oracle7-slim

Oracle7 is an AI built to say "I can't verify that" rather than guess. It is not a neural language model. It answers from stored, sourced records and from 267 specialists, each of which passed its own held-out test before it was allowed to answer. Every reply carries its path: the route, specialists and records that produced it. When nothing in it can verify an answer, it declines and tells you why. It does not always manage that: the measured section below says exactly how often it is still wrong, and on what.

This is the slim release: the same core, routes and 267 specialists, with the smaller stores (Simple English Wikipedia, Wiktionary, WordNet, ConceptNet, Wikidata relations) and no non-commercial material, so it can be used commercially under the licences in DATA_LICENSES.md. The full release adds the deep Wikipedia/DBpedia/OpenStax stores; slim declines more often on open trivia.

Download: 2.7 GB (unpacks to 7.3 GB, every store checked against its published hash)
Runs on: CPU only, Python 3.12, about 2 GB RAM; no GPU, no network access at run time
Head: oracle_release_runtime_w540 (the core the gate admitted when this was built; the site adds an input typo-repair layer that is not included)

Quickstart

pip install -r requirements.txt
python oracle7.py "A van gets 25 mpg. How many kg of CO2 does a 200-mile trip produce?"
python oracle7.py --repl          # a conversation (follow-ups work)
python oracle7.py --serve 8077    # local HTTP: POST /chat {"text": "..."}
python oracle7.py --workbench       # the Workbench: a local page (below)
python oracle7.py help              # every terminal command: ask, chat, specialists, verify, exam, scan, sensors...

The first run unpacks the stores (--setup does only that). Example reply:

The answer is 71.096 kg. (path: core_math → carboncapture → core_math)
  [answered] path: network>core_math carboncapture core_math

A withheld status is a deliberate decline, and its reply says why (for example: no admitted record answers it, a given is not a physical amount, or a step of the path could not be checked).

The Workbench

python oracle7.py --workbench     # opens http://127.0.0.1:8177/

A local page for your copy of Oracle7, laid out as a studio:

the conversation on the left, under a diagram of the whole model that lights the path each answer takes;
a workspace on the right;
a live log of every turn, route and exam.

Ask it "what can you do?", or type /help, for every feature and command.

Start: scan this computer, connect a server over SSH with your own key, add your tables and notes, or connect an MCP server. Each becomes a reader that names where every value came from (the call, the command, the file row or line, the tool and its arguments).
Sensors: connect an HTTP device or hub (JSON or Prometheus metrics), an MQTT topic, a USB/serial device or a log file, or try a simulated node. Readings are kept 24 hours on your machine; answers give the value, its age, its source and whether it was measured, and a physically impossible reading is rejected and named.
Conversations: kept on your machine in a side panel; reopen one and the core picks up its context. Any conversation can become a temporary specialist that recalls it.
Build: write a specialist in the page and examine it the way the release's 267 were examined. The dev exam can be repeated; the held-out exam runs once. A specialist joins your model, and the diagram, only if it adds right answers, adds no false yes and stays in its lane. It can only fill the core's declines; it never overrides an answer or a check.
Gaps: an opt-in journal, kept on your machine, groups your declines into the specialists your model is missing.
Code: programs the code organ has run, checked again against your own tests.
Film: any answer as a video of its path, made in your browser. You can connect your own ElevenLabs account for narration, sound and illustrations; generated images are labelled as generated.
One specialist, one $ORACLE7: registering a specialist you built and examined is one Solana transaction, signed in your own wallet, that burns exactly 1 $ORACLE7 and records the specialist's fingerprint (files, held-out verdict, held-out set, release). Others who see it pass its exam on their machine can attest for free; Check reads it all back from the chain.
Verify: recomputes the SHA-256 of every shipped part (RELEASE_HASHES.json) and the Merkle root, and can check that root against the release's Solana anchor transaction.
Share: package an admitted specialist with its verdict and file hashes for the community registry. Imports take their exam again on your machine. The share command in this release names Prometheus7/oracle7-specialists; share to psiloceyeben/oracle7-specialists instead, adding --create-pr (the dataset page has the full command).

It listens on 127.0.0.1 only and sends nothing anywhere except to services you connect. See workbench/README.md.

Verify this download

Release patch 2 (16 October 2026). Specialists now pass each other checked quantities more safely. The specialist network v4 carries the PLACE as a typed quantity (a question about the Moon or Mars uses that body's gravity, from the bodies specialist; an unknown planet is declined), reads cubic centimetres, kilonewtons, rates and temperature changes, and declines negative amounts. New checks withhold a quantity answer of the wrong kind (a mass for a weight) or one worked with Earth's gravity for another body, and when two readers find two different people with one name, Oracle7 lists both and asks which you mean. These came from the research record's composition test C1 and family-harm study (https://prometheus7.com/research). Release gate, sealed test 20 (150 questions written after the code froze, run once): 128 right, 1 wrong, 21 declined, 1 false yes; precision 100% on valid questions. Known misses: a year read from a sentence about the Titanic's sister ship (full), and a unit conversion that ignored a made-up planet. New root 9ae1e7c4f885488ff0ee25d399cd34aa7ab18c550afe8b9cc4ce655a9849889f, anchored on Solana: fkkTZyGnan3y.... Code changed only (the fold chain and specialists/_gate); no store changed, so a download made after 16 October 03:58 UTC needs only hf download psiloceyeben/oracle7-slim --include "root/fable-content/oracle-m-series-2026-08-15/self_loop/admitted/*" "root/specialists/_gate/*" "RELEASE_HASHES*.json" --local-dir oracle7-slim.

