Superpowers for scientific rigor — for anyone who needs a defensible written artifact.
Patent briefs. Policy positions. White papers. Literature reviews. Workshop papers. Grant applications. Full journal articles. Same discipline, same scaffolding, output shaped to the venue.
Closes the gap list from the v0.7-era external audit. Adds the orchestrated-loop recipe (the H3 enabler from the program paper), four new agents gene-transferred from amplifier-bundle-scientificpaper and AI-Scientist, two new behaviors, one new awareness context, and a healthcheck script. See docs/V0.8.0-PLAN.md for the durable record and scripts/v0.8.0-healthcheck.sh for one-command verification.
Headline additions:
recipes/orchestrated-loop.yaml— Phase 1 → Phase 2 → reconfigure → re-run, with residual-ledger schema, blinded adjudication, and no-op drift baseline (canonical reference:attractor.yamlStage 3).agents/paper-architect.md— structural-planning agent (IMRAD, abstract framework, section budgets, paper-type templates). Transferred fromamplifier-bundle-scientificpaper.agents/ml-paper-reviewer.md— venue-calibrated peer-review feedback (NeurIPS/ICML/ICLR/ACL/CVPR rubric + 5-ensemble + meta-reviewer + mandatory calibration warning).agents/literature-scout.md— agentic Semantic Scholar novelty-search loop.agents/idea-generator.md+recipes/idea-generation.yaml— autonomous research ideation with 3-axis Interestingness × Feasibility × Novelty scoring + reflection convergence + archive injection.behaviors/cache-only-verification.md+recipes/cache-only-verify.yaml— committee-audit re-verification path (--from-cachemode) without API keys.behaviors/cross-vendor-judge.md— codifies cross-vendor LLM-judge enforcement to mitigate the v3 program paper's reflexivity-hazard #1.context/experiment-calibration-awareness.md— ECE / Brier / reliability-diagram as first-class metrics.scripts/v0.8.0-healthcheck.sh— 43-check structural + capability + wire-up healthcheck. All passing.
All gene-transfers are idea-level. AI-Scientist's GPU experiment runner, aider dependency, and NPEET are intentionally NOT imported.
amplifier bundle add --app git+https://github.com/michaeljabbour/amplifier-bundle-research@main
amplifier bundle use research
amplifierThat's it. No YAML edits, no API keys beyond the one Amplifier already has.
> /question
What claim are you trying to defend?
> /study-plan
Locks your methodology before you see results. Writes a hash-sealed
pre-registration. (Named /study-plan because the generic /plan slash
command is reserved by amplifier-bundle-modes — the research-coordinator
routes natural-language "plan" requests here regardless.)
> /execute
Runs the analysis, gathers prior art, or pulls the evidence — depending on
what you're making.
> /critique
An honest peer reviewer on your work. Names specific limitations. Flags
overclaiming. Separates confirmatory from exploratory.
> /draft
Produces the artifact in the right structure for the audience.
> /publish
Formats for the target venue. LaTeX, DOCX, or plain brief. Blocks until
critique has run and any honest-pivots have been acknowledged.
Natural-language invocation (research-coordinator routes to the right recipe):
amplifier run "Run the patent-brief recipe on: Novel rolling-ROI control for AI agent sessions"
amplifier run "Run the policy-brief recipe on: Municipal guidance on generative AI in K-12"
amplifier run "Run the white-paper recipe on: Return on Inference — when model spend pays back"
amplifier run "Run the empirical-paper recipe on: Reflection tokens improve long-horizon reasoning"
amplifier run "Run the replication-study recipe on: Reproducing X et al. 2025 on dataset Y"
amplifier run "Run the benchmark-paper recipe on: RCE — Reasoning Chain Evaluation benchmark"
amplifier run "Run the grant-proposal recipe on: NSF CAREER — Inference-time compute allocation"
amplifier run "Run the literature-review recipe on: Systematic review of chain-of-thought methods"Or direct tool invocation:
amplifier tool invoke recipes operation=execute \
recipe_path=@research:recipes/patent-brief.yaml \
'context={"invention": "Novel rolling-ROI control for AI agent sessions"}'Scientific rigor is a set of habits: ask a precise question, lock the method before you see the data, report honestly against yourself, cite properly, format for the reader. Working researchers learn these over a decade. Everyone else reinvents the mistakes.
