AI Research Preflight

Deterministic, local-first Agent Skills for reviewing scientific-compute plans and declared experimental readiness before execution.

This repository deliberately separates two jobs:

Skill	What it does	What it never does
Research Compute Planner	Checks a sanitized research manifest, abstract DAG, data-governance metadata, pinned tools, and bounded resource estimates	Read raw research data, create a formulation, choose a scientific conclusion, or launch a cloud job
Formulation Readiness Check	Checks local formula, policy, and evidence manifests for arithmetic, units, declared bounds, pair coverage, process dependencies, and experiment gates	Generate or optimize a formula, infer undeclared compatibility, or claim safety, stability, efficacy, compliance, or real-world feasibility

Both Skills use Python 3.10+ standard library code. Their deterministic engines make no network requests, use no telemetry, and write nothing unless the caller explicitly redirects standard output.

Why preflight instead of autopilot

AI can help structure a scientific plan, but execution carries data-governance, cost, scientific, and experimental consequences. These Skills create bounded, inspectable contracts before a separate human-approved execution step.

The research Skill plans a provider-neutral handoff. It does not discover cloud credentials, select an account, submit a workflow, or estimate a cloud bill.

The formulation Skill checks only rules and evidence supplied by the user. A clean report means "ready for the declared experiments," not "the formula is feasible."

Preview and install

The planned first release is v0.1.0. These pinned commands remain marked as unverified until the public tag exists and clean installs are recorded.

Preview either Skill before installation:

gh skill preview ProfesseurHaipeng/ai-research-preflight \
  plan-research-compute@v0.1.0

gh skill preview ProfesseurHaipeng/ai-research-preflight \
  check-formulation-readiness@v0.1.0

Install at user scope for Codex:

gh skill install ProfesseurHaipeng/ai-research-preflight \
  plan-research-compute@v0.1.0 \
  --agent codex --scope user

gh skill install ProfesseurHaipeng/ai-research-preflight \
  check-formulation-readiness@v0.1.0 \
  --agent codex --scope user

GitHub warns that third-party Skills are not verified. Inspect each pinned tree and understand its scripts before installing.

Research Compute Planner

Use a sanitized JSON manifest containing metadata, an abstract step graph, version-pinned tool references, resource envelopes, and hard limits. Do not put raw sequences, images, personal records, access URLs, or credentials in the manifest.

Use $plan-research-compute to check this sanitized research-compute manifest.
Produce a provider-neutral preflight and do not read data or launch anything.

Run the deterministic engine directly:

python3 skills/plan-research-compute/scripts/research_compute_plan.py \
  request.json --format json

Its output distinguishes invalid input, blockers, review items, and a bounded plan that still requires human approval. Missing benchmarks remain explicitly unbenchmarked; the engine does not invent precise resource figures.

Formulation Readiness Check

Keep the private formula, the user's declared policy, and the evidence index in three separate local JSON files. Public examples use opaque IDs only.

Use $check-formulation-readiness to run a redacted local preflight against my
supplied formula, policy, and evidence. Do not generate or improve the formula.

Run the deterministic engine directly:

python3 skills/check-formulation-readiness/scripts/formulation_readiness.py \
  --formula formula.json \
  --policy policy.json \
  --evidence evidence.json \
  --as-of 2026-07-12 \
  --format json

Default reports use opaque identifiers and do not echo component names, raw amounts, policy thresholds, or evidence contents.

Shared boundaries

• No raw research data or personal records in public examples.
• No real formulas, private policies, component identities, or unpublished evidence in this repository.
• No HTTP calls, cloud SDKs, shell evaluation, telemetry, or hidden uploads.
• No wet-lab instructions, clinical recommendations, or scientific conclusions.
• No safety, stability, efficacy, cost, compliance, or universal-compatibility guarantees.
• Every executable result remains subordinate to domain review and explicit human approval.

See public scope, tested compatibility, and inspiration and provenance.