Kalshi Science and Technology — Mispricing Audit

Report date 2026-06-26 · Horizon 45 days · Book size $1,000 · Methodology v3 (conviction-weighted, EV-agnostic, 15% cluster cap)

Verdict: no deployable edge this week. 0 PICKS 8 REJECTS 100% CASH

The Science & Technology slate is dominated by three liquid clusters — AI-model release timing, GPU compute-price ladders, and SpaceX launch counts — and each fails a v3 screen for a $1,000 book. The AI-release ladder repriced violently in the last ~24 hours on OpenAI's June miss (edge already eaten). The compute ladders settle off an opaque third-party index I cannot independently verify. The one near-locked favorite (SpaceX ≥14 June launches) has a 5¢ spread that forces a negative-edge cross. Per v3, a near-empty book in an efficient week is the correct answer — padding it with low-conviction filler is exactly what lost money in the resolved track record. So: zero deployed, full cash, and the work is in the rejects below.

1 · How this was researched

Selection mode: category-match. $CATEGORY = "Science and Technology" matched trading_events.category directly and returned 171 active markets closing within 45 days — a healthy population, so no theme/keyword fallback was needed.

Pipeline:

Live prices/quotes captured 2026-06-26 ~15:00–15:30 UTC from the Kalshi public API; snapshot fields from the DB mirror. Probabilities are subjective estimates, stated explicitly per candidate as a calibration datum.

2 · Markets reviewed (Stage-2 survivors)

The 12 markets that cleared the mechanical Stage-1 filter, with the live read at pick time. “Eff. entry” is the price you'd pay for the side a trade would take. Every row resolved to a reject — see §4.

TickerMarketYES bid/askVol 24hOIClusterDisposition
KXGPT-OPENB-26JUL03GPT-5.6+ before Jul 315 / 1620,37014,289gpt-5.6-releasereject — edge eaten
KXGPT-OPENB-26JUL10GPT-5.6+ before Jul 1039 / 412,9893,344gpt-5.6-releasereject — 24h repricing
KXGPT-OPENB-26JUL31GPT-5.6+ before Jul 3168 / 696,14819,602gpt-5.6-releasereject — no depth
KXH200Q-26JUN30-4.390H200 price > $4.39 by Jun 3090 / 937,80213,129gpu-compute-pricereject — source
KXH200Q-26JUN30-4.590H200 price > $4.59 by Jun 3091 / 933,75413,086gpu-compute-pricereject — source
KXH200Q-26JUN30-4.790H200 price > $4.79 by Jun 3091 / 923,6485,275gpu-compute-pricereject — source
KXH200Q-26JUN30-4.990H200 price > $4.99 by Jun 3089 / 9313,77420,546gpu-compute-pricereject — at-the-money
KXH200Q-26JUN30-3.190H200 price > $3.19 by Jun 3091 / 943,7425,681gpu-compute-pricereject — source
KXH200Q-26JUN30-3.390H200 price > $3.39 by Jun 3093 / 964,5006,124gpu-compute-pricereject — source
KXH200Q-26JUN30-3.990H200 price > $3.99 by Jun 3094 / 952,4006,082gpu-compute-pricereject — source
KXH200Q-26JUN30-5.590H200 price > $5.59 by Jun 303 / 64,3544,671gpu-compute-pricereject — priced tail
KXSPACEXCOUNT-26JUN-13>13 SpaceX launches in June (≥14)89 / 947,84346,961spacex-june-launchesreject — spread

The H200 rungs share one underlying (the Jun-30 H200 index print) — they are one cluster, not nine independent edges. Adjacent compute ladders (A100, H100, B200, RTX 5090) were screened as the same class and carry the identical settlement-source problem. AI-leaderboard markets (“Best AI / Top Model / Top Coding AI” in June) were already resolved-in-practice (Anthropic/Claude leading, legs pinned at 0–1¢ or 97–99¢) and dropped at Stage 1. Gemini 3.5 Pro, GPT-6, New Glenn, Starship-from-Florida, FDA drug approvals, and nuclear-criticality markets were all dropped at Stage 1 for volume24h well below 1,500.

