02 · SCP Engine

Draft Zero of your SCP, delivered.
Every sentence traces.

The SCP Engine proposes the claims your scientific communication platform (SCP) should make, each with its verdict and its sources, and writes Draft Zero from them. Judgment and sign-off stay with your writers.

How it runs

Questions in. Claims with verdicts. A draft that traces.

Three steps from the library to the page. Each one opens to the evidence beneath it.

1 · Questions
The questions your platform has to answerAgreed with you at the start; the same questions the library was screened against. Written to the questions the library was built to answer.
2 · Claim Library
The statements the evidence supports, with a verdict on eachSupported, partially supported, contradicted, or insufficient evidence. Each source marked for what it does to the claim; the gap named where the evidence runs thin.
3 · Draft Zero
An initial draft in which every sentence tracesEvery sentence starts from a graded claim. A Word document with its reference list, ready for your writers to make it theirs.
Claim, graded obesity · illustrative
“Weight reduction is sustained through 68 weeks of continued treatment.”
Phase 3 RCT, 2022 68-week weight change vs placebo, primary endpointSupports
Pooled analysis, 2025 maintenance phase, pooled across trialsSupports
Real-world cohort, 2026 persistence lower outside trial conditionsPartially supports
Narrative review, 2024 background onlyContext
Verdict: supported · 4 sources · the real-world qualifier is carried into the draft
Illustrative
✦ Claim Library

A citation is not proof. We check that the source says what the sentence says.

A claim is a specific scientific statement: this therapy, this population, this comparator, this outcome. Most documents assume the link between a statement and its references. The engine establishes it. Every claim is graded against the library, each source marked for what it does to the claim, and the gap named where evidence runs thin. Your reviewers see the verdict before the sentence.

Verdicts supported · partially supported · contradicted · insufficient evidence.
On every claim the sources, what each does to the claim, the quoted evidence, and the gap if there is one.
Challenge any claim and it is rechecked against the library, with the full trace returned.
✦ Draft Zero

Every sentence starts from a graded claim.

Draft Zero is an initial draft in which every sentence traces: statement to claim, claim to evidence, evidence to source. It is a starting point, not a finished document. Your writers cut, reshape and argue with it. The trace follows every sentence they keep; anything they add can be checked the same way.

Arrives as a Word document with its reference list, for writing teams; an interactive edition for pharma and biotech.
Timing within days of the library.
Authorship the draft is a starting point; judgment and sign-off stay with your writers and reviewers.
Draft Zero · obesity · excerpt illustrative

Weight reduction is sustained through 68 weeks of continued treatment1, with gastrointestinal events predominating and consistent with the class2. Cardiovascular outcome data now support a claim the prior platform did not make3. 1 Phase 3 RCT, 2022 · grade High · supported 2 Pooled safety analysis, 2025 · grade High · supported 3 Outcomes trial, 2024 · grade High · supported · not in prior platform

Every sentence traces. Writers start here, not from a blank page →
Illustrative
The questions agreed at the start · obesity · illustrative
1What is the magnitude and durability of weight reduction?Clinical evidence · 9 topics
2What is the safety and tolerability profile, and how does it compare to the class?Clinical evidence · 4 topics
3What do cardiovascular outcomes add to the platform?Clinical evidence · 2 topics
4What is the burden of the disease the asset addresses?Unmet need · 4 topics
The same questions the library was screened against →
Illustrative
✦ Questions

Each claim is written with its evidence attached.

The questions decide what evidence is sought; the evidence decides what gets written. The engine writes to the questions you agreed, and only those. They are the questions the library was screened against, so every claim traces to a question, and every question to the papers that answer it. Change the questions and the library is re-screened; the draft follows. The draft cannot say more than the evidence base was built to support.

Agreed in one conversation at the start, written down, fixed before screening.
Changed later by agreement; the library and the draft are re-run against the new set.
Tested against general-purpose AI

Five drugs, five therapeutic areas, scored blind, claim by claim.

The same source materials given to three leading general-purpose AI models and to Posteros; outputs stripped of identifying marks and scored blind. A system that surfaces more of the evidence hands your writers a larger choice set.

Substantiated claims surfaced · best of three general-purpose AI models, best run
423
Posterossubstantiated claims across five drugs · 2.4× the comparator
175
Best of three general-purpose AI models, best runsame sources, scored the same way
What this does and does not show. A first measurement, not an RCT. The multiple varies by therapeutic area, from 1.4× to 4.2×, and the variation is part of the readout. No completeness claim: the figures compare whole systems against each other, never against a ceiling. Drug and model names are not given here.
Substantiated claims per drug · five therapeutic areas · scored blind, claim by claimPosteros study · June 2026
Best of three general-purpose AI systems (best run) Posteros Multiple myeloma 28 117 4.2× Heart failure 45 118 2.6× Generalised myasthenia gravis 36 80 2.2× Alzheimer’s disease 36 65 1.8× Psoriasis / psoriatic arthritis 30 43 1.4×

Drug and model names withheld on this page; stated in the methodology available on request. "Best of three" is the highest-scoring of three leading general-purpose AI models on each drug, each scored on its best output of several runs.

Design
Structured, blind, whole-system comparison: four systems on the full task, find the evidence and turn it into defensible claims.
Comparator
Three leading general-purpose AI models (June 2026 versions), each scored on its best output of several runs.
Test system
Posteros. One of the four systems is ours, and we say so; that is why the scoring was blinded.
Sources
Identical for every system: the approved label and the pivotal-trial publication for each drug. Every system free to retrieve more.
Key metric
The scientific claim: a substantive, market-relevant, evidence-based assertion traceable to a source. Filler is not a claim, so the score cannot be inflated with volume.
Readouts
Substantiated claims per system, audited one at a time; defensible references and evidence types per document; five drug-specific requirements per drug, scored Critical failure or Issue.
Blinding
Outputs stripped of identifying markers; reviewers did not know which system produced which document.
Sample
Five drugs, five therapeutic areas, 20 full documents, 423 claims, more than 240 references.
Scoring QC
The study itself went through independent medical review and came back with six precision corrections, kept as errata.
Limitations
A first measurement, not an RCT. No completeness claim: the maximum possible claim count in a domain is unknowable, so every figure compares whole systems against each other, never against a ceiling.
Replication
Documented and designed to be re-run by anyone, including on a molecule you choose.

Our own grade for this study. We grade quality for a living, so candor is the house habit: this is not an RCT-level study. It is a first measurement.

Start from the library.
Write from the claims.

Evidence Delta tells you when the claims need to change. Simulations tell you what experts will say about them.

Talk to us The Evidence Library