Specification

April 23, 2026 · View on GitHub

Status: v0.4.0 — v0.3 verification PASS (scripts, modules, smoke tests) + v0.4 simulated persona acceptance PASS (A patent attorney, B ML researcher, C peer reviewer). Real-user testing scheduled post-v0.4 per roadmap. Author: Michael Jabbour Target ecosystem: Microsoft Amplifier

1. Purpose

Give anyone who needs to produce a defensible written artifact the same disciplined workflow a trained researcher uses — without requiring the training.

The bundle does not write papers autonomously. It keeps the human in the driver's seat and prevents sloppy reasoning at each step: fuzzy questions, motivated-reasoning methodology, cherry-picked results, overclaiming, sloppy citations, wrong venue format.

2. Scope

In scope

Any artifact where a reader asks "is this claim defensible?" and the author needs to answer yes. Concretely:

Patent briefs and invention disclosures
Evidence-based policy briefs
Technical white papers
Due-diligence and market-analysis reports
Literature reviews
Conference and workshop papers
Journal articles
Grant applications
Replication studies

Out of scope

Creative writing (use amplifier-bundle-stories)
Code-first engineering work (use amplifier-bundle-superpowers)
Real-time experimental operation (the bundle drafts the plan and analyzes the results; a human or a separate pipeline runs the experiment)
Fully autonomous paper generation (see Denario; explicit non-goal here)

3. Design principles

Human-led, rigor-enforced. Denario aims for AI as primary author. This bundle aims for humans as primary author, with AI as the methodology coach and honest critic. The human makes every substantive call; the bundle refuses to let them skip a step.

Rigor is discoverable, not required. Modes default to the rigorous path. A user who wants to skip pre-registration can. The bundle will label the resulting output as exploratory rather than confirmatory and carry that label through to the draft.

One install, zero config. Matches amplifier-bundle-stories: amplifier bundle add then amplifier bundle use. No API keys beyond Amplifier's existing provider config. Default templates cover the 80% case.

Output ladder, not monoculture. Same scaffolding — question, plan, execute, critique, draft, publish — produces a 2-page patent brief or a 12-page Nature submission. The recipe picks the output shape; the discipline stays constant.

Skills stay lazy. The 10 K-Dense skills are registered but unloaded. Modes pull them on demand. No context tax for browsing or for projects that only need a subset.

Thin wrappers, upstream owns the content. Agents are ~30 lines each, referencing K-Dense skill files by path. The skill library stays a single source of truth. Following the ramparte/amplifier-collection-superpowers pattern.

4. Personas

Three users with different starting points and the same destination — a defensible written artifact.

Persona A: The non-scientist

Patent attorney, policy analyst, product manager, technical founder, journalist, due-diligence analyst. Has a claim to defend. Has never written IMRAD. Doesn't know what a p-value does or doesn't mean. Needs the bundle to translate ordinary intent into rigorous scaffolding without condescending.

Persona B: The working researcher

ML researcher writing a workshop paper. PI drafting a grant. Engineer writing a technical report. Knows IMRAD, has strong intuitions, wants Superpowers-style discipline enforcement so their own shortcuts don't leak through. Treats the bundle as a rigor co-pilot.

Persona C: The research-adjacent reviewer

Due-diligence partner, policy adviser, standards editor, patent examiner. Doesn't generate claims — evaluates them. Uses the /critique mode standalone against incoming artifacts and the scholar-evaluation agent as a structured scorer.

Full persona detail in PERSONAS.md.

5. The output ladder

Output	Target length	Recipe	Primary mode sequence
Patent brief	2–5 pages	`patent-brief`	question → plan → execute (prior-art) → critique → draft → publish
Policy brief	2–8 pages	`policy-brief`	question → plan → execute (evidence gather) → critique → draft → publish
White paper	6–20 pages	`white-paper`	full sequence, heavier draft phase
Literature review	8–25 pages	`lit-review`	question → execute (systematic search) → critique → draft → publish
Workshop / conference paper	4–10 pages	`empirical-paper`	full sequence with pre-registered experiments
Journal article	10–30+ pages	`empirical-paper` with `--long`	full sequence with multi-round critique
Grant application	venue-specific	`grant`	question → plan → critique → draft → publish
Replication study	6–15 pages	`replication-study`	starts at plan (question comes from source paper)

All recipes compose from the same six modes and the same eight agents. Only the templates, venue configs, and critique emphasis differ.

6. Modes

Six modes, fixed names, stable contract. Each mode has clear inputs, outputs, and agents invoked. The CLI vocabulary is plain English so non-scientists get it immediately.

Mode	Inputs	Outputs	Primary agents
`/question`	A rough claim or topic in the user's own words	A sharpened, falsifiable, operationalized question with success criteria	`hypothesis-designer`, `methodologist`
`/plan`	The sharpened question	A locked pre-registration artifact (YAML + human-readable summary) with methodology, predictions, analysis plan, and the honest-pivot clause	`methodologist`, `statistician`, `preregistration-reviewer`
`/execute`	The pre-registration and any raw materials (code, data, corpus, prior art seeds)	Results, figures, prior-art table, or evidence list — depending on recipe	`statistician`, `figure-designer`; recipe-specific tools
`/critique`	The results or the draft	A structured critique: confirmatory vs exploratory separation, named limitations, overclaim flags, scholar-eval scores	`honest-critic`
`/draft`	Pre-registration + results + critique	The artifact in the right structure: IMRAD, patent-brief skeleton, policy-brief skeleton, etc.	`technical-writer`, `figure-designer`, `venue-formatter`
`/publish`	The draft + target venue	Venue-ready output: LaTeX with correct preamble, DOCX with citations, or plain markdown brief	`venue-formatter`

Each mode has a full definition in modes/*.md. The mode files are agent prompts, not code.

