The Epistemic Stack
An adversarial pipeline for compounding, verifiable knowledge synthesis.
Last updated: September 2026
Problem
A language model can summarize a contested debate fluently without revealing whether the underlying disagreement is factual, methodological, or simply two sources answering different questions. Most AI-assisted research tools flatten this distinction. They produce a confident narrative that erases the structure of the disagreement, and that structure cannot be checked independently by anyone reading the summary.
Two failures recur in practice. The first is false contradiction: two credible sources answering different questions, a population-level correlation and a controlled mechanistic study, get flattened into "the evidence is mixed," when both can be true at once. The second is unverifiable confidence: a system asserts that a claim is well supported without exposing which sources it depends on, whether those sources exist, or whether the claim actually says what it is cited for.
Approach
The Epistemic Stack treats a contested topic as a typed dependency graph rather than a paragraph. Raw sources are decomposed into atomic claim elements, each tied to a verifiable source identifier, a DOI or an arXiv ID. Claims are connected by four relation types: SUPPORT, where one claim increases confidence in another; CONTRADICT, where two claims cannot both be true; DEPEND, where one claim's validity presupposes another; and CONTEXT_MUTATION, where two claims are each valid within a different interpretive frame, population versus mechanism, definition versus measurement, and should not be flattened into a contradiction.
CONTEXT_MUTATION marks two claims as non-contradictory because they answer structurally different questions, even when they share overlapping vocabulary. It is the relation type that stops a system from manufacturing a false binary out of a methodological frame mismatch, the pattern behind a large share of "the science is divided" headlines in nutrition and social-science reporting. It was added to the specification in response to early evaluator feedback on an earlier version of this work.
Once the graph exists, a deterministic Crux Score identifies which claims carry the most structural weight in a dispute, weighted by the credibility of the sources contradicting them, not simply by how many sources disagree. A companion signal, the Ag-Vacuum routine, points to the single missing piece of evidence that would most change the graph's shape, expressed as a concrete target rather than a generic call for more research. An automated Source Fidelity Check verifies, before a claim enters the graph, that its cited source exists and that the claim accurately represents what the source says.
Artifact
- A three-layer pipeline, Ingestion, Structure, Assessment, implemented as a runnable Python repository with zero external dependencies beyond the standard library.
- 26 source-verified claim nodes across three independent case studies, an open scientific dispute, a resolved frame mismatch, and a closed consensus with a documented objection, processed by the same unmodified pipeline.
- An automated Source Fidelity Check with a nine-state verification taxonomy, checking DOI and arXiv identifiers against Crossref and arXiv before a claim is trusted.
- A compounding protocol for merging graphs proposed by separate analysts, with deduplication, cycle detection, and a timestamped audit trail.
- A 12-test automated suite and a one-command reproducibility check that confirms the written specification matches the executable code exactly.
Evidence
- Applied without modification to three structurally different disputes, the pipeline produced three distinct, correct diagnoses: an open contested crux for the COVID-19 origins case, a resolved frame mismatch for the dietary eggs case, and a closed consensus with a documented rebuttal for the LHC safety case. Same code, three different shapes of disagreement, correctly told apart.
- Running the code, not just describing it, surfaced two real defects during adversarial testing: an early version of the Crux Score counted contradicting sources by number rather than weighting them by credibility, so a single unreviewed preprint counted as much as a peer-reviewed study; and the dependency graph, documented throughout as acyclic, had no runtime check enforcing that property until testing showed two nodes could be proposed as each other's sole dependency, artificially inflating both scores. Both were found, fixed, and are now covered by regression tests.
- A systematic citation cross-check re-verified all 26 nodes against their primary sources rather than against the notes taken when each node was built. Result: 20 aligned, 5 partially off and corrected in the current version, 1 unverifiable and flagged as such, 0 misattributed.
- Submitted to the Future of Life Foundation's Epistemic Case Study Competition in July 2026.
Two things this project does not yet do, stated plainly rather than smoothed over. Assigning CONTEXT_MUTATION instead of CONTRADICT is currently a judgment made by the human analyst during ingestion, not something the pipeline decides on its own; automating that classification is planned future work. And the compounding protocol described above has been tested on graphs built independently within the same team. A cross-team merge test, two analysts with no prior coordination, is the next validation step before this becomes a standalone artifact in its own right.