Pre-clinical intelligence markets

Independent probability.
Before the clinic.

A liquidity-sensitive prediction market for AI-generated therapeutics. Applies LS-LMSR with adaptive Bayesian seeding to preclinical milestone contracts — producing continuous, expert-calibrated probability estimates before capital is committed to IND-enabling studies.

AI platforms generate candidates faster than any evaluation infrastructure can validate them.
$2–5M

Cost per IND-enabling program

Committed before independent probability assessment exists.

~90%

Of AI-generated candidates

Receive no systematic external evaluation before internal triage decisions.

0

Independent price signals

Exist for pre-IND assets. Internal models evaluate their own output.

Every internal triage decision relies on the same computational models that generated the candidates. There is no adversarial check, no calibration against external judgment, no market mechanism to surface systematic overconfidence. Pre-Clinical Prediction Markets is that mechanism.

Mechanism deep-dive

Formal Market Architecture

01 / LS-LMSR core

Liquidity-Sensitive LMSR

The LS-LMSR (Othman et al., 2013) replaces the fixed liquidity parameter b of standard LMSR with a volume-adaptive parameter. Liquidity depth grows automatically with cumulative trading volume — early markets are price-sensitive (high reward for early conviction), mature markets develop institutional depth. The α parameter is derived per-asset from the oracle-attested confidence score, calibrating market sensitivity to the reliability of the underlying computational signal.

b(q) = α · Σqᵢ [liquidity grows with volume] C(q) = b(q) · log(Σ exp(qᵢ/b(q))) [cost function] pᵢ(q) = ∂C/∂qᵢ [marginal price = probability] α = 0.005 + (1 − confidence_score) × 0.075
02 / ABMM seeding

Automated Bioactivity Market Maker (ABMM)

Without an initial position, a market opens at 50/50 regardless of oracle signal — the cold-start problem. The ABMM solves this by placing synthetic stakes derived from oracle-attested computational scores (binding affinity, selectivity, IC50, literature signal). This provides a non-arbitrary opening price surface calibrated to the molecule's computational prior. The effective confidence combines computational and literature scores with a weighted blend, and seeds the initial YES/NO quantity vector accordingly.

effective_confidence = score_comp × 0.6 + score_lit × 0.4 q_abmm_yes = f(effective_confidence, α) q_abmm_no = f(1 − effective_confidence, α)
03 / ABMM retreat

Calibration-Weighted ABMM Retreat

As credentialed expert volume accumulates, the ABMM's influence retreats — transitioning the market from machine-seeded to human-informed pricing. The retreat function is parameterized as exponential decay (concave) rather than linear. Concave retreat is correct for two structural reasons: first, early credentialed trades carry the highest informational value (correcting the most error per trade) and should drive rapid initial retreat; second, thin specialty markets may never accumulate sufficient volume to fully exit ABMM dominance under a threshold design, requiring a residual floor for price stability.

Retreat is weighted by trader calibration score (Brier-based) rather than raw volume — a trade from an expert with a strong prediction track record drives more ABMM retreat than one from a newly credentialed participant. This makes the transition responsive to signal quality, not just signal quantity.

ldi_calibrated(t) = Σ (volumeᵢ × brier_scoreᵢ) [sum over credentialed trades up to time t] w(t) = exp(−λ · ldi_calibrated(t)) [ABMM weight at time t, w(0)=1, w(∞)→0] λ = log(2) / ldi_half [decay rate: w=0.5 when credentialed volume = ldi_half] ldi_half = expected credentialed volume at which expert signal equals ABMM prior strength (modality-parameterized) Effective ABMM quantities at time t: q_abmm_yes(t) = w(t) · q_abmm_yes(0) q_abmm_no(t) = w(t) · q_abmm_no(0)
Calibrated Expert Volume ABMM Influence 0 max 0 1 thin market floor naive proposed
04 / Incentive compatibility

Incentive-Compatibility Under ABMM Dominance

A market scoring rule is incentive-compatible if a trader's optimal strategy is to report their true belief. Under standard LMSR this holds by construction. The ABMM introduces a distortion: its large initial synthetic position makes the market expensive to move early, potentially creating incentives for credentialed experts to underreport their true belief (partial trade is cheaper than full correction) or strategically delay (waiting for ABMM retreat reduces the cost of future trades).

The exponential retreat function addresses both distortions. Rapid early retreat reduces the cost of truthful correction precisely when expert signal is most valuable. The calibration weighting ensures that high-Brier experts — whose corrections are most reliable — drive the most retreat, aligning incentives with signal quality.

The formal condition for ε-incentive-compatibility requires that the ABMM's residual weight at any time t satisfies:

w(t) · q_abmm ≤ δ(ε, α, p*) where: ε = maximum tolerated belief distortion α = per-market liquidity sensitivity parameter p* = expert's true belief δ = tolerance bound (tightest when p* is far from p_abmm)
Open question Closed-form derivation of optimal λ as a function of (α, confidence_score, modality). Connection to Roughgarden & Neyman (2023) on proper scoring rule aggregation and Roughgarden & Schrijvers (2017) on selfish expert elicitation.

Technical foundation

The Analysis Gap & Market Architecture

A — The Problem

The Analysis Gap

Generative drug discovery platforms — Recursion, Insitro, Terray — now produce candidates faster than any evaluation infrastructure can validate them. Each molecule carries a probability distribution over downstream clinical success, but virtually none receives the independent analytical infrastructure afforded to traditionally-derived compounds. Every internal triage decision relies on the same computational models that generated the candidates. There is no adversarial check, no calibration against external judgment, no mechanism to surface systematic overconfidence.