This bundle encodes the habits. A patent attorney, a policy analyst, a founder drafting a technical white paper, or a junior researcher drafting their first workshop paper — all get the same scaffolding and the same honest critic, shaped to what they're making.
Thin wrappers over K-Dense scientific-agent-skills, orchestrated with a Superpowers-style mode workflow, informed by Denario's multi-agent topology, packaged for the install ease of amplifier-bundle-stories.
See docs/SPEC.md for the full specification.
- 10 agents — hypothesis-designer, preregistration-reviewer, methodologist, statistician, honest-critic, technical-writer, figure-designer, citation-manager, venue-formatter, research-coordinator
- 6 runtime modes —
/question,/study-plan,/execute,/critique,/draft,/publish(register as slash commands viaamplifier-bundle-modes) - 9 recipes — patent-brief, policy-brief, white-paper, empirical-paper, benchmark-paper, replication-study, grant-proposal, literature-review, paperbanana-figure
- 7 behaviors — honest-pivot, exploratory-labeling, paperbanana, figure-generation, latex-authoring, conference-styling, research-modes (composite in
behaviors/research.md) - 6 conference formats + 3 non-academic venues — NeurIPS, ICML, ACL, IEEE, ACM, arXiv + USPTO patent, policy memo, NSF/NIH grant
- Reproducibility stack —
environment.yml,requirements.in,Dockerfile.research,execution-log.yamlschema,evidence-log.yamlschema for evidence-gather recipes - LaTeX templates — ACL and IEEE style files, IMRAD skeleton, patent-brief + policy-brief + white-paper + grant + replication-study markdown skeletons
- PaperBanana — multi-agent figure generation with 8 quality veto rules (Python module at
modules/tool-paperbanana/) - Utility scripts — LaTeX compilation, format validation, figure generation, template download
Generated via amplifier tool invoke recipes operation=execute recipe_path=@foundation:recipes/generate-bundle-docs.yaml. Regenerate on structural changes (validate-bundle-repo does this automatically when source_hash drifts).
End-to-end runs of the pipeline, preserved as repo artifacts with session YAMLs, final PDFs, and capability catalogs. Browse the index at docs/SAMPLES.md. First entry: examples/whitepaper-dogfood-run/ — a 16-page LaTeX white paper on the bundle itself, written with the bundle, for a patent-attorney audience.
To exercise the bundle from a local checkout without publishing, clone and register the working tree:
git clone https://github.com/michaeljabbour/amplifier-bundle-research.git
cd amplifier-bundle-research
# Register the local working tree as a named bundle
amplifier bundle add "file://$(pwd)" --name research-dev
# Run against it
amplifier run --bundle research-dev --mode chat "your prompt here"
# After local changes, refresh:
amplifier bundle remove research-dev && \
amplifier bundle add "file://$(pwd)" --name research-devSee bundles/dev.yaml for the standalone composition and docs/HANDOFF.md for the full development workflow.
v0.2.0 — fully specified, implemented, and Amplifier-hygienic. All agents, modes, recipes, behaviors, reproducibility templates, and three non-academic venue formats are written. bundle.md is in valid frontmatter form; agents carry proper meta: blocks with model_role; tool-paperbanana is protocol-compliant; modes register as slash commands; critique-required and honest-pivot-acknowledged gates enforce at /publish.
See docs/HANDOFF.md for what's verified, what's not yet been tested end-to-end, and the v0.4+ roadmap (user testing across three personas, dogfood paper, domain packs).
MIT. Credits and lineage in docs/LINEAGE.md.
This bundle is the broader research-lifecycle substrate (/question → /study-plan → /execute → /critique → /draft → /publish, with hypothesis sharpening, methodology lockdown, statistical tooling, and reproducibility infrastructure). It is the right choice for any researcher writing a defensible artifact end-to-end.
If you only want the paper-authoring slice — no modes, no pre-registration, no statistical machinery, just LaTeX/conferences/figures/citations — install amplifier-bundle-scientificpaper instead. It has a coherent independent purpose: the clean no-modes-required interface for "I just want to write a paper."
The two bundles share substantial infrastructure (Python scripts, conference formats, LaTeX templates, PaperBanana figure-generation module, paper-authoring behaviors) but differ in scope: research includes the full research-lifecycle pipeline (idea generation, novelty checking, peer-review feedback, statistical validation, the orchestrated H3 loop); scientificpaper covers only paper authoring. Detailed inventory + gene-transfer history in docs/V0.8.0-PLAN.md.