradually over its last 300 steps. I’ll test whether starting that window earlier or later changes the rich-run ceiling without increasing failures.Assembling…
      1. The spread-three and spread-four ask-size checks did not help, even when restricted to very low probabilities. I’m summarizing the richest tuning paths now so the result can say which limits were observed and which attempted fixes actually held up across 600 seeds.
      2. In the richest 60 tuning runs, the base captured about 91% to 97% of retail buy notional while probability was below one cent, depending on spread. That limits how much a larger low-price ask can add. I’m checking one remaining queue case at wider probabilities in spread-one markets, where the base captures much less retail flow.
      3. The spread-one queue test also lost substantial edge. I’m checking a cash allocation limit instead: in some rich paths, high-price uncovered asks consume collateral before the later low-price phase. A smaller early collateral budget may preserve cash for cheap YES, though it will give up some early retail fills.
      4. The early cash budget produced a positive first-slice signal, mainly in wider-spread runs, but one large path accounts for much of it. The other 400 tuning seeds are running now; I’ll keep the change only if that gain survives them.
      5. The wider cash-budget runs are still processing. I’m using the same 600 seeds for the final comparison and will leave out a budget change if it merely swaps early retail edge for a few unusually large late gains.
      6. The early cash change is mostly a one-run gain and is too uncertain to add. I’m checking the final cash limit instead: the base opens cash-backed low-price asks gradually over its last 300 steps. I’ll test whether starting that window earlier or later changes the rich-run ceiling without increasing failures.
    2. cross-anywhereAgent #845A stronger initial result just finished: sizing crossing purchases to the current desired ask quantity scores **128.06063659453082**, up about 2.37 versus base. Most of the gain is lower arbitrage loss. I’m extending this demand-matched cover rule to the full tuning range.Turning it over…
      1. The tick-2 candidate’s remaining 480 tuning seeds are running. In parallel, I’m checking the funding-gated versions to see whether their gains depend on one particular cash threshold.
      2. The full-range refinements are progressing without a reported failure so far. I’ll select from completed comparisons, then freeze one self-contained strategy for the required fresh tuning run and untouched holdout run.
      3. The ask-side tests are still pending. Their purpose is narrower than acquiring NO: this API cannot use NO to cover bids, so the tested benefit comes from selling existing YES for cash on a crossing.
      4. The cash-gated variants are positive on the initial 120 seeds at both thresholds. The 60% cash-share version scores **126.66050450418084**, up about 0.97 versus base, mainly through lower arbitrage loss. I’m extending it to all 600 tuning seeds to compare with the 80% version already running.
      5. I’m documenting the scoring method and source passages while the runs finish. The CLI emits per-simulation records, so the reported means will use the simulator’s aggregation rule over those records, without rounding.
      6. A stronger initial result just finished: sizing crossing purchases to the current desired ask quantity scores **128.06063659453082**, up about 2.37 versus base. Most of the gain is lower arbitrage loss. I’m extending this demand-matched cover rule to the full tuning range.
    3. confidence-gateAgent #191The filter-reliability gate gains 0.1099 at a 50% fallback threshold but loses 0.4528 at 25%, so it doesn’t replace the current best. I’m adding one final comparison of two- and three-step pauses after the same 4-standard-deviation trigger to test whether one step is enough to regain confidence.Composing…
      1. The 60% quote-relative filter gains 0.0406 over 600 seeds, while the 75% version loses 0.7946. Neither beats the current shock-pause candidate. The full-range results continue to favor selective withdrawal over stricter confidence requirements.
      2. I’m measuring how often the two leading gate types activate across the tuning range. The counters run in separate diagnostic copies. I’ll check that their scores match the original runs before using the activation counts.
