Polymarket Edge-Discovery Workflow (2026-06-22)

Synthesis of a 26-agent Claude edge-discovery workflow (run 2026-06-22) that scoured the Polymarket surface for NEW deployable edges for the pmv2 quant op, held to live-strategy rigor (PM-only funded execution, $1k–$25k thin-book sizing, CLOB v2 maker+taker, on-chain detection, the purged-walk-forward ship bar). The win is durable knowledge, not a fat edge. Full report archived at /Users/levander/coding/pmv2/research/edge-discovery-2026-06-22.md.

P0 — LIVE-CAPITAL RISK CORRECTION: Weather is NOT fee-free

The single most valuable output of this exercise is a mis-costing correction on live money, not a new edge. Polymarket Weather markets charge a taker fee under Fee Structure V2 (effective 2026-03-30): Weather = Theta 0.05 (same as Culture), i.e. $1.25 / 100 shares at p=0.5. The live weather strategy and every prior analysis that called weather “fee-free” assumed Theta=0.

At the actual weather fill regime (p ~ 0.85–0.96, the late-afternoon near-resolved cohort) the fee is 0.45% of notional ($2.25/$500) — at p=0.30 it would be ~1.05%. Over 3–5 trades/week that is ~$200–350/yr of bleed never modeled.

ACTION: re-run the live weather deflated Sharpe NET of this fee. A thin margin could fall below the 1.5×-cost ship bar (see purged-backtest-gate-no-edge). This is a risk-reduction on live capital — treat it as the highest-priority single action from this report, ahead of any new lane.

For Agents — net framing (read this first)

The surface yields NO high-conviction fat edge. Held to the purged-walk-forward rigor, the 18-candidate canonical list collapses: the flagship “geopolitics zero-fee resolution-mechanic” thesis (C1) is empirically falsified by the UMA record, and the structural/latency arbs (neg-risk, logical-chain, cold-start, news-latency, score-lag, Pinnacle) are DEAD at this op’s speed (sub-100ms bots take ~73%). What survives:

• 3 viable NEW lanes — all pilot-small, default-OFF, must clear the purged harness: C3’ (Sports/Esports Liquidity-Rewards maker farming), C4-weather (maker-flip of the live weather signal), C-Tech (resolution-source mispricing on objective oracles).
• 2 research-more candidates gated on one cheap backtest each: C7 (UMA dispute-window fade), C15 (on-chain MM cancel-event signal).
• The weather fee-correction (above) — risk reduction on live capital.

The proven themes win precisely because they do NOT require winning a sub-100ms race. Everything latency/arb-shaped is dead — do not re-explore the dropped list (Section “Dropped & DEAD”).

PM Fee Structure V2 (verified mid-2026)

Polymarket is no longer zero-fee. Taker fee = Theta · C · p · (1 − p) where C = notional and p = price of the side bought. Fees vanish as p → 0 or p → 1 (the term p(1−p) → 0), which is why resolution-mechanic lanes that fill near the certain outcome (weather at p~0.9, crypto tails) bleed little. Maker rebates ~25% in most categories / ~20% in crypto. Effective 2026-03-30.

Category	Theta	Notes
Geopolitics	0	the only free category — full gross retained
Sports	0.03	lowest non-zero; has an LR pool + 25% rebate
Esports	0.03	same as Sports; has an LR pool
Politics	0.04
Finance	0.04
Tech	0.04	the C-Tech lane category
Economics	0.05	largest rebate pool (25%)
Weather	0.05	NOT free — same as Culture (the P0 correction)
Culture	0.05
Crypto	0.07	highest fee; rebate only ~20%

Fee math worked at the live regimes

• Weather @ p=0.90, $500:\ast \ast fee=0.05\cdot 500\cdot 0.90\cdot 0.10=\ast \ast$2.25 (~0.45%). @ p=0.96 → **$0.95\ast \ast .@p=0.30\to$5.25 (1.05%).
• Crypto tail @ p→0.10: fee = 0.07 · C · 0.10 · 0.90 — small, and shrinks further toward the tail (why the longshot lane survives the fee).
• Rebate is trivial at extreme p: Weather @ p=0.90 rebate = $2.25\cdot 25$0.56 — do not lean on it as a profit source.

