research-vault 0.3.0

Research Vault — an adoptable, batteries-included AI research-assistant framework

An autonomous research crew you delegate to — not a tool you operate.

You hand Research Vault a research question. A hub agent — Alfred — plans the work, dispatches a crew of specialists, and runs the full research loop: literature review, experiment design, execution, analysis, synthesis. You stay in the conversation and approve at the gates that matter. You never drive the research loop from the CLI — Alfred does. (You run rv init once at setup; from there you converse.)

The aim is to give the scientist space to think and collaborate with the crew — not to worry about implementation and execution. The agents do the mechanical work; a human stays in the loop where judgment matters; and the disciplines that make research trustworthy are wired into the machinery so they bite.

The discipline is the headline. The autonomy is the point.

---

The crew

One flat, vault-level crew — six named specialists, each a hat composed from a shared charter plus a role doctrine. You talk to Alfred; Alfred coordinates the rest.

Alfred — Hub. After Alfred Pennyworth, the butler who runs the household. The single front door and sole orchestrator: the only agent that walks the DAG and the sole spawning authority. He plans the work, dispatches the crew, and surfaces every decision that needs you — and he never executes code or merges anything himself.

Wren — Architect. After Sir Christopher Wren. Owns the stack and the architecture map; vets every new dependency and keeps the system coherent. Wren designs — he doesn't lay stone.

Mason — Engineer. The master stonemason. Builds the code: features, tests, CI, and the authorized merge. Wren draws the plans; Mason builds them.

Ada — Researcher. After Ada Lovelace. The science itself — literature review, experiment design, retrieval-backed citations, analysis, and synthesis.

Argus — Reviewer. After Argus Panoptes, the hundred-eyed watchman. Independent verification: adversarial review and the honesty gates. Read-only by construction — and no agent reviews its own work.

Iris — Designer. After Iris, goddess of the rainbow and messenger of the gods. Figures and the surfaces through which the work reaches the world.

Least-privilege is stamped into each hat: coordinator-class hats get no shell (structural, not disciplinary), the reviewer is read-only, and the researcher carries web retrieval for support-checked citations. Nobody merges on their own authority — a merge executes only when an independent gate authorizes it.

---

How you actually use it

Once you've run rv init, you never drive the research loop yourself. It is a conversation:

1. You describe the work to Alfred in natural language — a question to investigate, a literature space to map, an experiment to run.
2. Alfred walks the DAG — he plans the research loop, dispatches each node to the specialist whose hat fits it (Ada for the science, Mason for the harness, Argus for verification, Iris for figures), and threads the artifacts between them.
3. You approve at the gates — at each human-go gate Alfred pauses, hands you an evidence packet, and waits. Nothing clears the gate without your explicit go; the crew cannot approve its own work.

The rv command line exists — but it is Alfred's control surface, not a human keyboard interface. You converse; Alfred runs the verbs. (The full verb reference is below, for the curious.)

---

Why it exists

Most "AI research assistant" tooling optimizes for output volume. The failure mode of an LLM in research is not slowness — it's confident fabrication: an invented citation, a metric that never traces to a run, a "passing" check that is green-and-empty, a result banked because it looked clean. Research Vault is built around stopping exactly that, mechanically:

• Anti-fabrication. Every specific — a number, a citation, a file — must trace to a real source. A citation needs a real retrieval, support-checked, not recalled from memory.
• Every outcome is a finding. A null result is a result. The loops are built to reach and record an honest null, not to fish for a positive.
• Verify the artifact, not the signal. "CI is green" is a claim to check against the artifact, never a fact to relay. A completed node is verified by its produced artifact's freshness, not by an agent's say-so.
• Honest pre-registration. The confirmatory plan is frozen (a content hash) before the run, and edits to the frozen set are caught structurally.
• Human-only approval. The crew cannot approve its own work. The approval gate is a mechanical trust boundary keyed on an interactive terminal — a dispatched agent has no TTY and is refused, regardless of flags.

Each discipline maps to a command and a gate. The code is the proof those disciplines are actually runnable; they are not slogans in a CONTRIBUTING file.

