Prototype | Paper-Pro workflow

The strongest end-to-end research paper workflow in this tool right now.

paper-pro is the production-style paper lane for Bohr: one lead agent, a specialized manuscript architecture pass, adversarial review, citation-integrity checking, evidence verification, compliance checks, and a final provenance sidecar. This page shows the exact sample paper, the agent graph behind it, and the cost discipline that keeps the workflow usable before it becomes expensive.

Read the full sample paper Open the technical workflow docs

Validated run

20 Apr 2026

Strict smoke workflow pass for the sample paper

Input tokens

133,875

Measured across manuscript subagents in the passing proof run

Output tokens

37,594

Compressed handoffs plus one final manuscript expansion

Subagent cost

$0.99469

Measured subagent-only telemetry; lead orchestration cost is slightly higher

Agent flow

How the paper moves from topic to final manuscript

The design principle is simple: only one stage should spend heavily on long-form prose at a time. Everything else stays short, explicit, and artifact-driven so the workflow remains debuggable and cheaper to operate.

Diagram 01

Paper-pro orchestration map

Diagram 02

Token bifurcation map

Bohr lead agent

Orchestrates the entire paper lane.

Scopes the run, selects the slug, coordinates artifact paths, hands the right context to each specialist, reconciles critiques, and promotes the final manuscript plus provenance sidecar.

paper-architect

Designs the paper before prose starts.

Turns a raw topic and evidence pack into a manuscript architecture: thesis boundary, section plan, evidence map, argument order, and the constraints that keep the writer honest.

writer

Expands the architecture into a complete draft.

Writes the actual paper from the architecture memo and evidence files, keeping claims bounded to what the local evidence package can really support.

citation-integrity

Audits citation-to-claim fidelity.

Checks whether each claim is supported at the right granularity, whether source-status labels are honest, and whether weaker sources are clearly marked instead of silently promoted.

reviewer

Acts like a skeptical peer reviewer.

Looks for overclaiming, missing caveats, weak transitions, false confidence, and places where the prose sounds stronger than the evidence base actually is.

verifier

Runs the evidence-truth pass.

Checks the revised manuscript against the local evidence bundle and records the honest status: pass, inferred, blocked, or still unverified.

paper-compliance

Confirms the paper contract is complete.

Checks that the required sections exist, the run produced the expected artifacts, and the manuscript is promotable into a final paper plus provenance file.

Prompt used

The exact sample invocation

This sample page is not based on a vague demo pitch. It comes from a real workflow test using the prompt below. 

/paper-pro notes/citation-verification-research-notes.md. Use slug citation-verification-proof. Primary local evidence files are notes/citation-verification-research-notes.md, notes/citation-verification-research-verified-excerpts.md, and notes/citation-verification-research-verification.md. Do not reuse draft, review, citation, compliance, or provenance artifacts from any other slug. Create fresh architecture, draft, citation-integrity, review, and compliance artifacts for this slug before finalizing.

Artifact contract

What the workflow is required to produce

Final paper

papers/citation-verification-proof.md

The complete sample manuscript rendered below on this page.

Provenance sidecar

papers/citation-verification-proof.provenance.md

Honest source accounting, verification scope, and unresolved weaknesses.

Session transcript

notes/citation-verification-proof-paper-pro-session.txt

Raw workflow run transcript from the successful sample generation.

Architecture memo

outputs/.plans/citation-verification-proof-architecture.md

The paper blueprint used to control scope before drafting.

Cost optimization

How token bifurcation and caveman mode keep the paper lane practical

The workflow does not treat every agent as a full essay writer. It splits work into cheap reasoning lanes and expensive prose lanes, then uses file-based handoffs to avoid re-paying for the same context over and over.

Cheap planning lane

Architecture, critique, and compliance passes stay short. They exchange memos and path references instead of pasting full papers back into parent context.

Expensive prose lane

Only the writer and final lead synthesis spend big on long-form prose. That keeps most agents in compressed reasoning mode and reserves expansion for the manuscript itself.

File-based handoffs

Artifacts land on disk, then later agents read the exact files they need. This avoids repeated transcript inflation and stabilizes resumability for long research runs.

Caveman mode

Short `[CAVEMAN MODE: ...]` prefixes and minimal context objects reduce chatty output. In this repo, the caveman skill cites roughly 75% output-token reduction as a design goal, not as a measured guarantee for every run.

Measured on this sample run

37,594 output tokens

The passing paper-pro proof run used 133,875 input tokens, 37,594 output tokens, and recorded $0.99469 in subagent spend.

Illustrative caveman design model

~112,782 output tokens avoided

The caveman skill documents roughly 75% output-token reduction as a design goal. If that target held for this run, a verbose equivalent would be about 150,376 output tokens. That is a modeled comparison, not an instrumented A/B measurement.

Modeled verbose output: ~150,376 tokens

Approximate spend implication

~$0.65 avoided

Using the observed run as a blended cost baseline, the same workflow without caveman-style compression would trend toward roughly $1.65 in subagent spend instead of $0.99. This is a run-level model, not provider-billed A/B telemetry.

Why it saves: short planning memos, file-based handoffs, and only one expensive long-form writer lane

Quality gates

Why this workflow is strong professionally

  • Fresh architecture, draft, citation-integrity, review, verifier, compliance, final paper, and provenance artifacts were all created for the `citation-verification-proof` slug.
  • The verifier memo reported no fatal or major unsupported claims in the revised manuscript.
  • Citation-integrity issues were revised before final promotion, instead of being hidden in post-hoc cleanup.
  • The sample remains intentionally narrow: biomedical evidence only, with weaker de Lacey support kept as historical context rather than central proof.

Source-status honesty

What the provenance record says explicitly

  • Mogull 2017 and Greenberg 2009 were treated as directly re-fetched PMCID-backed support in the local evidence package.
  • Tatsioni et al. 2007 was treated as directly re-fetched PubMed abstract support.
  • ICMJE was used only for broad normative context, not as proof of a line-by-line citation-checking mandate.
  • de Lacey 1985 remained explicitly weaker-support historical context in the final paper and provenance record.

Session trace

Proof-run transcript excerpt

This excerpt comes from the successful session transcript saved to notes/citation-verification-proof-paper-pro-session.txt. 

$ node ./scripts/test-paper-pro-workflow.mjs
Paper-pro workflow validation passed.
Generated local artifacts:
  papers/citation-verification-proof.md
  papers/citation-verification-proof.provenance.md
  outputs/.plans/citation-verification-proof-architecture.md

Provenance sidecar

The workflow ships paper plus evidence accounting

Provenance Sidecar

The local proof run writes source accounting and verification status to papers/citation-verification-proof.provenance.md.

The deployed website does not require this ignored generated artifact to exist at build time. When the artifact is present in a local workspace, this page renders its full contents automatically.

Sample output

Full paper generated by the workflow

The manuscript below is rendered from the actual sample artifact at papers/citation-verification-proof.md.

Citation Verification Proof

The full sample manuscript is generated by the local paper-pro proof run and stored at papers/citation-verification-proof.md.

This hosted prototype keeps the page build-safe because generated research artifacts are intentionally ignored by Git in this repository. The workflow contract, telemetry, agent flow, and provenance expectations remain visible here, while the local artifact path records where the complete manuscript is produced during a proof run.