games.sgit.ai

A game is the only artefact that makes somebody state a belief before they are told the answer.

Site version: v0.5.0

Every page below has a markdown twin generated from the same content as the HTML page. Fetch the .md and you have the page, without the chrome.

Pages

• games.sgit.ai — a game makes you say what you think, before you are told↗: Why this family builds games: a game is the only artefact that makes somebody state a belief before they are told the answer. The catalogue, the method, and how to build one as a vault.
• The catalogue↗: Every game this family has published, with a maturity rung that has a test behind it, and the vault each one ships in.
• What Can It Do? — the scoreboard↗: The scoreboard game: 40 questions, five levels, scored with a proper scoring rule so that saying yes to everything loses. 17 of the 40 are things no agent can do.
• Which Agent Is It? — the floor plan↗: The floor-plan game: think of an agent, answer cheap questions, chalk which wings of the building you think it can enter — then the doors open.
• Ideas & feedback — the reply channel↗: What players argue with becomes ideas, grouped into themes and answered with a published position — as a graph, joined to the mesh the questions come from, naming nobody.
• The method — four claims↗: Four claims about how to build a game that measures something: score calibration not knowledge, put a ceiling in the set, derive levels from a graph, and treat the mandate as a by-product.
• Calibration, not knowledge↗: Why the scoreboard scores how well you know what you know, with a proper scoring rule that makes saying yes to everything a losing strategy.
• The ceiling↗: 17 of the 40 questions are things no agent can do, because a control outside the environment bounds them. Without them the game could only measure underestimating.
• Levels are distance, not danger↗: A level is the number of hops from a capability to a one-way consequence, computed from the mesh — and the game says on the rail that it is distance, never a danger rating.
• Grant vs. mandate↗: The grant is what the agent can do. The mandate is what you wanted it to do. The delta is the thing nobody writes down — so the game extracts it as a by-product of play.
• The maturity ladder↗: Five rungs, each with a test that moves a game up, so a label on a catalogue card is a claim rather than a mood.
• Building a game as a vault↗: How to build a game that ships as an encrypted vault: the authoring contract, the folder-manifest trap that cost this family four releases, the telemetry lane, and what to publish beside the game.
• What our games send↗: What Can It Do? sends anonymous usage events over two write-only append lanes; Which Agent Is It? sends nothing. What is in an event, what is deliberately absent, what it proves, and how to stop it.
• The network↗: Where games.sgit.ai sits in the sgit.ai family, and which sibling site answers which question.
• Admin & engineering↗: How games.sgit.ai is built and released: one content file, generated HTML and markdown twins, and a seven-check gate.
• Release history↗: Every release of games.sgit.ai. The site version bumps on every push to dev, and CI refuses a push whose commit subject and version.txt disagree.