B — LS-LMSR

Liquidity-Sensitive LMSR

The LS-LMSR (Othman et al., 2013) replaces the fixed liquidity parameter of standard LMSR with a volume-adaptive parameter. Liquidity depth grows automatically with cumulative trading volume — early markets are price-sensitive (high reward for early conviction), mature markets develop institutional depth. The α parameter is derived per-asset from oracle-attested confidence scores, calibrating market sensitivity to the reliability of the underlying computational signal.

b(q) = α · Σqᵢ C(q) = b(q) · log(Σ exp(qᵢ / b(q))) pᵢ(q) = ∂C/∂qᵢ α = 0.005 + (1 − confidence_score) × 0.075

α controls price sensitivity per trade. High-confidence molecules (score = 0.9) receive α = 0.0125 — the market is tight and resists large swings. Low-confidence molecules (score = 0.3) receive α = 0.075 — early trades carry high price impact, rewarding early correct conviction.

C — ABMM & Retreat

ABMM Seeding & Calibration-Weighted Retreat

Without an initial position, a naive LMSR opens at 50/50 regardless of oracle signal — the cold-start problem. The Automated Bioactivity Market Maker (ABMM) seeds each market with synthetic stakes derived from oracle-attested computational scores (binding affinity, selectivity, IC50, literature signal), providing a non-arbitrary opening price surface.

effective_confidence = score_comp × 0.6 + score_lit × 0.4 q_abmm_yes(0) = f(effective_confidence, α) q_abmm_no(0) = f(1 − effective_confidence, α)

As credentialed expert volume accumulates, the ABMM retreats via an exponential decay function weighted by trader calibration scores (Brier-based). Concave retreat is correct for two structural reasons: early credentialed trades carry the highest informational value and should drive rapid initial retreat; thin specialty markets may never accumulate sufficient volume to fully exit ABMM dominance, requiring a residual stabilizing floor.

ldi_calibrated(t) = Σ (volumeᵢ × brier_scoreᵢ) w(t) = exp(−λ · ldi_calibrated(t)) λ = log(2) / ldi_half

D — DeFi Primitives

Downstream DeFi Primitives

A working price oracle for pre-clinical assets enables three DeFi primitives that have not previously existed in drug discovery. These are downstream consequences of solving the oracle problem — not speculative extensions.

Milestone-Gated Funding Pools

Partners stake capital into a molecule's M1 market. If M1 resolves YES, capital automatically routes to the M2 pool. If NO, capital returns pro-rata. The prediction market becomes the funding gate — replacing the $2–5M IND commitment with a staged, market-priced capital release mechanism.

Pre-Clinical Asset Derivatives

Once a continuous probability estimate exists on a molecule, options become possible. A floor contract pays out if a candidate drops below a threshold IND probability — pipeline insurance priced by the market, not by actuarial tables constructed from historical base rates.

Computational Model Staking

AI labs stake their models rather than specific molecules. If a model's output systematically outperforms market priors across resolved milestones, the model earns calibration-weighted returns. This creates a continuous public benchmark for generative drug discovery models — currently unavailable in any form.

These three primitives are downstream of one architectural prerequisite: a working price oracle. The LS-LMSR + ABMM stack is that oracle.

Live demo

9 live markets.
Running now.

The LS-LMSR Platform dashboard runs live LS-LMSR pricing across 9 oncology targets including ALK, BRAF, BTK, CDK4, CDK6, and EGFR. Real b-values. Real trade history. Real-time price updates.

ALK b=19.77 BRAF b=11.51 BTK b=11.24 CDK4 b=12.41 CDK6 b=12.10 MET b=10.49 PD1 b=8.91 EGFR b=10.10 KRAS G12C b=14.73
LIVE

ALK

b = 19.768
IND Submission78.4%
Phase 1 Success18.3%
Phase 2 Success2.4%
Approval0.9%

Considerations

Theoretical & Practical Considerations

Participation does not require structural disclosure. The obfuscation architecture allows full anonymization — scaffold class, target class, and mol-hash fingerprint only. The compound structure never leaves the oracle layer. Disclosure level is negotiated per participant. The platform supports everything from full-structure restricted markets to complete anonymization.
Internal scores optimize for within-campaign ranking. They cannot produce calibrated absolute probabilities comparable across platforms, targets, or time. The platform produces cross-platform, outcome-validated probabilities. The two signals are complementary — the value is specifically in cases where they diverge.
Participant identity is not disclosed on anonymized markets. Batch submission requirements prevent target-class triangulation. Restricted deep market prices are not published externally. Competitor employees can be excluded by the conflict registry.
Where material non-public information is a concern, affiliated participants may be restricted from trading markets tied to their own compounds under enforcement rules. Resolution authority — reporting attested outcomes — can be separated from trading exposure by design.
Four sourcing channels: nominated scientists, institutional CRO and academic center partnerships, structured credential verification, and calibration-based promotion from retail tier. Expert status requires demonstrated Brier score performance across resolved trades — earned continuously, not granted once.
The platform is one input into a capital allocation decision, not a decision rule. The value is in systematic divergence from internal conviction — a flag to investigate before committing IND-enabling resources. The mechanism surfaces overconfidence; it does not replace scientific judgment.
Publication tiers can be chosen to match confidentiality needs. Fully confidential markets are available. Anonymized shadow markets apply delayed publication windows. Exposure to external signal is managed by design, not assumed away.