amplifier-bundle-research
May 29, 2026 · View on GitHub
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.
Install
amplifier bundle add --app git+https://github.com/michaeljabbour/amplifier-bundle-research@main
amplifier bundle use research
amplifier
No YAML edits and no API keys beyond the one Amplifier already uses.
Why it exists
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, and 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 first-time researcher all get the same scaffolding and the same honest critic — shaped to what they are making.
How you use it
A six-step pipeline, available as slash commands. The research-coordinator also routes natural-language requests to the right step.
| Command | What it does |
|---|---|
/question | Sharpen a vague idea into a precise, defensible claim. |
/study-plan | Lock the methodology before results exist — writes a hash-sealed pre-registration. |
/execute | Run the analysis, gather prior art, or pull the evidence. |
/critique | An honest peer reviewer: names limitations, flags overclaiming, separates confirmatory from exploratory. |
/draft | Produce the artifact in the right structure for the audience. |
/publish | Format for the target venue (LaTeX / DOCX / brief). Gated until critique has run and any honest-pivots are acknowledged. |
Or jump straight to a document type
amplifier run "Run the patent-brief recipe on: Novel rolling-ROI control for AI agent sessions"
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"
Document recipes: patent-brief, policy-brief, white-paper, empirical-paper, benchmark-paper, replication-study, grant-proposal, literature-review, idea-generation, paperbanana-figure.
Just want to write a paper?
If you want the paper-authoring slice only — LaTeX, conference formatting, citations, and figures, without the pre-registration and experiment machinery — use the bundled lean variant:
amplifier bundle add ./bundles/paper-only.md
amplifier run --bundle research-paper-only "Draft a NeurIPS paper on ..."
(This absorbs the former amplifier-bundle-scientificpaper, which was a strict subset of this bundle and is now retired.)
What's inside
| Component | Count | Highlights |
|---|---|---|
| Specialist agents | 14 | hypothesis-designer, methodologist, preregistration-reviewer, statistician, literature-scout, idea-generator, honest-critic, ml-paper-reviewer, research-paper-architect, research-technical-writer, research-citation-manager, venue-formatter, figure-designer, research-coordinator |
| Runtime modes | 6 | the /question → /publish pipeline, registered as slash commands |
| Document recipes | 10 | one per artifact type, end-to-end |
| Discipline behaviors | 12 | honest-pivot, exploratory-labeling, stop-slop, cross-vendor-judge, cache-only-verification, plus the paper-authoring composite |
| Experiment-integrity tools | 6 | audit, power analysis, provenance check, resume/repair, stage analysis, hypothesis blocking |
| Figure generation | 1 | PaperBanana multi-agent pipeline with 8 quality veto rules (modules/tool-paperbanana) |
| Venue knowledge | 9 | NeurIPS, ICML, ACL, IEEE, ACM, arXiv + USPTO patent, policy memo, NSF/NIH grant |
Plus a reproducibility stack (environment.yml, Dockerfile.research, execution/evidence-log schemas), LaTeX templates and skeletons, and utility scripts for compilation, validation, and figure generation.
Architecture
The diagram (bundle.dot) shows the pipeline spine, the specialist agents grouped by the stage they serve, and the supporting tools and knowledge. Regenerate the PNG with dot -Tpng bundle.dot -o bundle.png.
Local development
git clone https://github.com/michaeljabbour/amplifier-bundle-research.git
cd amplifier-bundle-research
# Standalone dev composition (adds a provider so it runs without --app)
amplifier run --bundle ./bundles/dev.yaml "your prompt here"
See bundles/dev.yaml and docs/HANDOFF.md for the full workflow.
Provenance
Thin wrappers over K-Dense scientific-agent-skills, orchestrated with a Superpowers-style mode workflow, informed by Denario's multi-agent topology. Full specification in docs/SPEC.md; credits and lineage in docs/LINEAGE.md.
v0.9.0 (2026-05-29) — Autonomous experiment loop
Full empirical lifecycle inside a single session: data → analysis → honest conclusion. The loop runs propose → collect → analyze → decide → log against a hash-locked FROZEN apparatus, terminating in a held-out promotion gate confirmatory re-run + Benjamini–Hochberg + cross-vendor multi-LLM judge panel. Output is a defensible FINDING — confirmatory or honest null — not a hill-climb.
New:
skills/conducting-autonomous-experiments/SKILL.md— discipline guide for the looprecipes/autonomous-experiment-loop.yaml+ iteration body + promotion gateagents/experiment-runner.md— execution agent with integrity guardrailstemplates/experiment-ledger.yaml— schema for the hash-sealed ledgercontext/autonomous-loop-awareness.md— loop-state context for agentsscripts/experiment-loop/— helpers (hash-lock, ledger writer, BH correction)
Reached via /execute only when the locked plan declares experiment_loop; plans without one behave exactly as before.
Attribution: Pattern transfer, idea-level, from Karpathy autoresearch program.md (MIT): propose → run → measure → keep/revert → log against a frozen apparatus with git as ledger. GPU/PyTorch/CUDA training code intentionally NOT imported — only the discipline, hardened with pre-registration hash-lock, n>1 variance, guardrails, integrity audit, held-out promotion gate, and cross-vendor judge panel.
License
MIT.