Diagnosis Biotech

Part of the cognition series. Pre-registration, data, and code: github.com/kimjune01/universal-diagnosis

Three predictions. Three biotech companies. One framework. Head-to-head with Martin Shkreli.

The idea

Every biotech company is a learning loop for its molecule. The drug goes through trials. The company reads the results, decides what to change, and writes those changes back: new trial design, new endpoints, new manufacturing. That’s the backward pass. The Natural Framework calls it the consolidate stack.

Most analysts diagnose the drug: will it work? That’s the forward pass. They look at a snapshot. This trial, this endpoint, this p-value. The framework diagnoses the trajectory: is the company learning? Is the same failure recurring?

Two companies can have identical Phase 3 results and opposite outcomes, depending on whether the company’s backward pass is working.

But a functional consolidate stack is necessary, not sufficient. It also needs cache (cash) and runway (time). A company that reads its failures perfectly and knows exactly what to change is still dead if it can’t fund the next iteration. Cash is the timer on a broken forward pass. The consolidate stack decides what to do next; the cash decides whether it gets to.

The evidence

We diagnosed five companies from Shkreli’s public positions. Two retrospective (outcome known), three prospective (outcome unknown). Four parallel search agents per company, two independent merges, mechanical scoring. Pre-registered before any diagnosis.

Run 0: Retrospective (demonstration)

CAPR (Capricor Therapeutics) — Shkreli shorted. Said HOPE-3 “will not work.”

The temporal graph tells a different story. Four iterations: ALLSTAR (MI, failed) → HOPE-1 (DMD, pivoted) → HOPE-2 (IV, signal) → HOPE-3 (Phase 3, confirmed). The company changed the indication, the delivery route, the dose, and the endpoints across four trials. The consolidate stack was functional. ALLSTAR’s failure class (wrong indication) didn’t recur.

Framework: dying_pivoted, PASS. Shkreli diagnosed a snapshot. The framework diagnosed a trajectory. (SOAP notes)

Result: HOPE-3 positive (PUL v2.0 p=0.029, LVEF p=0.041). Stock +440%. Shkreli admitted “bad call.” Framework correct.

QURE (uniQure) — Shkreli was long. Sold November 2025.

AMT-130 topline was positive (cUHDRS 75% slowing, p=0.003). But the FDA reversed its prior agreement on external controls — first privately (November 2025), then publicly, with officials calling the data “distorted” and “manipulated” (March 2026). Stock -84%.

The temporal graph shows why: FDA published guidance in October 2022 explicitly discouraging external controls for neurodegenerative gene therapies. uniQure built their entire BLA strategy on external controls anyway. When the FDA enforced its own published rules, management argued the FDA’s inconsistency and raised ethical objections to sham trials instead of pivoting.

Framework: dying_dying. Shkreli got out before the break. The framework would have flagged the risk earlier — the 2022 guidance was the first event in the temporal graph. (SOAP notes)

Run 1: Prospective (real predictions)

These are live. Outcome unknown. Published March 20, 2026.

SPRB (Spruce Biosciences) — BLA submission for TA-ERT by December 31, 2026.

The molecule is mature (7 years, 22 patients, BTD). The company is new — acquired it from bankrupt Allievex after its own lead program failed. Consolidate stack is fragile (going concern, $48.9M cash) but functional (reading FDA feedback, hiring commercial leadership).

Framework: dying_pivoted, PASS. Shkreli: bull. We agree on direction, disagree on reasoning. He says “drug will be approved” (forward pass). We say “company is executing despite stress” (consolidate stack). The risk: PPQ batch slip → cash crisis. The molecule has outlived two prior companies. (SOAP notes, diagnosis)

ATYR (aTyr Pharma) — New sarcoidosis trial announced by September 30, 2026.

EFZO-FIT Phase 3 missed primary (steroid dose reduction, p=0.33) but hit secondaries (KSQ-Lung p=0.048, composite steroid withdrawal + improvement p=0.020). The secondary signals suggest the drug has activity on quality-of-life endpoints; the primary endpoint may have been the wrong measure. FDA Type C meeting mid-April 2026.

Framework: dying_pivoted, PASS. Shkreli: bear (“really bad drug”). His critique was mechanism-level (“really bad drug”), but the secondary endpoints suggest the molecule has activity that the primary endpoint didn’t capture. The framework predicts the company will pivot because the consolidate stack is functional: honest read of the miss, constructive FDA engagement, no panic pivot. Inverse of QURE — same kind of FDA pushback, opposite management response. (SOAP notes, diagnosis)

INMB (INmune Bio) — CORDStrom MAA to UK MHRA by September 30, 2026.

