# Brainstorm Desk

> A desk for early-stage research brainstorms at https://brainstorm-desk.skillsafe.ai/. Paste or drop a session's idea register (JSON), criteria (JSON) and scoring sheet (CSV); the browser runs JavaScript ports of the standard-library tools bundled with the scientific-brainstorming agent skill - validate_register.py, evaluate_matrix.py and session_scaffold.py - with stdout, stderr and exit codes identical to the Python, for free. A metered run (model gpt-terra) then drafts the three files from session notes, or runs the skill's adversarial review over a shortlist. No idea is ever selected: decision stays null. Derived from the agent skill @k-dense-ai/scientific-brainstorming (k-dense-ai/scientific-agent-skills).

## What the free lane computes (in the browser, nothing uploaded)

- validate_register.py: session and participant ids, each idea's provenance (origin, contributors listed as participants, recorded stage, source refs required for literature-inspired ideas, AI tool required for AI-assisted ones), linked assumptions (missing and unlinked), predicted observations, uncertainties, evidence and idea status, exact-after-normalisation duplicate statements, clusters and the decision log. Structure only - never truth, novelty, feasibility or approval.
- evaluate_matrix.py: score = 100 x sum of normalized weight x direction-normalized rating; the score interval implied by <criterion>_low/_high ratings; the score and rank range when each weight moves by --weight-delta (default 0.10) one at a time with renormalisation; ties broken by idea id for presentation only; decision null.
- session_scaffold.py: an empty schema 1.1 register from a form.
- Page cross-checks (labelled as the page's): ideas scored but not in the register, registered ideas not scored, presentation neighbours whose intervals overlap, a criterion carrying more than half the weight, no interval columns, gate columns (ethics, safety, consent...) marked as needing review, AI-assisted origins; and a browser status (ready_for_decision_owner / revise_first / not_reviewable).
- Checked against CPython 3.14 and 3.13 on 118,000 fuzzed runs with identical output (see NOTICE.txt).
- Exports: matrix.json and validation.json exactly as the scripts write them, a matrix CSV, a Markdown summary, the three files and the equivalent command lines.
- Private ratings: one blank rating sheet per participant (with each criterion's scale and direction); filled sheets merge back into the scoring sheet as the median, with min/max as <criterion>_low/_high, keeping the sheet's other columns.
- The loop: what changed since the last check of the same session (rank and score moves, flags resolved and added), and the register's assumptions with their owner and test.
- After a review (free, in the browser): an action checklist (Markdown or CSV) of confirmed flags, dispositions, gates, questions and untested assumptions with their owners; pause/stop dispositions applied to idea status after confirmation; the decision owner's own decision written as a decision_log record. The page never decides.

## The metered lanes (task field)

- draft: register, criteria and scoring sheet written from session notes as a scribe - no ideas of the model's own, pseudonymous participant ids, ratings copied from the notes or left blank for the team. The page re-runs the skill's tools on the draft and checks every rating against the notes.
- review: every flag answered; for each shortlisted idea the skill's adversarial-review template (strongest version, counter-observation, at least two alternative explanations, measurement and sampling failure, harm or misuse, mitigation, residual uncertainty, disposition, next action); fragility, gates, criteria issues, missing perspectives, questions and what is not established. Status never looser than the browser's; decision always null. Reviews can be added to the register as AI-assisted records.

Input: {task, notes?, register?, criteria?, scores?, facts? (JSON string built by the page), question?}. API: https://brainstorm-desk.skillsafe.ai/api.html

## Limits

Scores are judgments and decision aids, not evidence; a valid register is not ethics, biosafety, regulatory or institutional approval; the review checks no literature. The review covers at most 12 ideas per run.

## Source

Agent skill: https://skillsafe.ai/skill/@k-dense-ai/scientific-brainstorming (repository https://github.com/k-dense-ai/scientific-agent-skills, skills/scientific-brainstorming, MIT). Cite: Kassis et al. 2026, arXiv:2609.00065. See https://brainstorm-desk.skillsafe.ai/NOTICE.txt