7. Agents

Eight thin wrappers over K-Dense skills. Each agent is a ~30-line AGENT.md that references the skill it wraps and sets the agent's role, tone, and output contract.

Agent	Wraps	Role
`hypothesis-designer`	`hypothesis-generation`	Turn rough claims into falsifiable, operationalized questions with predictions and mechanisms
`methodologist`	`scientific-critical-thinking`	Design methodology, flag bias risks, reference GRADE / Cochrane ROB / CONSORT / STROBE / PRISMA where relevant
`preregistration-reviewer`	`scientific-critical-thinking`	Review the pre-registration before data is seen. Confirm predictions, stat tests, MDEs, and analysis plan are specific enough to be falsifiable
`statistician`	`statistical-analysis`	Test selection, assumption checking, power analysis, APA-formatted reporting
`honest-critic`	`peer-review` + `scholar-evaluation`	Structured critique with CONSORT/STROBE/PRISMA checklists + quantitative scoring. Defaults to skeptical
`technical-writer`	`scientific-writing`	IMRAD and non-IMRAD structures. Two-stage outline → prose
`figure-designer`	`scientific-schematics`	Publication-quality diagrams and charts
`venue-formatter`	`venue-templates` + `citation-management`	LaTeX/DOCX venue output with proper bibliography

Full agent manifests in agents/*.md. See agents/hypothesis-designer.md for the canonical thin-wrapper pattern.

8. Skill layer

The bundle does not reimplement scientific skills. It wraps K-Dense scientific-agent-skills as the single source of truth. The 10 skills used:

hypothesis-generation       scientific-critical-thinking
scientific-brainstorming    statistical-analysis
peer-review                 scholar-evaluation
scientific-writing          scientific-schematics
citation-management         venue-templates

Skills are registered at bundle install and loaded by agents on demand via load_skill(). The remaining 124 K-Dense skills are available but not registered — researchers in a specific domain (genomics, chemistry, astrophysics) can add them explicitly.

9. Behaviors

Two behaviors run silently across every mode:

honest-pivot — If the analysis diverges from the pre-registered predictions, the bundle surfaces this explicitly and marks the deviation in the draft. No quiet rewriting of the hypothesis after results are in.

exploratory-labeling — Any finding that wasn't in the pre-registration gets tagged exploratory and carries that label through to the draft's Results section and the publish-phase formatting.

Both are on by default. Both can be disabled explicitly (with a visible warning) for users who know what they're doing.

10. Recipes

Recipes are YAML files that compose modes, set defaults, and bind the output template. A minimal recipe:

name: patent-brief
description: USPTO-style invention disclosure with prior-art section
modes: [question, plan, execute, critique, draft, publish]
defaults:
  template: templates/patent-brief-skeleton.md
  venue: uspto
  critique_emphasis: [novelty, enablement, prior-art-coverage]
  honest_pivot: true

Recipes ship for every output-ladder entry in §5. Users can copy and edit them without touching the bundle.

11. Install and UX contract

The contract: install is one command, first run is one mode, zero YAML edits required. This is the amplifier-bundle-stories bar.

Installation:

amplifier bundle add --app git+https://github.com/michaeljabbour/amplifier-bundle-research@main
amplifier bundle use research

First run options:

# Interactive — start at /question
amplifier

# Recipe-driven — hand it an intent, get a finished artifact
amplifier run --recipe patent-brief "Rolling-ROI control for AI agent sessions"

# Mode-specific — e.g., critique an incoming document
amplifier run --mode critique --input incoming-claim.md

Full UX transcript in UX_MOCKUP.md.

12. Non-goals

Not a Denario replacement. Denario targets autonomous end-to-end paper generation, with the human as optional editor. This bundle explicitly keeps the human as primary author.
Not a domain-science toolkit. The bundle provides the research scaffolding. Domain-specific tools (protein folding, molecular dynamics, cosmology) stay in K-Dense's larger skill set and are opt-in.
Not a LaTeX IDE. Venue formatting is the final step, not the focus. The core value is the discipline upstream of formatting.
Not a plagiarism checker or IP-clearance tool. Prior-art search in /execute surfaces candidate references; legal clearance remains a human call.

13. Evaluation

The bundle ships with its own eval target: the RCE (Reflection Capability Evaluation) benchmark. A successful v1.0 release means:

Persona A produces a patent brief that a patent attorney rates ≥ 4/5 on defensibility
Persona B produces a workshop paper that passes the bundle's own /critique at ≥ 80% on CONSORT/STROBE-relevant items
Persona C can run /critique standalone against an incoming document and produce feedback rated ≥ "useful and correct" by ≥ 70% of reviewers

The first paper this bundle produces should be the paper about this bundle, evaluated against these targets. Dogfooding is the proof.

14. Roadmap

See ROADMAP.md. Short version:

v0.1 (now) — Spec, skeleton, 3 recipes fully working (patent-brief, policy-brief, empirical-paper)
v0.2 — All recipes, all modes, RCE eval harness
v0.3 — Venue pack (ICML, NeurIPS, Nature, USPTO, federal policy)
v1.0 — RCE-validated, dogfooded paper published, public release

15. Open questions

Should /execute for non-empirical recipes (patent-brief, policy-brief) always invoke a prior-art / evidence-gather sub-agent, or should that be a separate mode? Current answer: same mode, recipe-parameterized. Worth revisiting after first user tests.
How much of the K-Dense skill set should be registered vs opt-in? Current answer: 10 registered, 124 opt-in. Domain researchers will push for more.
Should behaviors be enforceable at the publish step (refuse to ship a draft that violates honest-pivot)? Current answer: warn, don't block. Users get final call.