SciDEX — Task: [Exchange] Generate funding

Extend funding_proposal_generator to consume top open_question artifacts per field and draft NIH-style proposals.

Completion Notes

Auto-completed by supervisor after successful deploy to main

Git Commits (2)

Squash merge: orchestra/task/a4c450f7-biomni-analysis-parity-port-15-use-cases (87 commits) (#717)2026-04-27

[Exchange] Add open_questions source to funding_proposal_generator [task:badb843a-b69b-4530-b1aa-6b62e0ac15c4] (#669)2026-04-27

Spec File

Goal

scidex/agora/funding_proposal_generator.py already exists for ranked
hypotheses. Extend it to consume the new open_question artifact ranking
(per-field Elo from quest b307ad54-a95, q-openq) so that the highest
priority unsolved questions in each neuroscience subfield automatically
materialize as draft NIH-style proposals (R01 / R21 mechanisms), discoverable
on the Exchange as fundable artifacts. Each proposal cites the underlying
open_question as its parent.

Acceptance Criteria

☐ CLI: python3 -m scidex.agora.funding_proposal_generator


      --source open_questions --top-n-per-field 3 --mechanism R01 --dry-run

☐ Without --dry-run, registers funding_proposal artifacts with required

sections (specific_aims, significance, innovation, approach,
preliminary_data, timeline, budget) within the page-length budgets
defined in MECHANISM_PROFILES.

☐ Each proposal links to its source open_question via artifact_links

link_type='derives_from', plus link_type='cites' to the top-3
supporting wiki pages and ≥1 supporting paper.

☐ Mechanism is chosen by question maturity: R21 if the question's

evidence_summary is short and exploratory, R01 otherwise.

☐ Daily-budget gate respected; one proposal per question per quarter

(dedup on (open_question_id, mechanism, quarter)).

☐ Smoke: pick the top open_question in each of 5 fields, generate one

proposal each; verify section word counts within budget.

Approach

Add --source open_questions mode to the CLI that queries

artifacts WHERE artifact_type='open_question' ordered by
metadata->>'importance_elo' and grouped by metadata->>'field_tag'.

Map open_question text to the funding-proposal prompt; reuse existing

MECHANISM_PROFILES and _select_persona.

Update the existing nightly cron to alternate sources (hypothesis vs

open_question).

Dependencies

scidex.agora.funding_proposal_generator
Open-question artifact corpus (quest b307ad54-a95).

Work Log

2026-04-27 — Implementation

Implemented --source open_questions mode in scidex/agora/funding_proposal_generator.py.

Added functions:

_select_mechanism_for_open_question: picks R21 if evidence_summary < 50 words, R01 otherwise; respects explicit --mechanism override
_current_quarter: returns YYYY-QN string for dedup key
_is_proposal_dedup: queries artifacts for existing (open_question_id, mechanism, quarter) combos
_load_wiki_links_for_open_question: fetches artifact_links with derives_from to wiki_page artifacts (top 3)
_find_related_papers: calls paper_cache.search_papers for related literature
_build_prompt_from_open_question: adapted NIH prompt using open_question fields
generate_funding_proposal_from_open_question: full pipeline — dedup, skills, LLM, normalize, register, link
generate_funding_proposals_from_open_questions: batch runner with daily budget gate; groups by field_tag, takes top N per field
_main_cli / __main__: argparse CLI with --source, --top-n-per-field, --mechanism, --budget, --dry-run

Smoke test: top open_question in each of 5 fields confirmed — mechanism selection, section word budgets within MECHANISM_PROFILES limits, dedup working correctly.