Release history

• v0.5.0 (2026-09-09) — Two entries in the catalogue. The map vault (mxhepww5): the pack's two matrices drawn live and a proposal form on every row, for a reader without a GitHub account. The Mavs PoC (0833bu5a): the same scoreboard on a draft pack that describes how Mavs AI works, four surfaces with and without Mavs in the path — a sketch, every row derived from mavsai.ai and pending their input. Both read keys in the tripwire's allow-list.
• v0.4.1 (2026-09-09) — The telemetry page names where the disclosure now lives: the what-we-know page became a screen inside What Can It Do? (its vault v1.2.0), and that vault's top bar no longer links to the other game. A copy fix, nothing structural.
• v0.4.0 (2026-09-09) — Two vaults. What Can It Do? now runs out of its own vault (pg87npy3, v1.0.0), reading the public pack from the player site on every load; the home page and the catalogue embed and link to it. The games vault 4evnlwrj keeps Which Agent Is It? and the home page at v0.26.0 with every permission removed and nothing sent, and holds the 9 September version of both games on branch release-2026-09-09. The telemetry page says which vault sends and that browser signals ship off; the floor-plan page says the profile handover did not survive the split, and where the version with it is.
• v0.3.3 (2026-09-09) — The vault-split plan moves to the repository it is about — what-can-it-do.games.sgit.ai, under briefs/ — and leaves here. One copy, in the place a reader of that game would look.
• v0.3.2 (2026-09-09) — A second brief: the plan for splitting the games vault. What Can It Do? moves to a vault of its own that reads the public pack from the player site instead of carrying a copy; the wider vault keeps a frozen branch at the 9 September cut and drops its tracking on main; a third vault lets a reader browse the map and propose a change over an append lane that drains into a pull request; the site learns to load a question pack and nothing else; and a fourth vault is a PoC that explains Mavs AI as a change of control tier on the rows where data leaves toward a model. Written against the vault at v0.23.2, which is eleven releases past what the sites were built from and now sends browser fingerprints the player site says it does not — the plan's first decision is about that. briefs/01__PLAN__vaults-and-packs.md, open for review.
• v0.3.1 (2026-09-09) — Points at the map. The claims each question rests on — what each product can reach, tool by tool, with the evidence behind every row — now live on the player site as a data pack that a pull request can change, with a grants matrix, a page per product and per capability, and eight starting mandates drawn against their profiles as deltas. This site's pages for the game and the catalogue link across; the method pages are unchanged.
• v0.3.0 (2026-09-09) — This site goes back to being the generic one. It points at the games; the material specific to any single game belongs with that game. So the Licence to Operate embed and the long RiskMandate exposition move to what-can-it-do.games.sgit.ai, where a reader who has just played that game actually wants them, and grant vs. mandate keeps the method claim and links across. The telemetry notice also moves off the top of the front page to the foot of it: what it describes is less than a default web-server access log, and a notice a reader has to step over to reach the game treats something ordinary as an obstacle. Also fixes the nav, which prefixed the page's up-path to absolute hrefs and so emitted ../https://… for any off-site menu entry — the footer had always handled that and the nav had not; the link check caught it the moment one was added.
• v0.2.0 (2026-09-09) — These games are part of RiskMandate, and the site now says so and shows where the join is. RiskMandate is the business risk layer for autonomous systems — it begins where security stops, its unit is the mandate (the right to act, granted by a named owner, scoped, time-bound rather than standing), and its signature mechanic is that a real risk has no deny button: the agent already has the access, so it can only be accepted, in a direction, for an interval, by somebody named, and underwritten upward. The games are the front end of one part of that — the bit that gets a real person to state what they wanted. Grant vs. mandate was rewritten around that model and now says plainly what is NOT built: the mandate draft exports into nothing, and the handover is a person retyping. Also embeds Licence to Operate, the published vault holding one agent's grant of 12, mandate of 4 and the 8-capability delta no policy covers, with every reply priced — the best worked example in this family, kept out of the catalogue because it is a simulation rather than a game and blurring that would cost the site its one distinction.
• v0.1.1 (2026-09-09) — The telemetry notice got proportionate. It was an amber warning panel, which was the wrong size for what it says: anonymous counting with no cookies, no analytics script and nothing that identifies anyone is LESS than a default web-server access log and far less than the analytics on almost every site a reader will visit today. Dressing that as a hazard implies a risk that is not there, and a page that over-warns about something ordinary teaches people to skim the next notice. It is now one quiet line in the reading column, still before the thing it describes and still enforced by the release gate. The telemetry page states the comparison outright rather than leaving the reader to guess the scale, and says why the notice exists at all: not because the counting is invasive, but because a vault sending anything departs from a platform promise, and departures get stated.
• v0.1.0 (2026-09-08) — First publish. The argument (a game is the only artefact that makes somebody state a belief before they are told the answer), the catalogue with a five-rung maturity ladder that has a test behind every rung, four method pages, the build guide, the telemetry page, and the network entry. The games are embedded rather than copied: the vault is the source, and assets/vault-app-embed.js opens it over the SG/Vault embed protocol with the vault-browser surface suppressed.