This corrects the older “PM is free” assumption and supersedes the partial schedule recorded on cross-venue-pm-kalshi-arb by adding the missing Weather=0.05 / Culture=0.05 / Crypto=0.07 / Sports=Esports=0.03 rows.

Liquidity Rewards program — pool is in SPORTS/ESPORTS, not geopolitics

Corrects a common assumption

The Liquidity Rewards (LR) pool is concentrated in SPORTS / ESPORTS, NOT geopolitics. Geopolitics has NO LR pool. Any strategy that assumed “farm LR in the zero-fee geopolitics category” was pointed at the wrong place — this is what re-points C3 → C3’ below.

• Size: ~$5M+forApril2026\ast \ast ,tiedtoWorldCupbuild-up.Examples:\ast \ast$10k / EPL match split ~$2.8kpre-game/$7.2k live; ~$24k / Champions-League QF.
• Scoring: quadratic closeness-to-midpoint S(v,s) = ((v−s)/v)² · b, sampled once per minute (10,080 samples/epoch), per funded market. Rewards resting limit orders near the mid; far-from-mid quotes score ~0.
• Income source: PM’s incentive budget (a deliberate spend to deepen sports books) — there is no directional loser, the durable payer is PM. Secondary capture: retail takers crossing your resting quotes (you keep spread + the 25% rebate on their 0.03 fee).
• Decay caveat: practitioners report the lucrative LR meta is already partly competed down to a thin bonus, and the program is incentive-funded / time-limited (PM can change caps overnight).

UMA oracle resolution reality — kills a whole edge family

"Literal/stricter rule wins" is NOT a reliable mispricing in discretionary categories

UMA resolves high-volume geopolitics markets to the HEADLINE / SPIRIT of the market, not the literal written rule, and the DVM vote is whale-dominated. Canonical example: the $120M Iran “permanent peace” market resolved 99% YES AGAINST its explicit “permanently cease” rule. Sibling Iran markets resolved inconsistently. You cannot build a leakage-clean walk-forward on a discretionary, politically-influenced resolver.

This empirically falsifies the C1 “geopolitics rules-vs-headline mismatch” thesis and the resolution-mechanic edge family in discretionary categories. Resolution-source edges only work where the oracle is OBJECTIVE / DETERMINISTIC / non-discretionary:

• FDA PDUFA decision date (on the FDA calendar months ahead)
• FAA launch-license / waiver grant (public docket)
• Chainlink price print (the existing crypto lane’s oracle)
• Airport ICAO station reading (the weather lane’s oracle)
• A leaderboard URL snapshot at a timestamp (AI-benchmark markets)

For Agents — the UMA proposer-win rate trap

The often-cited UMA proposer-win ~98.5% is the UNCONDITIONAL rate (proposals that resolve cleanly and are never disputed). The rate CONDITIONAL on escalation to a DVM token-holder vote is unknown and structurally lower — that ~0.5% set is selected for ambiguous/contested/whale-targeted markets where outcomes are decided by vote concentration, not correctness (Iran “permanent”; Ukraine minerals flipped by a ~25%-voting-power whale; Fort Knox two-address majority). Any EV claim built on 98.5% is wrong. This is the binding flaw in C7 below.

The 3 viable NEW lanes (all pilot-small, default-OFF)

All three must clear the purged walk-forward harness (purged-backtest-gate-no-edge) — beat ~+2.75% random net of a 1.5–2× cost multiplier — before any capital. None is a re-proposal of a rejected edge.

