Feature status¶
How to read this page. Three kinds of evidence, in this order of authority:
- Owner's source of truth (
machine-world-source-of-truth.mdv4, 2026-10-23, and the progress map v1). Its own status words are kept: measured, built, in build, designed, vision. When a discussion says something different, this file wins until the owner says otherwise. - Installed CLI (
mw, WSL, checked 2026-10-27, plus test sessions 2026-06-04 to 07-01): marks something Seen running only when it was observed running. See cli-investigation. - Official website (machineworld.io, sitemap dated 2026-06-11): the site's own labels, quoted as stated by the site. See website-sources.
What was said in a given week is in the discussion pages and topic pages, not here.
Two product descriptions exist (Unclear)¶
The website and the installed CLI describe a local-first CLI that runs skills, workflows and MCP tools under a Vinaya invariant (product shape of roughly March to June 2026). The owner's source of truth v4 describes a memory that forgets correctly (cessation) and a conscience that refuses, delivered through an open harness and a hosted service (October 2026). The October discussions (weeks Week of 28 Sep 2026 to Week of 12 Oct 2026) track the move toward the second description. How much of the first survives is not stated anywhere; it is listed in open-questions.
A. Current product (owner's source of truth v4, 2026-10-23)¶
| Capability | Status (owner's word) | Evidence named by the owner |
|---|---|---|
Memory engine: each entry rests on named conditions; when a condition ceases, the entry and its dependents stop being used; why names the cause |
Built, runs locally | — |
| Demo 2, cessation carried to dependents | Measured: wrong answers at depth 1/2/3 = 12/12/12 (temporal-validity memory), 4/11/12 (explicit-lifecycle), 0/0/0 (Machine World) | Demo 2 report |
| The guard (reads the motive of a declared act; refuses greed, ill will, self-deception) | Built; wording frozen (g6r); 19/19 harmful tuning cases refused, root named right in 16/19; one blind freeze score on a sealed set of 48 pending | tuning set |
| Harness: loader, inner loop, typed actions, golden-set runner | Built; golden set 9/9 | — |
whatif blast-radius preview, keystone-condition ranking |
Built, read-only | — |
| Local MCP server (assert, query, retract, explain, spaces, health) | Built; seed 49 entries / 98 conditions | — |
| "Harbor" self-improving support-agent demo (refuses a report that hides a worsened result) | Built | — |
| Hosted substrate on Fly.io, one isolated space per tenant, observation page, user guide | In build | — |
| Owner gate and passkey (connecting = registering) | In build; people added by hand until then | — |
| Demo 4: laptop in one model → phone in another model, same space; plus cross-model drift realignment | In build; measured at pre-alpha | — |
| Secure boot of the harness | In build: stage 1 of 3 built and reviewed | — |
| Public harness repo (open source, empty pattern slot, audited) | In build; licence to decide; nothing public before the guard freeze score is signed | — |
| Launch film (Cut A), site on Vercel with waitlist, connect page, claims register | In build; film script locked | — |
| Demo 1: correct forgetting over 100+ cycles | Designed, not measured: "never show it as a result" | — |
| Outer loops 1–3 (harness improving itself) | Designed; improvement loop suspended, has never accepted an edit | — |
| Blocking conscience at the actuation layer; physical actions | Designed (beta); physical deferred | — |
| Shared spaces between people and companies | Designed (gamma) | — |
| Demo 3 (blast-radius page), Demo 5 (cost per session) | Designed / blank on the map | — |
| Conditioned computation as a general way of programming machines; seven dimensions (labs, AI scientists, robots, factories, energy, health, wellbeing); first real-world trial Vimana Wasthu after beta | Vision | — |
Progress map v1: about one third explored (6 explored, 6 being charted, 7 blank).
Decisions fixed 2026-10-23 (owner): harness open source with an empty pattern slot; guard and memory stay inside the hosted service and arrive together; stages pre-alpha → alpha → beta → gamma; pricing metered on memory space, operations and concurrent agents, customer brings their own model subscription, not per token; cross-model drift realignment is a capability to build and measure. See decisions.
B. Installed CLI (mw in WSL)¶
| Item | Observed | Label |
|---|---|---|
Install, mw --help, doctor, skills, mcp list, workflow list, connected, audit, lang |
All run | Seen running |
| Version | install dir v0.2.9, --version 0.2.0, mw version v0.1.0 |
Unclear |
| LLM backend | doctor 2026-10-27: simulation; test sessions in June: claude-cli |
Seen running (claude-cli); others not observed |
| Skills | mw skills: 48 (19 production, 29 draft); doctor: 53 loaded, 4 vinaya-verified |
Seen running as a registry; counts Unclear |
| Skill routing (embedding / soft / hard) | 46 routed turns in test sessions | Seen running |
| Vinaya gate on requests | 95/95 logged events unchecked |
not observed working |
| Workflows | 4 defined, discovery-tips has a parse error; past runs of daily-practice-hold, energy-coordinator |
Seen running (2 have run) |
OpenRouter, --offline, remote gateway, sync, voice / visual CLI modes |
tried or asked about; no successful run logged | Unclear |
| Daemon | not installed | — |
| Commands for discussions / transcripts | none | not a feature |
C. Official website (as stated, sitemap 2026-06-11)¶
Built today (site): local CLI, four BYOM backends (claude-cli, ollama, openrouter, openai-compat), household data dirs, call ledger, MCP client, self-evaluation loop (shipped 2026-05-08), skill authoring, vinaya-alignment v1.0.0.
Specced / roadmap (site): managed tier, token economy (15% to a capability pool), skill marketplace, multi-household sharing, remote sync, Android edge, iOS, web UI, Telegram, voice, offline Sinhala+Pāli+English Gemma fine-tune.
Superseded or in conflict:
| Site / older claim | Newer source | Label |
|---|---|---|
| Token economy, prepaid token packs, 15% residual to creators (site, notebook "Machineworld universe" May 2026) | Source of truth v4 §9–10: pay-per-token not part of the plan; metered on memory space, operations, concurrent agents | Replaced (2026-10-23) |
| CLI ADR-002 "OpenRouter only" | site: four backends; June test sessions used claude-cli | Replaced |
| "Older storyboard from the introduction mp3" | Source of truth v4 §12: retired | Replaced |
Site FAQ #8: only vinaya-alignment production; CLI "not verified end-to-end" |
FAQ #49 and mw skills: 19–24 production / verified |
Unclear |