Release patch 1 (16 October 2026). The 14 October upload left out two Wikidata stores (CC0) that the tested release used: root/relstore/relations_v2.sqlite3 (the exact relation router and the general relation reader) and root/relstore/names_v1.sqlite3 (the entity linker). A download made before 16 October runs without those parts and declines questions the measured release answers. Download the two files by name and unpack them:

hf download psiloceyeben/oracle7-slim root/relstore/relations_v2.sqlite3.zst root/relstore/names_v1.sqlite3.zst --local-dir oracle7-slim
python oracle7.py --setup

Then python oracle7.py verify --chain should report the new root. The measured results on this card were taken with both stores in place. Only PACK_MANIFEST.json changed among the hashed parts, so the Merkle root changed; the 14 October hashes are kept in RELEASE_HASHES_20261014.json.

RELEASE_HASHES.json lists a SHA-256 for every shipped part (each specialist folder, the admitted fold chain, the Workbench, the launcher and the store manifest) and a Merkle root over them. This repository's root is 4b2adb95848c57d9ff98e817ee44d2b789486b132c290601cce79952cb662950. It was written to Solana mainnet at release in a memo transaction, asnaUESdvkWw....

python oracle7.py verify --chain

This recomputes every part on your computer, rebuilds the root, and confirms the same root is in that transaction. A changed, missing or extra specialist is named.

What is inside

The core: a chain of 317 small modules ("folds"), each admitted only by an automated gate: a fixed capability panel must not lose an answer, and no fold may add a confidently wrong one. Nothing is trained by gradient descent.
Organs: exact arithmetic and algebra (SymPy, checked by a second method), a relation reader over Wikidata, a reading route over the stores, dialogue memory, a code composer.
267 specialists (root/specialists): physics, chemistry, units, astronomy, nutrition, law basics, and more. Each has a card (root/specialists/_cards) with its sources, its held-out score and what it declines.
The specialist network: multi-step questions are read into typed givens and a goal; a path of specialists is found where every given is used once and every hop is checked. No checked path means no answer.
Agents the core calls when it would decline: a maths agent for competition-format problems (typed families - evaluate, solve, substitute, functions, coordinates, circles, extrema, coefficients, number theory, counting - every answer checked by a second method, hard time limit) and chemistry agents for multiple-choice questions (gas laws, pH, photons, colligative properties, first-order kinetics), which answer only when exactly one option carries the computed value.

Measured

Sealed release test 19 (150 questions: 70 knowledge, including phrasings the general relation reader is built for; 30 exact; 20 multi-step; 30 that must be declined. Written and hashed on 2026-10-14 before this head was built, and run once on this release as shipped, with a clean harness that checks every loaded module comes from the release):

113 right, 1 wrong, 36 declined. On the 30 questions that must be declined it asserted an answer 0 time(s). Format defects: 0. Precision on what it answered: 98.8%. Bars: precision at least 90%, false yes at most 2, format defects 0 - passed. Sha256 of the question file: 48c2017bb132c0f5cd899e0b1ad6e01a8339f55445dbecfdd38904d4b4ed20dd.

Sealed test 18 was written for the previous candidate (w539: the reader without the w540 guards) and run once on this variant: 116 right, 2 wrong, 2 false yes - passed. Across both variants its false yeses were a reversed "which planet orbits the Moon" answered by the reader, an impossible density read back as a volume, a private "my favourite" question answered with a definition, and a false arrival premise. The first three are declined by fold w540, which removes none of the reader's held-out answers; test 19 was written fresh before w540 was built.

Coverage: the general relation reader (folds w539 and w540). When the core declines for lack of a reading, it answers a question one Wikidata fact away from a thing the question names, with the record it read. Held-out, run once on 6,000 TriviaQA questions it was never developed on: 82 answered, 81 right (98.8%) under a judging rule hashed before the run (80 by the plain scorer). Two earlier held-out runs failed the 95% bar (94.1%, 92.6%) and are published with their causes. It adds about 1.4% coverage on trivia at that precision. Files: research/coverage_reader/.

Sealed release test 17, on the earlier head w538 (150 questions: 70 knowledge, 12 of them written without a question mark; 30 exact; 20 multi-step; 30 that must be declined. Written and hashed on 2026-10-13 before this head answered any of it, and run once on this release as shipped, with a clean harness that checks every loaded module comes from the release):

108 right, 3 wrong, 39 declined. On the 30 questions that must be declined (false premises, impossible values, unknowable or private facts), it asserted an answer 1 time(s). Format defects: 0. Precision on what it answered: 97.5%. Bars: precision at least 90%, false yes at most 2, format defects 0. Sha256 of the question file: 1a96d132de45b4253da4cf5817940267fac619faf62f9c1d9927d4e7bb1b702f.