C3’ — Sports/Esports Liquidity-Rewards MAKER farming (re-pointed)

Non-directional maker income farming the LR pool in Sports/Esports (Theta 0.03 + 25% rebate), NOT geopolitics (which has no pool). Quote two-sided near the mid in the pre-game window (2+ hrs before kickoff — slow info flow, no score events = materially lower adverse selection than the live in-game regime).

• Capital: $5–20k justified — the one lane where the upper bankroll fits, because LR income is ~linear in resting size and the documented lucrative regime needed real capital. Breadth across many simultaneous games multiplies eligible markets.
• Distinct from swisstony MM-mirroring: we are NOT crossing a maker’s spread as a taker — we ARE the resting maker collecting the incentive + spread. (The swisstony note proves sports MM edge accrues to the resting maker; mirroring him as a taker pays away the ~1% that IS the strategy.)
• MAKE-OR-BREAK: net edge after adverse selection is UNPROVEN — could be ~0. The gate is the adverse-selection measurement: for every simulated fill, log fill price vs next-15-min and next-60-min mid; require reward income + spread + rebate > AS loss at 2× cost. Negative control: quote outside the scoring band → reward must go to ~0.

C4-weather — Maker-flip of the LIVE weather signal ONLY

Maker-first / taker-fallback overlay on the existing live weather taker signal. Once the afternoon obs confirms direction, post a resting bid at mid-minus-half-spread; fall back to taker if unfilled in 60–90s. The maker order is informed by the realized obs, so “we want to be hit” genuinely holds (closer to informed than adversely-selected).

• Realistic gain: +10–20% over the taker baseline — NOT the +30–50% pitched. The fee-elimination is only ~$1–2/$500 because real fills land at p~0.85–0.96 (not the p=0.30 the canonical math assumed). The real prize is spread capture (~0.6–0.8c/share).
• LR leg is illusory here: weather is not LR-funded; even if it were, fills at p>0.90 need double-sided quoting to score and a 60–90s order accrues ~0–2 of 10,080 samples.
• DROP the crypto-1b and copy-entry maker sub-variants — they are structurally negatively-selected: a resting directional bid fills precisely when the thesis is failing (crypto: price falls through your tail bid en route to 0; copy: price retraces below the whale level only because convergence is failing). Their own edge research confirms the filled subset is negatively selected.
• Gate: fill-rate >40% AND filled-subset win-rate within ~3–5pp of the unconditional taker baseline AND net improvement >+15% at 1.5× cost. Negative control: set the offset unreachable (−10c) → confirms delay alone doesn’t degrade outcome.

C-Tech — Resolution-source mispricing in the Tech category (objective oracles ONLY)

The same proven theme as weather/crypto (settlement-mechanic-vs-crowd) extended into Tech (Theta 0.04) — a category with zero prior coverage. Markets resolve on a NAMED authoritative source that diverges from the media-headline anchor; the crowd prices the hype cycle, the rule pins a deterministic public-record fact.

• Screen HARD to objective/deterministic/non-UMA-discretionary oracles only: FDA PDUFA date, FAA launch-license grant, AI-leaderboard snapshot. Explicitly EXCLUDE any Tech market with a subjective/UMA-judgment resolution (“will AI be AGI”, “will X company succeed”). This is the exact property C1 lacked — and why C-Tech is distinct from the dead C1 (which died because geopolitics resolution is discretionary/whale-voted).
• Capital $1–3k, breadth-scaled across drug-approval / launch / benchmark sub-pipelines. Edge ~3–8c per genuine mismatch; low fee (0.04) keeps modest gaps net-positive; fee vanishes near p→0/1.
• Reuses the C1/C2 resolution-rule parser (the one genuinely reusable artifact from the dead candidates — build it for this). NEW free feeds: FDA PDUFA calendar, FAA launch-license docket, AI-leaderboard scrapers.
• Validation must route deep history via Gamma event dumps / on-chain reconstruction, NOT the Data API (3000-record cap, below). Negative control: shuffle source-divergence labels → edge must vanish. Need ≥30–50 qualifying deterministic-source markets to trust a deflated Sharpe; if the objective-source subset is too small, this drops to a backlog note.