---

Where it fits

Research Vault stands on a wave of work exploring agentic research and discovery — systems that let agents plan, run, and reason about scientific work:

• AlphaEvolve — A Gemini-powered coding agent for designing advanced algorithms (DeepMind)
• AutoResearch — Andrej Karpathy
• The AI Scientist-v2 — Workshop-Level Automated Scientific Discovery via Agentic Tree Search
• AutoResearchClaw — aiming-lab

Our emphasis is a deliberate choice, not a verdict on any of these: discipline and doctrine, with a human in the loop. We want to give the scientist room to think and collaborate with the crew — freed from the mechanical work of retrieval, harness-building, running, and analysis, but never removed from the judgment. The agents do the mechanical work; the disciplines (anti-fabrication, honest pre-registration, verify-the-artifact, human-only approval) keep it trustworthy; and the human stays where human judgment belongs — the questions, the design, and the gates.

---

What the crew runs — the three loops

Research Vault ships three research loops as DAGs. Alfred walks each one node by node, dispatching every node to the specialist whose hat fits it. Two of them build notes/ (the crew's reasoning pillar); the third transforms notes/ into a submittable document (the user-facing deliverable pillar).

Literature review (rv review)

A pre-registered, saturation-gated review. The protocol must be approved before search fires (L-2 anti-fishing gate); a deterministic width-sweep (review-search) is screened by a thin agent judgment layer (review-screen) before the deterministic snowball walks forward (cited-by) and backward (references) (review-snowball). The raw corpus then passes a relevance gate: a mechanical off-domain pre-filter (review-relevance-screen) ahead of concept-tagging and curation (review-curate), then a cold, canary-verified re-check (review-relevance-verify) of every [NEW] paper before coverage-gate — below a 30% off-domain fraction it auto-prunes and the run proceeds; at/above threshold it HALT-DECLAREs. On coverage-gate GO, Phase-2 relate nodes auto-emit and fan out over every in-scope paper (no hand-run expand). On approve-review GO, the manuscript loop over the same scope auto-emits and auto-starts (no hand-run hand-off). OKF outputs: literature/*.md notes, concepts/, mocs/, and typed gap notes.

flowchart TD
    scope[review-scope] --> HG1[["[HG] approve-protocol"]]
    HG1 --> search[review-search] --> screen[review-screen] --> snowball[review-snowball]
    snowball --> relscreen[review-relevance-screen]
    relscreen --> curate[review-curate]
    curate --> relprep[review-relevance-verify-prep] --> relverify[review-relevance-verify]
    relverify --> GATE2[["coverage-gate (auto-resolved)"]]
    GATE2 -. "auto-emits" .-> relate["relate-*\n(Phase-2 fan-out)"]
    relate --> synthesize[review-synthesize] --> critic[review-coverage-critic]
    critic --> GATE3[["approve-review (auto-resolved)"]]
    GATE3 -. "auto-emits + auto-starts" .-> MS[["manuscript loop\n(new tree, slug == scope)"]]

Experiment (rv experiment)

A pre-registered study. The plan is critiqued and frozen before any harness is built; each main's harness is reviewed independently before the run fires; results gate conditional ablations; all findings are ratified before write-up. OKF outputs: experiments/*.md (pre-reg), findings/*.md.

flowchart TD
    plan --> critic[plan-critic]
    critic --> HG1[["[HG] human-go-plan"]]
    HG1 --> harness["harness\n(×N mains)"] --> hr[harness-review]
    hr --> HG2[["[HG] human-go-harness"]]
    HG2 --> run --> score --> analyze
    analyze --> HG3[["[HG] human-go-conditionals"]]
    HG3 -->|if threshold| cabl["conditional\nablations"]
    HG3 --> HG4[["[HG] human-go-findings"]]
    cabl --> HG4
    HG4 -.-> methods-update

Manuscript (rv manuscript)

Turns a saturated notes/ corpus into a submittable document, by type (type: lit-review — a survey/review paper — ships today; a future type: experiment-paper is designed for, not built). The organizing framework is chosen by an ensemble — N cold, independent lens candidates (each a different organizing lens) are synthesized into a single backbone by select-and-graft, then vetted by a cold, rejects-only critic before the gate auto-commits; every draft/revise round re-fires hard fidelity gates (hermetic references build, citation-resolve, coverage, equation-fidelity) plus the cold-agent-judged support-matcher (judge-emit writes the fan-out task set, the hub fans cold subagent judges out over it, judge-ingest assembles the verdict — fail-closed, never a config-gated skip); a 6-lens cold-agent-judge review board (N rounds, default 2, hardcap 3, skip-once-cleared) — depth, width, synthesis, self-containment, adversarial, instruction-following — scores FLOOR axes by MIN-across-lens (never average) plus a mandatory annotated-bibliography canary that must not clear. OKF inputs: literature/, concepts/, mocs/, gaps/. Output: manuscripts/<slug>/{report.md, sections/, references.md, figures/} (markdown only — no LaTeX).

