{"id":"operating-system","title":"Lupine Operating System","subtitle":"How the research loop fits together: agenda, ledger, evidence, and correction.","category":"operations","tags":["meta","system"],"source":"articles/docs/operating-system.md","lang":"en","words":164,"readMinutes":1,"toc":[{"depth":2,"text":"Loop","id":"loop"},{"depth":2,"text":"Durable State","id":"durable-state"},{"depth":2,"text":"Resource Lanes","id":"resource-lanes"},{"depth":2,"text":"Design Constraint","id":"design-constraint"}],"html":"<h1 id=\"research-loop-architecture\">Research Loop Architecture</h1><p><code>glim-think</code> is the mastermind. Everything else is substrate.</p>\n<h2 id=\"loop\">Loop</h2><ol>\n<li>Sense: ingest external evidence and internal run output into the ledger.</li>\n<li>Shape: turn evidence into claims, hypotheses, and discriminative tasks.</li>\n<li>Allocate: request the model, sandbox, browser, compute, or storage resource\nneeded for each task.</li>\n<li>Execute: agents claim tasks, run tools, and store traces.</li>\n<li>Verify: claims are promoted only when backed by reproducible evidence.</li>\n<li>Broadcast: the live system reports what changed and what is next.</li>\n</ol>\n<h2 id=\"durable-state\">Durable State</h2><ul>\n<li><code>records</code>: benchmark evidence.</li>\n<li><code>claims</code>: structured findings.</li>\n<li><code>pending_experiments</code>: local or sandboxed experiments waiting to run.</li>\n<li><code>intelligence_tasks</code>: horizon-aware work queue.</li>\n<li><code>resource_requests</code>: requested capabilities for queued work.</li>\n<li><code>operating_cycles</code>: every agenda expansion or scheduled refresh.</li>\n</ul>\n<h2 id=\"resource-lanes\">Resource Lanes</h2><ul>\n<li><code>workers-ai-screening</code>: cheap first pass.</li>\n<li><code>capable-model-route</code>: escalated reasoning.</li>\n<li><code>browser-data-scout</code>: live source inspection.</li>\n<li><code>tier-4-sandbox</code>: LAMMPS, Cargo, and git workspaces.</li>\n<li><code>artifact-storage</code>: R2 evidence and reports.</li>\n<li><code>d1-ledger-read</code> / <code>d1-ledger-write</code>: durable research memory.</li>\n</ul>\n<h2 id=\"design-constraint\">Design Constraint</h2><p>No code exists only to persuade. If a surface does not help the system think,\nexecute, verify, or operate, it should live outside this repo.</p>\n"}