3 · Top picks — detail and thesis

No contracts cleared the v3 screen. There are zero numbered picks this run. This is a deliberate output, not a gap in coverage: see the three structural reasons in the lede and the per-candidate rejects in §4. The recommended action on a $1,000 book is to hold 100% cash and revisit when (a) the AI-release ladder settles into a stable price the market and I disagree on, or (b) a compute market migrates to a verifiable settlement source.

Two candidates are genuinely close and sit on the watchlist with explicit triggers — see Execution notes in §5.

4 · Recommended $1,000 portfolio

Allocation

PositionCost% of capitalMax payoutEV
Cash reserve$1,000100%$1,000$0
Deployed (contracts)$00%$0$0
Total$1,000100%$1,000$0

Total deployed $0 · Cash held $1,000 (100%) · Max book payout $1,000 · Blended EV 0% · Dollar edge $0. Per v3 the cash floor is ~10% of capital; here the whole book is the floor because nothing else cleared.

Cluster exposure

ClusterDeployed cost% of capitalCap (15%)Status
gpt-5.6-release$00%$150within cap
gpu-compute-price-jun30$00%$150within cap
spacex-june-launches$00%$150within cap
gpt6-release$00%$150within cap

Conviction exposure

TierPer-pick capDeployed# picks
HIGH$150 (15%)$00
MEDIUM$70 (7%)$00
LOW$30 (3%)$00

Risk profile

Execution notes & watchlist

5 · What I rejected and why

Every Stage-2 survivor that didn't make the cut is logged here and machine-tracked to settlement. If these rejects resolve as well as picks would have, the selection screen adds nothing — that's a finding worth knowing either way. Stage-3.5 entry-band cuts are called out explicitly.

KXGPT-OPENB-26JUL10 · would take NO · would-be MEDIUM eff. entry ~61¢ · favorite band

My P(NO, i.e. no GPT-5.6 by Jul 10) = 70% · Market-implied P(NO) ≈ 59% · Cluster gpt-5.6-release

The cleanest directional thesis on the board — fade near-term GPT-5.6 release after OpenAI's heavily-anticipated June window collapsed (prediction-market odds for late-June fell from ~83% to ~18%). The problem is timing: this market already did the move. YES repriced from ~85¢ to ~41¢ in the last ~24 hours on the June miss (verified in the 5-min snapshot history). Buying NO now is chasing a same-direction move that already happened.

Reject: Stage-3 48h-move rule — the thesis is freshly priced in; my 70% vs the market's 59% is opinion, not an information advantage over a price that just absorbed the exact news I'd be trading on.

KXGPT-OPENB-26JUL03 · would take NO · would-be deep favorite eff. entry ~85–90¢

My P(NO, no release in ~6 days) = 90% · Market-implied P(NO) ≈ 84% · Cluster gpt-5.6-release

Directionally I agree the market won't ship in a specific 6-day window with no imminent signal. But after the same 24h repricing, NO already costs ~85¢, and the deployable depth sits at 0.88–0.90 effective (the deep yes-bid wall at 10–13¢), leaving ≤2¢ of residual edge.

Reject: deep favorite, edge eaten — fillable price ≈ true probability. Not worth the capital lock.

KXGPT-OPENB-26JUL31 · would take YES · would-be MEDIUM eff. entry ~69¢ · favorite band

My P(YES, GPT-5.6 by end-July) = 71% · Market-implied P(YES) ≈ 68% · Cluster gpt-5.6-release

July is the consensus release window, and the 91¢→69¢ repricing may have slightly overshot. But the top-of-book is ~1 contract at the 0.69 ask, then 0.73 / 0.79 — sizing 100+ contracts means >5¢ of slippage, which erases a 1–3¢ edge. Also freshly repriced.

Reject: Stage-3 slippage screen — can't size without paying ≥3¢. Watchlist-only via a patient 0.66–0.67 limit.