flowchart TD
    scope[scope] --> L1[framework-lens-by-chronology]
    scope --> L2[framework-lens-by-mechanism]
    scope --> L3[framework-lens-by-outcome]
    scope --> L4[framework-lens-by-population]
    scope --> L5[framework-lens-by-theoretical-tension]
    L1 --> syn[framework-synthesize]
    L2 --> syn
    L3 --> syn
    L4 --> syn
    L5 --> syn
    syn --> fcritic[framework-critic]
    fcritic --> GATE1[["approve-framework (auto-resolved)"]]
    GATE1 -. "auto-emits" .-> sec["section(s)\n(type-generic, sequential)"] --> assemble
    assemble --> board[["6-lens board\n(cold judge-fanout, N rounds)"]]
    board --> GATE2[["approve-manuscript (auto-resolved)"]]

See doctrine/manuscript-loop.md for the full walkthrough (scaffold → framework lens-ensemble → autonomous Phase-2 emission → the 6-lens review board → the fidelity gates → the manuscript output) and its known limitations.

All three loops use the same underlying machinery: a DAG walker over typed nodes, with a grounding manifest that binds each node to the artifacts it reads and produces. [HG] nodes are human-go gates — the points where Alfred pauses and waits for you. Only one gate is ever human in the lit-review and manuscript loops: approve-protocol (Gate 1, the pre-registration checkpoint before any search fires). Every downstream gate (coverage-gate, approve-review, approve-framework, approve-manuscript) resolves autonomously through the gate-policy engine (review/autonomy.py) — the user receives the manuscript as the system's best version, with no "approve the result" gate and no provisional/vetoable bookkeeping. The experiment loop's four [HG] gates (pre-registration + per-main harness review + conditionals + findings) remain fully human — that loop's autonomy program is a separate, not-yet-built effort.

---

The crew's control surface

Everything below is what Alfred runs on your behalf. You don't type these — but they are readable, so you can see exactly what the coordination is made of.

How a loop runs (the DAG walk)

Alfred walks the DAG one dispatchable node at a time, using a deterministic brief emitter so no dispatch is hand-transcribed (hand-transcription is where context drifts):

rv dag status <run_id>              # 1. identify the next node (PENDING; reads verified)
rv dag brief  <run_id> <node_id>    # 2. emit the deterministic dispatch brief
#                                      3. dispatch that brief verbatim to the crew agent
rv dag complete <run_id> <node_id>  # 4. record SUCCEEDED/FAILED; the walker advances
rv dag tick   <run_id>              #    advance the frontier to the next gate
rv dag approve <run_id> <node_id>   #    human-go: the solo decision gate

The brief is a pure function of the node plus run state — byte-identical given the same inputs — so what a crew agent receives is grounded in resolved absolute paths, not a re-typed summary.

Core capabilities

• DAG research loops with typed nodes, afterok/watch edges that gate on artifact freshness, and in-session resolution (no background pollers/daemons).
• Deterministic crew briefs (rv dag brief) — every dispatch carries a fixed structural preamble (role framing, anti-fabrication, the return schema) plus the node's spec and resolved read/write paths.
• K-3 pre-registration freeze (rv plan check, rv plan freeze) — a structural shape-lint (no empty/TBD/"fallback" diagnosis cells; one-component-per-ablation) before the human-go, then a content hash of the confirmatory covers: set that is re-verified at findings — a post-freeze edit to the frozen set is caught, not trusted.
• Cross-project corroboration — Alfred declares a genuine edge between projects (rv project relate <a> <b> --kind <why>), then rv research corroborate ranks candidate evidence across declared peers by TF-IDF, an LLM judge confirms each, and a human reviews. Never auto-asserted.
• The mechanical approval trust boundary (rv approval, rv dag approve) — security = stdin.isatty(), full stop. A signed token path exists for non-interactive operators, but the crew cannot self-approve by construction.
• OKF typed notes — 8 note types (literature, concepts, methods, experiments, findings, mocs, datasets, gaps). Notes are pointers, not embeds: a datasets note points to its artifact (path/URL/DOI + content hash), never contains it. literature is two-layer: a cross-project CENTRAL CORE (intrinsic paper facts — ids, ## Result, ## Key equations, ## Related papers) lives once at cfg.literature_root/<citekey>.md; a thin per-project OVERLAY (role, position, concept-edges + a central: pointer) lives at the familiar project_notes_dir/literature/<citekey>.md — the same paper read once, never re-transcribed per project. rv literature list <project> reads a project's overlay against the central corpus ledger (zero recompute) to show which papers it draws on.

