pepmg_

● Evidence architecture · current September 2026

From question to cited answer.

pepmg searches several evidence lanes at once, admits only source-grounded passages, then makes a deterministic checker—not the writing model—the final authority.

LIVE ANSWER PATH

Search wide. Admit narrowly. Check mechanically.

production route
QUERY What does the evidence actually support? Intent and safety classification happen before retrieval.
RETRIEVE IN PARALLEL
01

pepmg corpus

Internal evidence pages + semantic retrieval

02

PubMed

Up to 5 biomedical records

03

Serper → Brave

Up to 10 web hits · one per domain

04

Regulators

openFDA + FDA material

05

Trials

ClinicalTrials.gov

06

Europe PMC

Study and citation coverage

  1. 01FUSE

    Rank + deduplicate

    Reciprocal-rank fusion combines the search lanes. PubMed and Europe PMC records for the same study collapse into one source.

    Authority breaks ties
  2. 02CLEAN

    Sanitize the evidence

    Prompt injection, prices, stray dose fragments, and unverified PMIDs are removed. Each admitted passage is capped at 1,500 characters.

    Maximum 6 passages
  3. 03WRITE

    Synthesize narrowly

    The model writes a short answer from numbered, sanitized passages and must cite them inline. It does not get vendor material as answer context.

    Cerebras → DeepSeek → Workers AI
  4. 04PROVE

    Run the fact gate

    Every number, year, PMID, efficacy claim, and safety claim must exist in the cited passage. Unsupported dosing or vendor recommendations are rejected.

    Failure → deterministic fallback
✓
OUTPUTShort answer + visible source cards

Dose cards, when present, are built deterministically from regulatory or primary peer-reviewed records with the population or species preserved.

● Source primacy

The closest source wins.

Search rank can find a source. It cannot upgrade one.

  1. 1Regulatory records + primary peer-reviewed studiesHighest authority
  2. 2Systematic reviews + responsible secondary reportingContext, not primacy
  3. 3Crowd sourcesDiscovery only
  4. 4Vendor materialLabeled and visible; excluded from model context

● Model map

Models write. Gates decide.

Production answer synthesis currently falls through Cerebras-hosted models, then DeepSeek, then Workers AI. Gemini 2.5 Flash exists only in a feature-flagged agent experiment; Gemini Pro is not in the live answer route.

PRIMARYQwen 3 235B → gpt-oss 120B → GLM 4.7 FALLBACKDeepSeek Chat → V4 Flash LAST RESORTWorkers AI · Llama 3.1 8B FINAL AUTHORITYDeterministic claim gate

● Deep editorial path

Field Notes use a separate, slower lane.

A publishing workflow—not the live answer path.

  1. 01Source brief≥5 relevant sources · ≥2 primary or official
  2. 02DraftClaude Opus writes against the source packet
  3. 03Critical reviewClaims, outcomes, contradictions, PMID/DOI, source primacy
  4. 04Freshness lockReview record + content hash · expires after 24 hours
  5. 05Release gateCitations, evidence, build, tests, trial merge
  6. 06Verify liveDeploy completes only after the public URL passes

● What this proves

Traceability, not certainty.

The pipeline is designed to keep each public claim attached to evidence and to fail closed when it cannot. It does not make a weak study strong, prove that a result will replicate, or turn research evidence into personal medical advice.

See the executable evidence checks → · Browse the evidence library →