KXH200Q-26JUN30-4.390 (representative of the H200/compute ladder) · would take YES · deep favorite eff. entry ~93¢

My P(YES) ≈ 96% (conditional on a clean print) · Market-implied ≈ 91% · Cluster gpu-compute-price-jun30

This is the most tempting trap: deep book (thousands of contracts), a 4-day horizon, and a threshold ($4.39) 13% below the implied spot, so the GPU rental price physically can't fall through it in time. But the market settles off the Ornn index (ornnai.com), whose live H200 value I could not independently verify (the site is JS-loaded and routes to a data portal I can't read). The implied spot (~$5.0–5.1, inferred from the cliff to the $5.19 rung at ~3¢) is the market's number, not mine — so buying here is selling tail-insurance against an index I don't understand, with no information advantage.

Reject: Stage-3 source/definitional screen — settlement source not independently verifiable; no edge beyond the market's own implied spot. The entire A100/H100/H200/B200/RTX 5090 ladder class is rejected for the same reason.

KXH200Q-26JUN30-4.990 · would take YES/NO · coin-flip near threshold at-the-money

My P(YES) ≈ 60% · Market-implied ≈ 91% (but threshold ≈ implied spot) · Cluster gpu-compute-price-jun30

Unlike the lower rungs, $4.99 sits right at the implied settlement spot — small index noise flips it. That makes it effectively a coin-flip on the index's exact print, dressed as a 91¢ favorite.

Reject: Stage-3.5 — at-the-money to an unverifiable source; flip risk on settlement noise, no defensible read.

KXSPACEXCOUNT-26JUN-13 · would take YES · would-be HIGH eff. entry 94¢ ask · deep favorite

My P(YES, ≥14 June launches) = 93% · Market-implied ≈ 89–91% · Cluster spacex-june-launches

The strongest mechanical case on the board: by Jun 24 SpaceX had flown the 7th of 8 planned June Vandenberg launches plus ~6 Florida launches, so ≥13 are already in the books with the cadence (~one every 2–3 days) pointing to ≥14 by Jun 30. The market's own ladder agrees — modal final count is exactly 14. But the quote is 89 bid / 94 ask: crossing to 0.94 for a ~93% event is negative edge, and a positive entry (≤0.92) requires a patient limit into a gapped book with no guarantee of a fill.

Reject: Stage-3 slippage screen — not deployable without ≥3¢ of spread cost. Moved to Watchlist B with a ≤0.92 limit trigger. (Note: the higher rungs ≥15 are already active picks from a prior run.)

KXGPT-OPEN-26JUL01 · would take NO · priced/illiquid GPT-6 by Jul 1

My P(NO) ≈ 98% · Market-implied ≈ 96–98% · Cluster gpt6-release

A full-integer GPT-6 jump before Jul 1 is ~nil and the market agrees (YES ~2–4¢). Thinly traded (vol24h ~545) and effectively fully priced as NO.

Reject: Stage-1 liquidity + no edge — correctly priced, nothing to add.

KXSPACEXCOUNT-26JUN-14 · would take YES · excluded already live

≥15 June launches ≈ 9% tail · Cluster spacex-june-launches

Listed for completeness: this rung is already an active pick in users' feeds from a prior run, so re-emitting it would duplicate the feed. Independently it's an unattractive ~9% tail.

Reject: already an active pick (duplicate-prevention rule).

6 · Sources

Data sources: Kalshi read-only DB mirror (trading_events, trading_markets, market_snapshots) for population, rules, and 5-min price/volume history; Kalshi public API for live orderbooks; web search/fetch for primary-source grounding. Methodology v3 — conviction-weighted & EV-agnostic sizing, 60–90¢ favorite entry band, 15% per-cluster cap.

Disclaimer: Probabilities are subjective estimates, not guarantees. Prediction-market contracts can resolve to zero and you can lose 100% of capital at risk. This is research, not financial advice. v3 picks are cleared for auto-copy and count toward the public track record — which is precisely why a $0-deployed, all-cash week is reported plainly when the slate is efficient, rather than padded to hit a count.