Score a research brainstorm without picking a winner

What this does, and what it does not

The scientific-brainstorming agent skill ships three standard-library Python tools. validate_register.py checks an idea register's structure: session and participants, each idea's provenance (human, AI-assisted, literature-inspired), linked assumptions, predicted observations, uncertainties, evidence and idea status, clusters and the decision log. evaluate_matrix.py computes a fully disclosed weighted additive matrix from a criteria file and a scoring sheet: score = 100 x sum of normalized weight x direction-normalized rating, the score interval implied by low/high ratings, and the score and rank range when each weight moves by --weight-delta one at a time. It leaves decision null. session_scaffold.py writes an empty register. This page runs JavaScript ports of all three. Checked against the Python on fuzzed inputs under CPython 3.14 and 3.13 (see the notice for the count), stdout, stderr and exit code were identical in every case, including the ones where the Python stops with an uncaught exception, which the page reports as a crash rather than inventing a report. Under CPython 3.12 the only difference is the wording of one JSON error (a trailing comma).

The page adds its own cross-checks, labelled as such: ideas scored but missing from the register, presentation neighbours whose intervals overlap, a criterion carrying more than half the weight, and gate columns (ethics, safety, consent...) marked as needing review. A validator pass is not ethics, biosafety or institutional approval, a score is not evidence, and a rank is not a decision - the skill is explicit, and so is every review. The paid lanes read only your files, the tools' results and your notes. The draft lane may not add ideas or ratings; the page re-runs the skill's tools on every draft and checks every review against them. Derived from the agent skill @k-dense-ai/scientific-brainstorming (k-dense-ai/scientific-agent-skills, MIT; see the notice).

Around the tools, the page does the free bookkeeping a session needs next. It hands out one blank rating sheet per participant and merges the filled sheets back as the median, with the lowest and highest rating as the interval. It shows what changed in ranks, scores and flags since the last check of the same session, and lists the register's assumptions with who checks each one and how. After a review, it turns the result into an action checklist, pauses or stops the ideas the review recommends once you confirm, and writes the decision owner's own decision into the decision log. The page never makes that decision.