PaperCI

Continuous integration for evidence-backed scientific stories and hypotheses.

PaperCI helps researchers answer two linked questions:

Given the results we actually have, what is the strongest scientific story we can responsibly tell?

Which ambitious, falsifiable mechanisms are worth testing next without pretending they are already established?

It converts experimental results into traceable evidence cards, proposes competing paper arcs, tests every claim against its evidence, and makes the most important remaining uncertainties explicit for human review. A separate frontier track turns those boundaries into speculative Hypothesis Cards with competing explanations, predictions, falsifiers, decisive tests, and evidence-upgrade paths.

PaperCI is designed to improve when foundation models improve. Models may propose better interpretations and stories; the project owns the stable evidence format, claim rules, provenance, evaluation cases, and human review workflow.

Try it in two minutes

PaperCI is pre-alpha. The source tree is preparing v0.5.0a1; install the latest verified GitHub artifact (v0.4.0a1) to try the current released behavior, then create a complete synthetic project without an API key or network call. PyPI publication is currently blocked by the repository's protected deployment rules; the protection gate has not been bypassed.

$ python -m venv .venv
$ source .venv/bin/activate
$ python -m pip install \
    https://github.com/XiaofengZhou16/PaperCI/releases/download/v0.4.0a1/paperci-0.4.0a1-py3-none-any.whl
$ paperci demo
$ cd paperci-demo
$ paperci lint --fail-on never
$ paperci compare
$ paperci compare-hypotheses

The demo creates two synthetic result files, traceable evidence and claims, three competing stories, three frontier hypotheses, generation manifests, and paperci-report.md. It intentionally contains an overstated mechanism claim. paperci lint flags PCI-MECH-001; the frontier track retains ambitious mechanisms as explicitly speculative and falsifiable research directions.

See the ten-minute pilot to evaluate PaperCI with a safe synthetic case and report first-run friction.

Build a project manually

From a source checkout installed with python -m pip install -e .:

$ paperci init demo --id demo-study --title "Demo study"
$ paperci import-table demo results.csv --statement-column finding --unit-column unit
$ paperci add demo --statement "Target expression was higher after exposure." \
    --source notes://pilot --locator result-1
$ paperci claim demo --text "Exposure is associated with higher target expression." \
    --type association --strength supports --support E001
$ paperci claim demo --text "Exposure directly activates the target mechanism." \
    --type mechanism --strength demonstrates --support E001 --depends-on C001
$ paperci propose demo --arcs 3
$ paperci lint demo --fail-on never
$ paperci compare demo
$ paperci hypothesize demo --count 3
$ paperci compare-hypotheses demo
$ paperci explain PCI-MECH-001
$ paperci report demo -o demo/paperci-report.md

For a non-interactive smoke test from this repository:

paperci doctor examples/minimal-project.yaml
paperci validate examples/minimal-project.yaml
paperci propose examples/minimal-project.yaml --arcs 3 --dry-run
paperci lint examples/minimal-project.yaml --fail-on never
paperci compare examples/minimal-project.yaml
paperci hypothesize examples/minimal-project.yaml --count 3 --dry-run
paperci report examples/minimal-project.yaml -o paperci-report.md

The repository example also intentionally contains the same overstated mechanism claim, so paperci lint should flag PCI-MECH-001.

The report contains:

• distinct candidate arcs when the claim register supports different commitments;
• the evidence path behind every major claim;
• unsupported, overstated, or contradictory claims;
• a figure-question sequence linked to existing evidence and claims;
• alternative explanations that remain open;
• explicit gaps, run provenance, and human-review boundaries.

ProjectSpec 0.4 also records claim prerequisites with depends_on. The built-in provider follows this acyclic graph to recover long story chains instead of selecting only isolated claims. Nested outcomes can record parent_unit and per-group clusters, so tumour-, organoid-, well-, or cell-level counts do not silently stand in for independent animals or donors.

Three adoption modes

The user can start in sketch and progressively add provenance. A perfect data model is not a prerequisite for receiving value.

What PaperCI is not

• not a journal acceptance predictor;
• not a replacement for domain experts, statisticians, or reviewers;
• not an autonomous manuscript factory;
• not permission to promote association, motif enrichment, or prediction into causality;
• not a system that silently uploads unpublished results.

Design principles

1. Evidence before eloquence. A fluent claim with weak support still fails.
2. Competing stories, not one confident answer. The system exposes trade-offs.
3. Hard gates before taste. Validity is pass/fail; editorial appeal is comparative.
4. Local first. Core inspection and linting work without a network or model API.
5. Model agnostic. Model adapters are replaceable and all runs are recorded.
6. Git friendly. Canonical YAML/JSON is readable, diffable, and reviewable.
7. Human promotion. Generated evidence and claims never self-verify.
8. Dual-track reasoning. Current claims and frontier hypotheses remain separate records.
9. No journal score. Ambition is multidimensional; novelty and publication fit require review.

Repository design package

• Product contract
• Technical architecture
• Open specification
• Core lint rules
• Story evaluation
• Table import
• Mechanistic-biology profile
• Ecosystem and differentiation
• MVP and roadmap
• Release process
• Community and governance
• Combined JSON Schema
• Minimal example

The repository contains the local evidence-bound story workflow plus an offline frontier-hypothesis baseline. The built-in hypothesis provider creates transparent research scaffolds; it does not search literature, nominate a named molecular actor that is absent from the input, or predict journal acceptance. Remote and local foundation-model adapters remain future work.

Commands

paperci lint --format sarif can feed findings into GitHub or another SARIF-aware code-review system. CI may use --fail-on error; exploratory work may use --fail-on never while retaining every finding.

Project status

Pre-alpha / v0.5.0a1 candidate. The offline CLI separates evidence-bound claims from frontier hypotheses, validates explicit reasoning dependencies, and includes the first benchmark-derived semantic and nested-unit rules. Hypothesis novelty remains unchecked unless a dated, traceable literature assessment is recorded. The perfect score on the single constructed Nature fixture is a regression result, not a general performance estimate. No output should be used in a submission or experimental decision without independent scientific review.

License

Code, schemas, tests, and repository documentation are licensed under Apache-2.0. Imported benchmark artifacts retain their source-specific licenses and must include a case manifest before public redistribution. See LICENSE.

Contributing

Start with CONTRIBUTING.md. Scientific rules require failing and passing fixtures, an actionable remediation, known exceptions, and review ownership. Security or privacy concerns should follow SECURITY.md, not a public issue containing unpublished data.

Repository: github.com/XiaofengZhou16/PaperCI

Citation

If PaperCI contributes to published work, cite the exact release or commit used so the scientific rules and ProjectSpec behavior remain reproducible. GitHub can export the repository citation from CITATION.cff; release-specific citation metadata will be attached to public releases when archival deposition is available.