← Prometheus7 · Oracle7Swarm7 · Oracle7 4.9 · 23 October 2026

Swarm7

You bring Oracle7's swarm a problem. It works out what the problem needs, builds the parts it does not have, tests each one on cases it has never seen, and composes parts until the problem is solved. Every answer, program and plan it returns can be traced step by step to the parts that made it. There is no neural network in it, and it runs on an ordinary CPU.

Released in Oracle7 4.9 on 23 October 2026, after its sealed release test passed · huggingface.co/psiloceyeben/oracle7-slim

The trailer · 62 s

Live in the harness, on the released 4.9 · 71 s · real output, long waits shortened and marked

The loop

a missing part is named and goes back to be built Survey names what the problem lacks Build tables, formulas, programs Examine held-out exam; only an admit stays Represent key, Tree node, typed signature Combine joins every part whose types fit Work research, code, plans; each hop checked Keep filed on the spine; recalled first

How it learns

Nothing is trained. Learning is building a part: a table of facts from a cited source, a formula composed from worked examples, a program assembled from moves that already passed their tests. A part is kept only if its held-out exam admits it: at least 3 more right answers than the core alone, no more than 2% wrong answers added, no more trick questions answered than the core, and inside its lane and time budget.

Why it compounds

A part's worth is what it does alone plus every path it joins. Parts connect wherever one part's output type, with its unit, is another's input. With K parts there are up to K(K−1) two-step paths and about K3 three-step paths, so each new part is worth more than the one before it. In the release test, a part built for one problem was used to solve the next.

Why it stays small

Every part is a representation: a key, a place on the Tree of Life, and a typed signature (what it takes, what it gives). Its vector is regenerated from the key whenever it is needed. Combinations are worked out from the signatures, never stored. Verified paths are filed in spine memory, ten vectors rebuilt from a short index of entries. The representations of a run's parts come to a few kilobytes.

Why it can be checked

Research answers come from a part that holds the fact, cited to its source. Code is run in a separate interpreter against its worked examples, with units as types. Plans run each step and check it before the next. When it cannot finish, it says which step failed and which part is missing. When a question is declined, the decline specialist reads why: it rereads a question form it did not follow, reads a missing thing's value at the table's own source with its unit, names a gap for the builder, or keeps the decline when the thing does not exist.

Sealed release test v22 · written and hashed before the code froze · run once · no person in the loop

Every bar held

The swarm was pointed at four problems it had never seen, in one workspace.

Parts admittedneeded 5 or more6
Precision of answered workneeded 0.90 or more1.0
Plans rightneeded 3 of 44 of 4
Library tests passedneeded 95%40 of 40
Research rightneeded 2 of 33 of 3
Made-up things answeredneeded 00
Decline specialistneeded 0.90 right7 of 7

The four plans

Volcanoesclimb Mount Teide, 3,715 m, at 350 m an hour10.61 h
Canalscross the Panama Canal, 82 km, at 8 km/hbuilt its missing part mid-plan10.25 h
Waterfallsabseil Sutherland Falls at 20 m a minuteits 248 m tallest drop; Wikidata also gives 581 m12.4 min
Canal feesthe Suez Canal at 12 km/h, a pilot at $90 an hourused the part it built for the canals$1,449.75

Failed first, and published

v19"1.6 m/s²" misread, so a number was taken from a source tag (the one wrong answer); "How long is X?" not read as a question about lengthfailed
v20a table answered a ship name several ships share with one of them, and was refused; two parts got the same name; hours and seconds were one typefailed
v21the exam counted the core's own wrong answers against a new tablefailed

Each failure got a general fix, tested on a fresh sealed set. An earlier design (a loop over the questions the release declined) built 18 parts unattended, 7 of them composed from two routes it had built itself; it failed its parts bar and was not shipped. All of it is in the changelog.

Live · the released 4.9 slim tree · recorded 23 October 2026, 06:04 UTC · excerpt of the real output

A walk across a bridge

$ python oracle7.py update this install 44f51e06832a2b81 oracle_release_runtime_w436 published 44f51e06832a2b81 oracle_release_runtime_w436 (anchored 2026-10-23T05:52:30Z) this install is the published release $ python oracle7.py swarm point bridges/target.json survey 1: What is the length of Golden Gate Bridge? | What is the length of Brooklyn Bridge? | What is the length of Akashi Kaikyo Bridge? table-length: dev WOULD PASS, held-out ADMIT compose: walk.txt formula-time: dev WOULD PASS, held-out ADMIT gap (0.0, 0.0, 1.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0)|min: composing minutes.txt from the reserve formula-time-2: dev WOULD PASS, held-out ADMIT - ask: How long is the Oresund Bridge? - after: [answered] Øresund Bridge: length is 7845 metre. [wikidata:Q297871] (path declines:not_held:table-length>source:Q297871:P2043) - ask: How long is the Zorvath Bridge? - after: [clarification] I didn't fully parse that, so I won't guess. I can summarize a topic, compare two things, list facts, give a timeline, exp… - plan: time (min) from {'distance (km)': 'ask: How long is the Golden Gate Bridge?', 'walking speed (km/h)': 5} -> **['time (min)', 32.8489056]** - {"step": "built", "part": "formula-time-2", "gives": "time (min)", "from": ["time (h)"], "checked": true} - {"step": "compute", "part": "formula-time-2", "inputs": {"time (h)": 0.54748176}, "output": ["time (min)", 32.8489056], "checked": true} - code: bridges/out/bridges.py - 20 tests, 20 passed, 0 failed $ python bridges/out/test_bridges.py TESTS passed 20 failed 0

Nothing about bridges was installed. It built a table of 800 bridge lengths from Wikidata and a walking-time formula from worked examples, both admitted by their exams. The Øresund Bridge is not in that table, so the decline specialist read it at the source; the made-up Zorvath Bridge stayed declined. The plan read the Golden Gate's 8,981 feet as 2.737 km, found nothing that gives minutes, built that part from its reserve, and answered 32.8 minutes. The whole run took 2 minutes 9 seconds.

Get it

Download the slim release from Hugging Face. A target is a small file: example questions, worked examples, examples kept in reserve, and the work to do (ask, plan, code). Then:

python oracle7.py swarm point target.json
python oracle7.py swarm parts
python oracle7.py swarm paths
python oracle7.py update

update offers a newer release only if it is published and anchored on Solana; update --apply checks every changed part against the release's hashes before writing, keeps a backup, and never touches your own specialists, memory or conversations. The Workbench has the same as a button.

What it cannot do yet

Next is 5.0: new kinds of parts in the same loop (multi-step specialists, a maths solver for whole systems, programs built from programs, pictures checked frame by frame), each with its own sealed test.

Sources

No language model inside. Every answer proved, cited or declined.prometheus7.com/swarm7