← Prometheus7 · Oracle7Roadmap · updated 2026-10-28

The road: 6.0 to 20.0

What Oracle7 lets you do, version by version, from 6.0 to 20.0. Green is built and proved by its own sealed test; a hollow red ring failed its test and is published that way. Each future version says what you can do with it, how it will be proved, and whether that proof is one we can run ourselves or one that needs other people. Two roadmaps, build and vision, are below the spiral. Answers stay free at every version.

Reach Index · release 5.5 1,755,340 checked chains from 223 parts, counted two independent ways ×4 parts → ×89 distinct functions · every version below publishes its own count

So far: 3.2 to 10.1

4.0The Law: every answer spelled out step by step. Released 20 Oct 2026.
4.5–4.9Reach and the swarm: it builds, examines and combines its own parts. 4.9 released 23 Oct.
5.0Organs: steps, maths, code and film, each checking its own work. Passed its sealed test.
5.1Reading: questions in plain words, answered only when every reading agrees. Passed its sealed test.
5.5Represented compute: capability counted exactly, 2,000 of 2,000 sampled chains re-verified. Passed its sealed test.
6.0Growth: 615 parts built unattended from Wikidata's formulas; right answers 2 → 168, none wrong. Passed.
6.5Many places: parts move between machines and every check runs at the receiver; 0 of 80 tampered parts admitted. Passed.
7.0The Slow Fire rules in software, on simulated places. Passed. The vine runs them live in simulation until its sensors arrive.
7.5Teacher: answers opened into lessons whose every step is checked. Passed.
8.0Calibrated forecasts of its own growth. Failed twice; published, not overridden.
8.5Discovery: hypotheses registered before the data, reproduced a second way. Passed.
9.0Reading: 731 right and 11 wrong on 2,000 unseen questions, 1.77× the 5.1 reader. Passed.
9.5Continuity: every past answer replayed exactly as of its date; tampering detected. Passed.
10.0One law: answers proved, cited or declined, re-checked by a tool that ships with it; 313 right, 3 wrong, 0 of 130 false answers. Passed.
10.1Clauses read and joined, organs asked together, handoff and check for other agents, ember arlynx's fixes; 493 right and 9 wrong on 1,200 fresh questions (1.13× 10.0). Passed v37e. Released 28 Oct.

Now: 10.1 is released (28 October), anchored on Solana and on Hugging Face. Next is 10.2: coverage that grows with what it knows, the same checks as tools for any agent harness, and every module under the anchored root. Scores, with every wrong answer subtracted, are on the scorecard.

done, passedfailed, publishednextmajor, plannedhalf, planned

Two roadmaps

The road above mixes two kinds of proof. Some versions are proved by sealed tests we write, seal and run ourselves. Others can only be proved by other people: by use, outside examiners, contributors, communities and certifiers. So the road is now two roadmaps.

About the dates. They are estimates at the pace of the last nine days, halved, because most of the next work extends what already exists. They are not promises:

The road from 3.2 to 10.0 was written on 19 October and read like years of work. It took nine days: sixteen versions, two of them failing first.
Building may go faster than estimated, or slower where the domain is new.
The vision will take longer than the build, and its pace is set by more than us.

Every version ships only when its own sealed test passes; a failure is published and the date moves.

The build roadmap: what we can prove ourselves

Version	Adds	Proved by	Estimate
10.2	Coverage from verified parts, agent tools (MCP), provenance under the anchored root; slim and full	Sealed test v38: right minus wrong positive on what it adds, precision ≥ 0.95	by 29 Oct
10.5 + 11.0	The examiner kit; a nightly sleep cycle that replays the day, turns failures into examined parts and publishes a morning report; the part format	Our own failures admitted as parts at unchanged precision; a sealed week of nights with gains and no new wrong answers	~30 Oct
11.5	Derivations checked by Lean	None claimed proved that Lean rejects	~1 Nov
12.0	Models that run over time	Backtests on held-out real series; intervals hold at their stated rate	~2 Nov
12.5	Receipts	Re-verified offline from the receipt alone	~3 Nov
13.0	The laboratory	Findings registered first, reproduced a second way	~4 Nov
13.5	The vine on real sensors	v33's bars held for a season	software ~5 Nov; proof ~Feb 2027
14.0	Engineering designs	Independent open-source simulators	~7 Nov
15.0	Self-sustaining growth	Cost per verified ability falling across three releases	mid Nov
15.5	Off-grid	The cumulative test on a small board, offline	~9 Nov
16.5	Whole courses	Every step checked across a published syllabus	~mid Nov
17.5 – 19.5	Private libraries, the long ledger, check anything, everywhere	Each by its own sealed test	through end Nov

The vision roadmap: the Slow Fire

Oracle7 began as a way of answering questions that keeps its receipts. The vision, set out in The Slow Fire, is where that discipline goes. Its horizon is ten thousand years, and its model is the mushroom: many small nodes rather than one brain, growth by connection, no centre to capture.

Proof on infrastructure, where its virtues are obvious and its limits do no harm.
The first real plant. The vine runs the Slow Fire rules live in simulation now, and moves to real sensors next.
Waking the world. Sensors for soil, water, air and trees. The sensors are sold; the intelligence is held in common.
Everyone builds. A teacher, a parent, a lab, each with a model of their own, on their own hardware.
A web with no centre. Facts travel with their receipts, so no one has to trust the node they came from.
The oak on the corner. It greets you before it tells you anything, says only what has been measured, and can choose not to be read.
Advocates for watersheds, reefs and forests, speaking in planning meetings and courts with the same discipline.
A protocol that outlasts its products. Bind, unbind and superpose can be rebuilt from a few pages.
Beyond the planet. Wake dead worlds, listen to living ones.

Each of these needs people, and they carry no dates. The test for every step: could the gift be taken back? If anyone, Prometheus7 included, could switch the network off or turn it against the people using it, it has been built wrong.

What holds at every version

Answers stay free, and it runs on your own machine.
No language model, Transformer, attention block or gradient-trained component, at any version.
Every answer is proved, cited or declined. A wrong answer costs more than a decline.
Capability is counted: every release publishes its Reach Index, measured, never estimated.
Every release passes a sealed test written before the build and run once; failures stay failed and are published.
Your data stays yours. Places and people can say no, and the record keeps the refusal.