Both programs missed primaries (CORDStrom EBDASI p=0.15 favoured placebo; XPro1595 Phase 2 ITT miss). Both pivoting to secondaries. BLA already slipped 12 months. Going concern. CEO transition mid-crisis.

Framework: dying_pivoted, FAIL. Shkreli: bear. We agree on direction but for different reasons. He says the company is worthless. We say the consolidate stack rewrites narratives faster than it ships filings. The temporal pattern — primary miss → secondary salvage → timeline slip — predicts another delay. Both merge instances agreed unanimously. (SOAP notes, diagnosis)

What the framework adds

An analyst looks at a drug. The framework looks at the company’s learning loop.

The predictions are binary and timestamped. The diagnosis involves judgment (LLM agents, human selection); the scoring is mechanical — check the source on the date, apply the pass condition. Full protocol.

Where we agree, we’re more confident

The blind-blind-merge operates at two levels here. The two merge instances that produce each diagnosis are truly blind to each other and to Shkreli — they see search agent outputs, not analyst positions. That’s real blindness. The comparison to Shkreli is weaker: he’s a public benchmark, not a merge partner. He can read this blog, and we selected targets from his positions. But the framework’s direction is determined before we check his call, so the convergence — when it happens — comes from independent reasoning on overlapping data through different lenses: he diagnoses the molecule, we diagnose the company.

At N=3, none of this is statistically meaningful. But the structure is interesting: INMB has independent convergence from different analyses. ATYR is the pure disagreement. If the framework is adding signal, it should show on ATYR.

The scorecard updates when catalysts resolve. ATYR and INMB by September 30, 2026. SPRB by December 31, 2026.

Method

Each diagnosis: four parallel search agents (two forward pass, two backward pass), different data sources, semi-blind. Top two selected, merged independently by two instances of GPT-5.4 into SOAP notes. If merges disagree, first merge is primary. Human reviews but doesn’t synthesize. Full protocol.

The temporal graph is modeled as a sequence-based dynamic graph (Peters et al., 2019). Each event is a public record with an archival date. The graph grows one event at a time. No pre-defined eras — the evidence builds the timeline.

The framework is the Natural Framework. The consolidate heuristic: check for trauma recurrence. count(similar_failures) > 1 = broken consolidate stack.

Hypothesis: vol mispricing from broken read_outcomes

If management systematically overframes ambiguous data, the market’s model of the company may be too narrow. When expectations are set by optimistic framing, ambiguous outcomes produce larger surprises. If this pattern holds, realized vol should exceed implied vol around catalyst dates for companies with broken read_outcomes.

Testable methodology:

1. For each company, identify catalyst dates (trial readouts, FDA actions, earnings with pipeline updates)
2. Pull 30-day implied volatility from listed options before each catalyst
3. Measure realized volatility in the 5-day window around the catalyst
4. Compute the realized/implied ratio per event
5. Segment by framework diagnosis: broken read_outcomes vs. functional read_outcomes
6. Test: do companies with broken read_outcomes have higher realized/implied ratios?

Qualitative evidence from this pilot:

• CAPR: management said “no significant deficiencies” at mid-cycle review (May 2025). CRL two months later. Stock cratered. Realized > implied?
• QURE: management said FDA aligned on accelerated approval (December 2024). FDA reversed eleven months later. Stock -49% in one day, eventually -84%. Realized > implied?
• ATYR: management framed Phase 3 miss around secondaries in the same press release. Stock -83% in one day. Realized > implied?

The directional prediction isn’t “stock goes up” or “stock goes down.” It’s “the surprise will be bigger than the market expects” — because the company’s own self-assessment is unreliable. That’s a long vol position, not a directional bet.

This is a hypothesis, not a finding. The backtest requires historical options data we haven’t pulled. But the framework generates the hypothesis mechanically — any company diagnosed with broken read_outcomes is a candidate for vol mispricing around its next catalyst.

Future signal: insider transactions

SEC Form 4 filings are dated public records. When insiders buy or sell before an 8-K, the read_outcomes pipe is functioning privately even if the public framing is spin. The divergence between insider behavior and management’s public narrative is itself a diagnostic signal — it tells you whether the company knows more than it says.

We didn’t incorporate insider transaction events in this pilot. But the temporal graph can absorb them: each Form 4 is an event with an archival date, a pipe (read_outcomes), and an implicit status (insider selling before bad news = read_outcomes functional internally, broken externally). Adding this signal to future rounds would strengthen the read_outcomes diagnosis without requiring any change to the framework.

Three bets. Clock’s running. Scorecard at github.com/kimjune01/universal-diagnosis.