workbench/data/systems.json

{
 "about": "Five systems built the same way as Oracle7: no language model, every decision with a receipt, measured on held-out data once. Each is its own repository in the release; numbers are from each one's model card.",
 "items": [
  {"id": "oracle7-deepspace", "name": "DeepSpace", "does": "Fault response for deep-space spacecraft. It reads telemetry, forms confirmed facts, then fault hypotheses, then admissible responses, each with a receipt. A self-maintenance gate admits rules learned from its own verified episodes.",
   "measured": ["98.9% correct on 1,700 fresh single-fault scenarios, 0 harmful actions (a limit-to-safe-mode baseline: 29.4%, 200 harmful)", "99.2% on 240 compound faults, 0 harmful", "9 learned rules admitted; all 4 planted wrong rules rejected"],
   "runs": "Python standard library, a fraction of one CPU", "use": "spacecraft, satellites, any machine with telemetry and a fault table"},
  {"id": "oracle7-trafficcop", "name": "Trafficcop", "does": "A decision service. Given a state and a set of options it returns one choice with a receipt: impossible options are struck, obligations are bound, attested outcomes rank the rest, and it abstains when nothing is warranted.",
   "measured": ["30 of 32 held-out routing decisions right (always-first: 8, random: 7)", "order-consistent on all 16 held-out cases", "about 0.1 ms per decision; it replaced a language-model decision service behind the same contract"],
   "runs": "Python standard library", "use": "routing work between tools, agents and specialists"},
  {"id": "oracle7-chameleon", "name": "Chameleon", "does": "Builds websites and components by reconstruction. Its operators are repair operations that act only when their preconditions hold, and a build counts as complete only when contracts written before it ran pass in a supervised browser.",
   "measured": ["0 false completions on held-out repositories (1 of 4 completed, 3 reported blocked with the reason)", "13 complete / 13 blocked on dev and transfer runs"],
   "runs": "Node and Python, a headless browser", "use": "the coding system of the family: site and component builds that prove themselves"},
  {"id": "oracle7-vision", "name": "Vision", "does": "Sees without trained weights. Fixed image descriptors are bound into one vector (HRR, D = 2048) and compared with stored, labelled examples; every answer cites the examples that voted, and it abstains when the vote is close. It reads text by rendering glyphs from open fonts.",
   "measured": ["MNIST 98.58%, Fashion-MNIST 90.46%, CIFAR-10 62.99% (official test sets, run once)", "at about 70% coverage: 99.85% / 98.77% / 76.39%"],
   "runs": "numpy, Pillow", "use": "checking images, reading screens and labels, verifying what Artist drew"},
  {"id": "oracle7-artist", "name": "Artist", "does": "Makes images by citation, with no generative model. A prompt is parsed into objects, counts, colours and positions; every object is a real photographed instance, cited by its image and annotation id; the layout satisfies the prompt by construction and Artist re-measures every clause from its own pixels.",
   "measured": ["GenEval-style 50.9% on 2,212 held-out images, scored by an independent judge (an approximation of the official scorer)"],
   "runs": "numpy, Pillow", "use": "stills for films and posts where every part has a source"}
 ]
}