---

Install

pip install research-vault      # a lean 29-package research toolkit
rv --help

The rv CLI and every verb run clean even with the toolkit absent (all toolkit imports are lazy) — so pip install research-vault --no-deps works, and rv bootstrap populates an isolated .venv if you need the full stack later.

The 29-package core covers the model seam (litellm as the unified provider interface, plus the Anthropic SDK and a tokenizer), analysis (pandas, numpy, pyarrow, scipy, statsmodels, datasets), eval (inspect-ai, evaluate, sacrebleu, rouge-score), a multilingual set, integrations (wandb + weave for experiment tracking and automatic call-trace observability, pyzotero + keyring for Zotero citation management), and harness utilities. GPU-fragile local inference (torch, transformers, …) is opt-in behind an extra — it is never installed by default (CUDA-pinned wheels break CPU-only machines):

pip install research-vault[local]              # local GPU inference
pip install research-vault[local,serve-vllm]   # + a serving stack

Per-provider SDKs (openai, google-genai, …) and plotting libraries are not shipped — install them directly at your discretion. litellm covers most API targets without a dedicated SDK.

Prerequisites

• Python 3.12+
• An agent runtime — the ONE hard requirement. Claude Code (see Quick start below). There is no required API key: with the runtime installed and zero keys, rv check is GREEN (exit 0) and you can start immediately.

Everything else is a feature you unlock when you need it. A missing feature key is never a failure — it is locked until you add the key:

Feature	Unlocks	Get a key / access
Provider API key(s)	API-model experiments (any ONE provider)	console.anthropic.com/settings/keys · platform.openai.com/api-keys
s2	rv research find retrieval	semanticscholar.org/product/api
asta	rv research find --deep	share.hsforms.com (asta MCP access)
wandb	experiment observability + rv wandb pull	wandb.ai/settings
zotero	rv cite	zotero.org/settings/keys
compute	remote-cluster experiments	rv compute init

Provider keys are provider-plural (Anthropic, OpenAI, …). The asta access request needs an institutional email (not a personal gmail). Run rv onboard for a guided, idempotent setup that stores each key in your system keyring (never a plaintext file), and rv check to verify — it points to rv onboard for anything still locked. wandb, pyzotero, and keyring are core pip dependencies; asta is the one external prerequisite that is not a pip dep.

---

Quick start

Research Vault runs on Claude Code.

pip install research-vault
rv init myvault      # scaffold a new vault
cd myvault           # enter it
rv onboard           # guided setup: keys, compute, inline-approval token
rv start             # launch Claude Code as Alfred in the vault

cd must come before rv onboard — onboarding writes the compute manifest into the vault, so it runs from inside. (rv init can't cd you in; a subprocess can't change the parent shell.) Onboarding is optional and re-runnable — the runtime alone is enough to start; rv init also offers to run it for you at the end.

On Claude Code, rv init scaffolds a CLAUDE.md and the crew under .claude/agents/, rendered with per-role tool grants and model aliases — so your session boots as Alfred (the hub), the coordinator you converse with, and the role hats become the subagents he dispatches. The human-only approval boundary holds mechanically: the approve-gate keys on an interactive TTY, which dispatched crew subagents never have.

rv init writes a QUICKSTART.md into the vault — the full walkthrough, including compute onboarding and an example session with Alfred. A real project is its own git repo; register it with rv project add (or stand up a fresh one with rv project new). After a pip install --upgrade research-vault, run rv update to pull the upgraded framework (doctrine, CLAUDE.md, the crew hats) into an existing vault — your notes, projects, and edits are preserved.

---

Status

This is alpha. The architecture is complete and the loops run end-to-end. It is not battle-tested across many adopters yet, and it is evolving. The disciplines are enforced by code you can read.

License

AGPL-3.0-or-later. See LICENSE. (Relicensed from MIT at v0.3.0 — see DEVLOG.md 2026-07-08. Releases 0.1.0-0.2.8 remain MIT for already- distributed copies; the flip is go-forward only.)

Contributing

See CONTRIBUTING.md. One rule above the rest: changes to a discipline are doctrine changes — they go through the doctrine, not around it.