Skip to main content
RetrievalResearchCitations·2026-04-12·8 min read·case_id · uc_research_synthesis_v1

Two hundred papers, one coherent codex.

A computational biology lab dropped a literature corpus of 200 peer-reviewed papers into Gnostikon. The platform extracted confidence-banded claims, tied each one to its source paragraph, and surfaced disagreements between authors as Aletheia conflicts — instead of silently choosing whichever line came first.

200
Source papers ingested
↑ corpus expanding monthly
1,847
Gnoses extracted
↑ across 11 sub-domains
12%
Aletheia conflicts surfaced
author disagreements
96%
Citations resolved to source paragraph
↑ vs. 67% baseline
Pipeline · corpus → coherent codex

A literature corpus arrives messy. The codex it becomes is auditable — every claim chains back to a paragraph.

stage 1
Ingest PDFs
parse · OCR · structure
stage 2
Extract claims
segment · cite-link
stage 3 · active
Validate vs. ethos
scientific_rigor · v3
stage 4
Surface conflicts
Aletheia · 224 disagreements
stage 5
Index in canonical
searchable · citable

The problem

The lab had a folder of two hundred peer-reviewed papers and an internal review process that produced a fifty-page summary nobody trusted. The summary's claims were strong; the citations behind them resolved to entire papers, not paragraphs. When a new lab member challenged a claim, the path back to its evidence was a manual hunt through PDFs.

The lab's bottleneck was not literature search — it was citation chain integrity. Reviewers spent more time verifying that a cited paragraph actually said what the summary attributed to it than they spent on the science.

The approach — the corpus is the codex

The lab dropped the corpus into Gnostikon and let the platform build a codex from it. The pipeline is mechanical: ingest each paper, segment its content into candidate gnoses, validate each one against the pinned scientific_rigor ethos, surface disagreements between authors as Aletheia conflicts, and index the validated set.

Three commitments shape the result:

  • Every claim chains back to a paragraph. Provenance is paragraph-grained, not paper-grained. Reviewers click a claim and land at the source line.
  • Confidence is banded, not boolean. The platform reports confidence ranges per claim; reviewers can sort the codex by uncertainty as easily as by topic.
  • Author disagreements are first-class objects. When two papers contradict each other, the platform does not pick a winner; it tags both gnoses with a conflict reference and lets the lab adjudicate.

Our reading group used to argue about what the literature said. Now we argue about what to do with it. The platform got us out of the citation-chasing business in about a week.

Helena V.Principal investigator, lit_chronobiology

Shape of an extracted gnosis

Each gnosis carries a claim, a paragraph-level citation back to the source paper, the author and publication metadata, the confidence band the platform assigned, and an aletheia.conflicts array listing other gnoses in the codex that contradict it. A gnosis is, in effect, a peer-reviewed claim with a structured confidence record and a graph of disagreements.

The codex as a whole is queryable by topic, by author, by confidence band, or by conflict density. A high-conflict topic is a research opportunity; a low-conflict one is settled science.

Operations & telemetry

The corpus pass produces one umbrella operation that records every paper ingested, every gnosis extracted, every conflict surfaced, and the codex's final state. Subsequent ingests append; the codex is never rebuilt from scratch.

op_idop_2026_04_12_T1a3b
corpus_idcorp_chrono_2026_q2
papers_ingested200
gnoses_extracted1,847
cite_resolution_rate96%
aletheia_conflicts224
ethosscientific_rigor · v3
scopelit_chronobiology
revision_pairbackend 2c4a · client 8819
end_statecodex indexed · 1,847 gnoses live

Outcomes after one corpus pass

  • 200
    Papers ingested in a single corpus pass.

    The lab walked out of the ingest cycle with a queryable knowledge base instead of a folder of PDFs. Search now returns claims, not documents.

  • 1,847
    Gnoses extracted, every one banded by confidence.

    Each claim carries the source paragraph, the cite chain, and an Aletheia confidence score. Reviewers can sort by uncertainty as easily as by topic.

  • 12%
    Of extracted claims contradict at least one peer.

    These conflicts now sit at the centre of the lab's reading-group agenda. The platform did not resolve the disagreements — it made them legible.

  • 96%
    Cite resolution to source paragraph (not document).

    Manual citation chains landed at 67% in the prior workflow. Bumping to 96% removed an entire class of "I think this came from…" review notes.

Implementation playbook

Four decisions worth making early when you build a research codex on Gnostikon.

1 · Pin a strict ethos before you ingest.

scientific_rigor should forbid: [speculation, undated_sources] and require: [provenance, confidence]. The ingest pipeline will reject gnoses that don't meet the contract — which is exactly what you want from a literature corpus.

2 · Resolve citations to paragraphs, not papers.

A claim cited to a paper is not citable; a claim cited to a paragraph is. Run citation resolution at ingest, even if it means an extra parse pass. The reviewer time you save is many multiples of the compute cost.

3 · Treat conflicts as features, not bugs.

A literature corpus with no Aletheia conflicts is a corpus the platform did not actually read. Surface conflicts in the codex's UI — they are the most useful signal a literature pass produces.

4 · Re-validate the codex on every new corpus pass.

When a new paper enters the codex, prior gnoses may need their confidence bands updated or their conflicts re-evaluated. Run validation across the affected scope, not just the new gnoses.