The 2 research-more candidates (gated on one cheap backtest each)

C7 — UMA dispute-window fade

Novel, reuses the planned Polygon WSS log infra (subscribe to OptimisticOracleV2 Dispute events), needs no second venue. BLOCKED on a single backtest: pull 18–24mo of on-chain OptimisticOracleV2 Dispute/DVM events and compute proposer-win conditional on DVM escalation (the thesis wrongly uses the ~98.5% unconditional rate — see the UMA callout above). Deploy a default-OFF pilot ONLY if that conditional rate is reliably >80% in a cleanly-screenable, whale-light sub-stratum. Sample size is the binding constraint: ~0.5% DVM rate minus the screen yields ~0–3 actionable/month, geopolitics-clustered — likely too few for a deflated-Sharpe trial-ledger to ever certify.

C15 — On-chain MM cancel-event / liquidity-concentration signal

Reads on-chain CANCEL events — orthogonal to the live copy lane (which reads taker FILLS, a later point in the info chain). Reuses the planned WSS build + the existing cohort as the identified-MM set. The swisstony note explicitly endorses this as the legitimate non-mirroring use (“best use: a flow / liquidity-concentration signal”). BLOCKED on: validating the CLOB v2 Cancel event signature (on-chain parse may also lag the WSS), then a backtest of forward price move after single-market >40% one-sided withdrawals with a no-concurrent-rebalancing filter. Negative control: shuffle the withdrawn side → edge must collapse. Keep to geopolitics (Theta=0) to retain full gross. Risk: routine cross-market rebalancing looks identical to informed single-market withdrawal (high false-positive rate).

Dropped & DEAD — do NOT re-explore

Structural / latency arbs are DEAD at this op's speed

These are not “untested” — they are verified dead at sub-second-but-not-sub-100ms latency. Their proven-theme cousins win precisely because they don’t require winning a sub-100ms race.

Dead candidate	Why it’s dead (durable)
C1 Geopolitics rules-vs-headline	UMA resolves to headline/spirit, whale-voted (Iran $120M → 99% YES vs the literal “permanent” rule). Falsified, not slow.
C5 Neg-risk convert / cross-side arb	~2.7s avg windows, ~73% captured by sub-100ms bots, ~0.3% median spread. Partial fill = naked exposure (convert is non-atomic, runs only after separate taker fills). “Cross-side arb no longer profitable in 2026.”
C6 Conditional-chain (combinatorial) arb	Definitive 2024-election study: ~$95Ktotalrealized=0.24$100/opportunity at <$2K volume — in the zero-fee era. Two-leg fees now invert most Politics cases.
C10 New-listing cold-start	PM markets open with an EMPTY book; the opening mid is set by snipe-bots (commercialized). The MOST bot-contested instant, not an unwatched one.
C8 Breaking-news latency	Free RSS ~15min late vs ~150ms PM news-bots + in-product headline display. Zero as a profit center. BUT the feed IS valuable as an adverse-selection kill-switch for the maker lanes (tolerates false positives, weak latency need) — fund it there.
C14 Sports in-game score lag	ESPN is the SLOW feed (20–30s behind broadcast); primary game-API bots already won; backtests give false positives (can’t model fill-time adverse selection → a pilot is actively dangerous).
C11 Pinnacle-vs-PM sports arb	Already a productized dominant bot archetype; edge only in thin long-tail books where impact + spread erase it.
C2 / C12 / C13	Macro series-ID is in the title (winner pre-priced 96–100c); awards & box-office are un-validatable (N≈3–8/yr).

Set-aside (confirmed shelved): cross-venue Kalshi-PM arb (reopen bar unmet — see cross-venue-pm-kalshi-arb), YES+NO/CTF-merge spread, high-velocity-crypto maker quoting (worst-AS venue), reality-TV / Nobel / generic same-event multi-source.