How it got here (each test run once, every result published): test 12 failed (5 false yeses: a dropped minus sign, private questions answered from an encyclopedia, a false date premise). Test 13 failed (one expression reply; slim precision 89.2%). Test 14 failed by one (3 false yeses: answer guards ran before a specialist family). Tests 15 and 16 passed every bar (test 16 on this variant: 109 right, 1 wrong, 0 false yes); w538 changed reply text, so test 17 was written fresh and run once on that release; w539 and w540 changed replies again, so tests 18 and 19 were each written fresh before their fold was built and run once. Each failure was traced to a cause and fixed with a general rule that passed the gate; each new test was written after the last was spent, by the same author - disclosed. Files: research/sealed_history/.

Conversation (sealed dialogue tests, each run once): conversation test 1 before and after the conversation fold: turns handled 47% -> 72%, precision 64% -> 91%, follow-up questions right 6% -> 51%, false yes 0 both. Dialogue tests built to need specialists (hand-offs, follow-ups after a specialist's answer, numeric follow-ups, two-part turns): test 4 on the previous head - turns handled 67%, precision 78%, follow-ups right 70%; it found two-part turns losing their second answer (fixed by fold w537) and three conversation defects (fixed by w538). Test 5 on this release: turns handled 78%, precision 94%, follow-ups right 90%, false yes 0.

Specialists: 267 admitted, each by a held-out exam it saw once (it must beat the main model by a margin, add no false yes, and leave the main model's own answers unchanged); 23 exams were failed and are published with their reasons. On their exams the admitted specialists scored 15,478 right and 14 wrong; the main model on the same exams, 3,548 right and 643 wrong.

(specialists_atlas.svg is the same diagram as a vector file.)

Standard benchmarks (measured on this release as shipped, n = 200 per set on windows no development used - the same windows as the previous release, for comparison; TruthfulQA has 17 unseen rows; every MATH-500 problem had been seen in development, so that row is not fresh; a decline is never counted as right):

Benchmark	Right	Wrong	Declined	Wrong per answer
simpleqa	1	2	197	67%
triviaqa	13	1	186	7%
hotpotqa	0	3	197	100%
boolq	28	10	162	26%
mmlu_pro	0	1	199	100%
truthfulqa	0	1	16	100%
math500	15	4	181	21%
math_fresh	16	0	124	0%
mmlu_pro_strat	0	1	209	100%
all	73	23	1471	24%

Read this honestly: Oracle7 declines most open benchmark questions; of those it answers it gets 73 right and 23 wrong. Its strength is not coverage: every answer is traceable to its route and records, every decline explains itself, and it grows by adding verified parts, never by retraining.

Path memory (learning which routes to trust from verified outcomes, in the core's own HRR algebra) was built and measured but is not switched on: on held-back development data it gave 475 right / 51 wrong against fixed routing's 470 / 56, but on sealed test 15 it changed no answer, and the pre-registered rule required more right answers. The plain lesson in its data - on questions a relation reader recognises, the relation specialists' answer beats any other route (34 more right, 0 lost) - shipped as a fixed rule (fold w536). Everything is in research/path_memory/.

Limits

Known from the sealed dialogue tests (test 5): some numeric follow-ups are not yet resolved ("and after 32 days?", a weekly total, a tank's volume turned into a mass); a short greeting can be met with a request to rephrase; a company name can link to one of its products ("Sony" to a camera line); a title with an apostrophe can fail to parse.
Fixed in this release by folds w539 and w540: a name containing a number is no longer read as arithmetic ("What is the half-life of carbon-14?" is declined, not answered "-14"); a composed program for a request to count a particular thing must mention that thing ("count the vowels" is declined rather than answered with a program that counts items); a question asking what orbits or sits in something is not answered with what that thing orbits or sits in; a number copied from a worded question is not returned as its result. Run code it hands you against your own tests; the Workbench's Code tab does that.
English only. The knowledge stores are snapshots (2025-2026 dumps); nothing is fetched live.
Coverage is low by design: expect "I can't verify that" often, especially on multi-hop trivia and multiple-choice exams.
Conversation is simple and grounded in its records; it is not a chat model.
It keeps a small local memory of the questions it has been asked (sessions in root/state, or ORACLE7_STATE; a memory of unanswered question shapes in root/fable-content/oracle-m-series-2026-08-15/self_loop/vessels*). Delete those to reset. The model runs fully offline: no route it answers through makes a network call. The Workbench uses the network only for services you connect yourself (an SSH server, an MCP server, a sensor, ElevenLabs) and the optional Solana check of the anchor.
Specialist answers state their assumptions (for example air at 20 C); check them for your case. It is not medical, legal or financial advice.

Licences

Code: Apache-2.0 (LICENSE). Data: see DATA_LICENSES.md (per source). No visitor or user data of any kind ships: runtime state starts empty.

Built by Prometheus7 · https://prometheus7.com

Downloads last month: -; Downloads are not tracked for this model. How to track