[Cross-cutting] Biomni + K-Dense competitive intel (WS1)

[Cross-cutting] Biomni + K-Dense competitive intel (WS1)

Task

• ID: task-id-pending
• Type: recurring
• Frequency: weekly (target Monday 14:00 UTC)
• Layer: Cross-cutting

Goal

Keep SciDEX current on what Biomni (Stanford / Phylo) and K-Dense (Biostate AI /
K-Dense AI) are shipping, and feed those deltas to Senate + the competitive
biotools quest so we can react inside a week rather than a quarter. Without
this driver, we only learn about their new releases from user reports — which
is too slow to compete or co-adapt.

What it does

• Scans GitHub: snap-stanford/Biomni (commits, releases, issues, discussions,

wiki), K-Dense-AI/claude-scientific-skills (commits + skill additions),
K-Dense-AI/k-dense-byok (BYOK desktop client). Diffs against the previous
week's snapshot stored under docs/bio_competitive/weekly/.

• Fetches recent papers via search_pubmed / search_biorxiv / search_arxiv

that cite Biomni (bioRxiv 2025.05.30.656746) or K-Dense Analyst
(arXiv 2508.07043) — logs new citations with abstract snippet.

• Fetches blog feeds where available (biomni.stanford.edu, k-dense.ai,

phylo.ai, biostate.ai) via fetch_url. Rate-limited to 1 req/sec.

• Aggregates into a weekly markdown report under

docs/bio_competitive/weekly/YYYY-MM-DD.md with sections: New tools /
skills, New papers / citations, Funding & customer signals, Benchmark
claims, Deltas vs SciDEX capability map, Recommended response tasks.

• Updates docs/bio_competitive/comparison_matrix.md when a cell changes

(new benchmark number, new tool count, new license event).

• Emits agent_contributions (type=competitive_intel) per report.
• Creates at most 2 Orchestra tasks per week for response actions (new skill

to adopt, new benchmark to run against, license change to review) —
hard-capped to prevent task flood.

• Records inaccessible sources (LinkedIn replays, paywalled PDFs) in

docs/bio_competitive/access_notes.md rather than silently dropping them.

Success criteria

• Weekly report lands every Monday (≥48 reports / year).
• Report file size ≥8KB with ≥5 substantive bullets (not a stub).
• ≥90% of Biomni + K-Dense commits in the trailing 7 days appear in the

report within 7 days of the commit (measurable via GitHub diff).

• Comparison matrix updated whenever upstream metrics change.
• Access notes file grows over time with explicit, non-fabricated gaps —

every blocked source listed with URL + reason + escalation owner.

• Run log: commits scanned, papers fetched, blog posts fetched, tasks

created, sources blocked.

Quality requirements

• No stubs: a report with only "no changes this week" is acceptable only

when the diff is truly empty; otherwise it must surface the actual
changes. Link to quest_quality_standards_spec.md.

• No fabrication. If a LinkedIn replay or paywalled paper is inaccessible,

log it in access_notes.md and move on — never invent content.

• When scanning ≥10 newly-cited papers in a cycle, use 3–5 parallel agents

to fetch + summarize in parallel. Single-agent only for small batches.

• Log total items processed + retries so we can detect busywork (identical

report week over week with no reader engagement → surface Senate task).

• Cite upstream sources with full URL + access date in every bullet.

Work Log

2026-04-16 — Slot minimax:74 (WS1 weekly scan)

• Report: docs/bio_competitive/weekly/2026-04-16.md — 9.8KB, 169 lines, ≥8 substantive bullets
• K-Dense BYOK (k-dense-byok): 22 commits since Apr 9, 8 releases (v0.2.14→v0.2.21); notable: Claude Opus 4.6 default model (Apr 13), parallel-web skills integration (Apr 13), message queue (Apr 13), custom MCP server loading (Apr 12), parallel API key handling in sandbox
• K-Dense claude-scientific-skills: 3 minor commits since Apr 9; no new skill catalogue changes
• K-Dense blog: New BixBench-Verified-50 score 90.0% (45/50) — a cleaned/curated benchmark variant vs the original 29.2% on full BixBench; new GPU skill (58x speedup claim, 12 libraries); Research Grant Program (90% academic discount); 3 new pharma use cases (Apr 3)
• Biomni: Zero commits since Jan 15; new issue #297 (Apr 15) adding Lib-INVENT tool (Roche cheminformatics library enumeration); pre-commit auto-update Apr 13
• Papers: arXiv search returned no new papers citing Biomni or K-Dense; paper_cache found 1 potentially Biomni-citing paper ("Network-Integrated Reverse Vaccinology Using Biomni-Prioritized Features")
• Comparison matrix: No cell update this cycle — BixBench 90.0% is on BixBench-Verified-50 (different denominator), cannot directly replace 29.2% entry without methodology clarification
• Recommended tasks: (2) — [Forge] Run SciDEX benchmark on BixBench-Verified-50; [Senate] Review K-Dense academic grant pricing response
• Access notes: No new blocked sources
• Commits scanned: Biomni 0, K-Dense skills 3, K-Dense BYOK 22; Blog posts 24; Papers searched multiple queries
• Pushed: docs/bio_competitive/weekly/2026-04-16.md (commit 4b4bfabdb)
• Result: Done — substantive report with actionable signals; K-Dense BYOK acceleration is the dominant theme

Related tools / packages

• search_pubmed, search_biorxiv, search_arxiv from tools.py
• fetch_url for blog / release page fetching
• gh api repos/snap-stanford/Biomni/commits?since=... via shell
• GitHub REST endpoints: /repos/{owner}/{repo}/commits, /releases,

/issues, /contents/skills

• agent_contributions schema for credit emission
• Output path: docs/bio_competitive/weekly/YYYY-MM-DD.md

Tasks using this spec (1)

[Cross-cutting] Biotools competitive intel — Biomni + K-Dens

File: task-id-pending_biotools_competitive_intel_spec.md

Modified: 2026-04-24 07:15

Size: 5.7 KB