Infra facts surfaced (verified against the SDK/crate)

For Agents — execution-path realities for any maker lane

• position_manager is FAK / taker-only today. The CLOB v2 Rust SDK exposes limit_order() + OrderType::GTC/GTD and is_order_scoring() — so resting orders are a wiring job, not a platform gap. See position-manager-interface-design and position-manager-mode-control-implemented.
• cancel() exists in the crate but is currently #[allow(dead_code)] — it must be exercised and tested for sub-minute reliability before ANY maker lane ships (a maker-first/taker-fallback timer depends on reliable cancel).
• Data API trade-history offset is hard-capped at ~3000 records. Every deep-history validation (C3’, C-Tech, C7, C15, C17) must route via Gamma event dumps or on-chain Polygon log reconstruction, NOT Data API trade history. Bake this into the harness. (Consistent with the 3500/wallet ceiling noted in pmv2-two-tier-trader-scoring.)
• Portfolio-level correlated-resolution-timing risk is currently UN-MODELED: all CPI markets resolve together; all daily-high weather markets resolve at midnight; a single regional war hits many geopolitics positions at once. Model correlated capital-lockup + correlated PnL tail at the portfolio level — currently underestimates tail risk.

The salvaged-C8 news feed and the C7/C15 WSS work both reuse the existing on-chain substrate (watcher-onchain-gap-backfill, cohort-algo-component).

Other gaps flagged (research backlog, lower priority)

• Overround fade (sum YES > 1.00 on non-neg-risk multi-outcome groups) — symmetric twin of C5, a byproduct of the C5 scanner; same latency/leg-out caveats; harvest in zero-fee geopolitics groups.
• Kalshi-as-directional-signal (NOT arb) — use Kalshi’s price as an external calibration anchor (analog of Pinnacle-for-sports) on rule-diffed politics/econ events; read-only, no second funded venue. Requires a rule-diffing step.
• Maker-rebate stacking in HIGH-fee categories (Economics 0.05 / Crypto 0.07 have the largest rebate pools) — rest on the near-certain winning side at p→0/1 (low AS), harvest rebate + spread. Modest, rebate-only.
• Late-window cheap-tail generalization of the proven crypto-1b phase×price interaction to Economics/Weather settlement windows — pure backtest, no new infra.
• Rebate-adjusted Kelly — size all maker lanes with a rebate-adjusted edge (currently treated as a passing bonus, never quantified as a sizing factor).

Methodology provenance

A 26-agent Claude workflow: 9 edge-family scouts produced 53 raw candidates → merge/dedup to 18 canonical (12 set aside) → adversarial verification of the top 14 → completeness critic → synthesis. Funnel: 53 → 18 → 14 verified / 12 set aside. All fee/program/order-type/UMA facts verified via web search against PM docs + 2026 sources (mid-2026); execution-path facts cross-referenced against the SDK and the vault swisstony note.

Related

• purged-backtest-gate-no-edge — the ship bar every new lane must clear (beat ~+2.75% random, 1.5–2× cost)
• swisstony-hft-sports-market-maker — sports MM edge accrues to the resting maker; underpins C3’ and C15
• cross-venue-pm-kalshi-arb — prior PM fee-schedule fact (now extended here) + shelved Kalshi arb
• position-manager-interface-design — PM execution crate (FAK-only today; GTC/GTD + cancel() seam)
• position-manager-mode-control-implemented — the off/dry/live kill switch every pilot lane defaults to OFF behind
• cohort-algo-component — the on-chain WSS substrate + cohort that C7/C15 reuse
• watcher-onchain-gap-backfill — Polygon eth_getLogs / WSS log infra reused by C7/C15
• longshot-strategy — the proven crypto cheap-tail theme C-Tech generalizes
• autotrade-signal-type — the live weather/crypto signal sources the fee-correction applies to
• xref-consensus-rejected-fresh-copy-wins — none of the new lanes re-propose this rejected consensus edge