M02 · Talaria Revive
M02 · Forensic failure analysis

This asset failed. Was it the molecule — or was it the trial?

A failed Phase III is not one event. It is a mixture of failure modes, and only some of them indict the drug. Revive runs the autopsy: it decomposes a failure into six independent modes, isolates the ones a redesigned trial could actually fix, and reads whether the asset leans revivable. Where the dominant mode is a design flaw, the molecule usually still works — the trial did not.

Every case below is a hand-authored synthetic composite. No sponsor, no molecule, no NCT. No engine is invoked. The numbers are illustrative and exist to show the shape of the analysis, not to predict any real asset.

Case file

Pick a failed asset to autopsy

Schematic · illustrative · synthetic composite · not a Talaria prediction. No sponsor, no molecule, no NCT. No engine is invoked.

SYN-COMPOSITE-DERM-2031
Molecule · under examination
01 · Failure decomposition

Six independent modes

Each mode is its own sigmoid, scored independently. A single trial can light up several at once, so these do not sum to 1 — this is a forensic mixture, not a softmax. The dominant mode is badged primary; the runner-up is marked secondary.

Failure-mode classification independent sigmoids · synthetic composite

Schematic · illustrative · synthetic composite · not a Talaria prediction.

02 · Headline

Lead with the redesign probability

Two numbers, deliberately kept apart. The simulated revival probability is a proper probability — the chance a redesigned trial clears its primary endpoint. The composite revival index is a product of sub-1 forensic factors, so it is small by construction; it is a relative signal, never a percentage chance. Reading the index as a gauge is the mistake this layout exists to prevent.

Simulated revival probabilityDT-SC
0.63
± 0.10 uncertainty

Simulated probability the redesigned trial clears its primary endpoint. This is the human headline — the number to read first.

Schematic · illustrative · synthetic composite · not a Talaria prediction.

Composite revival index
0.101
CI [0.00, 0.25]

Relative forensic signal — not a % chance of revival. A product of correctability factors, each below 1, so a strongly-revivable asset still lands near 0.05–0.15.

Do not render on a 0–100% revival gauge. It is a ranking signal, not a probability.

03 · Correctability

Which failures a redesign can actually move

Only modes that fired — scored probability at or above 0.40 — are candidates for correction. Each carries a correctability prior (how fixable the category is in principle), a lever score (how much the modifiable features move it), the model's simulated outcome, and the resulting final forensic signal.

Correctability · firing categories only final = prior × prob × lever × (1 − simulated_outcome)
Correctability per firing failure category
Category Prior Prob Lever Sim outcome Final signal Most relevant modifiable features

Schematic · illustrative · synthetic composite · not a Talaria prediction.

04 · Revised design

The modification bundle

The top modifiable features across the firing categories, with the expected lift each carries in this schematic. Complexity and cost premium describe what it takes to run the redesigned trial.

Revised-design modification bundle · top 5 feature · category · expected lift
Revised design modification bundle
#FeatureCategoryExpected lift

Schematic · illustrative · synthetic composite · not a Talaria prediction.

05 · Read

The verdict

Verdict read · presentation-layer

A presentation-layer read, derived for the reader — not a spec output and not a ranked recommendation. The levers are shown unranked and unvalued; ranking them would require a reconciled canonical-metric record, and there is none behind this page.

Founding cohort

Revive reads the autopsy the record leaves behind

This bench is a fixture. The Revive engine is not — it decomposes real failures into their modes and separates the correctable trial from the failed molecule. Talaria is pre-deployment. We are opening a founding cohort ahead of it.

Join the founding cohort