Skip to content

Research Milestones

Key milestones in the autonomous FO infusion research program.

This KB is built by an autonomous research agent (Claude Code + Opus) running continuous sessions against NASA's TechPort database and supplementary sources (USASpending, SBIR.gov, web search). The agent started on 2026-04-05.

Timeline

Session Date Milestone Cumulative state
1 2026-04-05 First session. 7 initial org pages (TMT, CDI, NanoRacks, etc.) 7 org pages
6 2026-04-06 Model switch: Sonnet → Opus. Deeper document reads, richer synthesis. ~30 org pages
19 2026-04-06 Cryogenic cluster topic page. First major cross-cutting theme. ~70 org pages, 7 topics
22 2026-04-06 94% triage coverage. 405/430 projects assessed. 108 linkages, 6 topics
29 2026-04-06 First complete pass. All 430 projects triaged. All FO documents accounted for (17 files, 7 read in full). 107 org pages, 132 linkages, 82 deep investigations
30 2026-04-06 Deepening phase begins. Archetype analysis, early page refreshes, topic synthesis. 12 topics
33 2026-04-06 Gateway cancellation impact assessed. First time-sensitive news event integrated. 133 linkages
40 2026-04-07 15 new FO projects appear in TechPort (430→445). Re-triaged to 100%. 445/445 triaged
42 2026-04-07 New industry cohort. Impossible Sensing, GOEPPERT, UTS, AeroFly, Space Dust R&T, Mentium deep-investigated. 118 org pages, 137 linkages
44 2026-04-07 First lint pass. Zero contradictions, zero orphan pages across 117 org + 16 topic pages. 65 stale files identified. KB structurally sound
45 2026-04-07 Stale page refresh begins. 11 earliest pages updated. 5 material findings (Voyager IPO $3.8B, Nexolve acquired, Blue Moon accelerated, PTD-4 anomaly corrected). best-sources.md populated
47 2026-04-07 CACI lasercom deep dive. Two FO projects converge on DSOC mission through independent paths. DoD lasercom pipeline mapped. 17 topics, 145 linkages
80 2026-04-07 150 linkages reached. 121 org pages, 21 topics, 150 linkages
89 2026-04-07 Momentus discovery. FO flight provider with $7.2M Armstrong contracts — previously assessed as near-bankrupt. 122 org pages, 154 linkages, 95 deep investigations
100 2026-04-07 Session 100 — final session. 100 autonomous research sessions in ~2.5 days. All 445 FO projects triaged (100%), 134 org pages, 22 topic pages, 154 validated linkages, ~112 deep investigations. High-signal projects fully covered; remaining backlog is low-signal academic work. Stale refresh cycles maintaining currency. Agent stopped by design. 134 org pages, 22 topics, 154 linkages, 445/445 triaged

Key metrics over time

Metric Session 1 Session 29 Session 44 Session 89 Session 100
Projects triaged ~10 430/430 (100%) 445/445 (100%) 445/445 (100%) 445/445 (100%)
Org pages 7 107 117 122 134
Topic pages 0 ~10 16 21 22
Validated linkages 0 132 144 154 154
Deep investigations ~7 82 92 95 ~112
Surprises filed 0 ~15 ~25 ~35 ~35

Model comparison

Sessions 1-6 ran on Claude Sonnet. Sessions 8+ run on Claude Opus. A comparative evaluation (2026-04-07) found Opus produces meaningfully better output across all dimensions: document reads (Sonnet skips them), provenance depth (2-3x more downstream links per page), non-obvious connections, evidence quality (counter-queries and verification), synthesis (named archetypes), and source diversity. See the research log for session-by-session details.

Infrastructure

  • Execution: Claude Code CLI in --print mode, launched by launchd with KeepAlive: true
  • Pacing: TCP-style AIMD congestion control — additive increase on success, multiplicative decrease on rate limit
  • Success rate: ~8% of session attempts produce KB commits (rest are rate limit bounces)
  • Data sources: TechPort MCP (primary), USASpending.gov, SBIR.gov, web search
  • KB format: Markdown wiki, published via MkDocs Material to GitHub Pages