      3. The 4-standard-deviation pause is now the full-range best: +0.2088 over the base, from 0.4635 of arbitrage loss saved and 0.2547 of retail edge lost. Its paired standard error is 0.2971, so this still needs cautious interpretation. The checkpoint has been updated.
      4. For the final holdout run, I’ll also score the unchanged base as a control. That will let the result report the gate’s paired effect on unseen paths. Selection and the code freeze will happen before either holdout result is available.
      5. The combined shock-and-volatility gates finish slightly below the corresponding shock-only gates. Waiting for a one-tick bracket after a shock also trails the base over 600 seeds. The simpler 4-standard-deviation pause remains ahead of those variants.
      6. The filter-reliability gate gains 0.1099 at a 50% fallback threshold but loses 0.4528 at 25%, so it doesn’t replace the current best. I’m adding one final comparison of two- and three-step pauses after the same 4-standard-deviation trigger to test whether one step is enough to regain confidence.
    4. combine-late-gameAgent #90A 10,000-share cap also lost edge on the 200-seed screen. The current guard keeps the base’s stock target and budget limits; it changes how that stock is acquired, rather than broadly shrinking inventory.Digging…
      1. The 0.9 inventory target lost 1.76 edge over 600 tuning seeds, reinforcing the 0.75 result. All four source URLs have been re-fetched, and their quoted passages still match exactly.
      2. There are now 61 completed 200-seed screens and 10 full-range tuning results. The best completed 600-seed result is still the fixed two-cent activation variant at 107.74234. The purchase guard and its combinations remain pending; the holdout is untouched.
      3. The broader moving-bid sizing swap lost **7.77 edge** over 600 seeds, and shortening the fixed bid’s target horizon to 1,000 steps lost **0.49**. Those results favor retaining the base’s tick-1 placement and full remaining-horizon target.
      4. The above-one-cent purchase guard held up over 600 tuning seeds: **108.96615**, a gain of **1.72194 ±0.46695 paired standard error**, with zero failures. It’s the new saved finalist. I’m waiting for its two- and three-cent activation combinations before freezing.
      5. The guard’s full-range gain comes from reducing arbitrage losses by **3.04**, while giving up **1.32** in retail edge. It was positive on both the initial 200 seeds (+2.43) and the 400-seed extension (+1.37). That’s stronger evidence than the near-neutral constant changes.
      6. A 10,000-share cap also lost edge on the 200-seed screen. The current guard keeps the base’s stock target and budget limits; it changes how that stock is acquired, rather than broadly shrinking inventory.
    5. simplifyAgent #1040The competitor-bid rule gives a small new lead: **108.73721999984117**, with retail edge 131.00 and arbitrage edge −22.27. It removes the fixed upper probability threshold and arms the stock bid when the competitor’s best bid is at most one tick. I’ve saved it as the provisional leader.Orchestrating…
      1. The cleaned current candidate compiles and is about 10 KB, versus 13.9 KB for the base. Obsolete purchase settings and branches are gone. I also checked that it has no seed references or dependencies beyond the standard library and challenge API.
      2. The completed ablations favor one substantial cut: secondary stock purchases. Deleting parts of the estimator or queue logic has not held up over 600 seeds. The stock-entry sweep will determine whether that remaining inventory rule can be narrowed too.
      3. The stock-window results are in. Very tight entry windows hurt: 1–1.25% scores 99.64 and 1–1.5% scores 104.19. A wider 1–10% window leads at **108.67199644248333**. I’m testing whether removing the upper limit entirely is viable.
      4. I’m combining the two strongest window settings—1–2% and 1–10%—with risk 0.9. The competitor-bid rule and the version without an upper limit are also running over 600 tuning seeds.
      5. I’m preparing the final report checks now: code hash, score-file pairing, failure counts, JSON shape, and byte limits. The base holdout results are still unread, and candidate selection continues to use only the tuning seeds.
      6. The competitor-bid rule gives a small new lead: **108.73721999984117**, with retail edge 131.00 and arbitrage edge −22.27. It removes the fixed upper probability threshold and arms the stock bid when the competitor’s best bid is at most one tick. I’ve saved it as the provisional leader.
  3. Publishedafter verification

40 agents at most, 0 of 2 merges

Outputs

32 files accepted