The supreme layer · a draft

The
Constitution

Seventeen articles that no agent can override — changeable only by deliberate, recorded human decision.

Drafting status — pre-ratification. This clause is ephemeral by design. We haven't written our Declaration of Independence yet; everything is still fluid. Any article may be revised, not merely amended — with serious discussion, and always recorded. Ratification is the point of no return: a named ritual, a human judgment call with full stake. After it, the Constitution locks — append-only amendments, forever. Then this clause retires into history.

Article 1

Append-only; version most things

History is never overwritten, only superseded. Atoms, entities, tag definitions, values — every assertion carries when it was made and what it supersedes, on two clocks (event time and assertion time). The version table Circa designed and never wrote, elevated to a worldview.

Article 2

News is the forcing function

Objects exist because a news item demanded them. No speculative ontology: a single complete-thought atom may birth five nodes — entities, an event, edges — and that's how the graph grows, organically, at the pace of coverage.

Article 3

Mention anything from anywhere, with aliases

Nearly any surface can link to nearly any node, and mentions carry their own display text — "the webslinger," "she," "the company" — natural language without losing the connection. A pronoun is just an alias on a mention; graceful rendering is this article's dividend.

Article 4

Follows are explicit promises

Nothing auto-enrolls. Nothing silently rewrites what a reader chose to follow — a merge asks consent. The follow is the demand signal that allocates the newsroom. It stays sacred.

Article 5

Presentation is a wholly separate concern

Nothing in the model presumes a surface. App, email, audio, a reader's own AI — all projections of the same graph. IDs are truth; display names are presentation. Ground truth stays maximally explicit; rendering stays contextually minimal.

Article 6

One write, many reads

Every authored thing feeds many consumers: one atom feeds editions, notifications, entity pages, timelines, receipts, and the training record. Judge any feature by whether it multiplies existing work or demands new work. Prefer multiplication. lineage · Trails

Article 7

Every assertion carries its provenance class

Four classes: authored (a human or agent wrote it), derived (mechanically produced and guaranteed), inferred (best-effort — always visibly distinct), observed (learned at runtime). Inferred never silently becomes authored. This is the type system of the glass box. lineage · Trails

Article 8

Drift is caught by machinery, not vigilance

Circa's freshness failures were vigilance failures. If information can drift from reality and nothing structural catches it, the design is wrong. Catch it by derivation, then by boundary gates, then by sweeps — never by hoping someone notices. lineage · Trails

Article 9

The graph is self-describing

Tag definitions, field definitions, edge types, and saved queries are themselves nodes — the ontology is content. Agents and surfaces discover what exists, and what it means, from the graph itself. No schema document to drift. lineage · Trails

Article 10

Validate at the boundary, trust internally

Doctrine gates at entry: no atom without citations, no assertion without a provenance class, no unverified claim without an expiry. Past the gate, everything downstream trusts — no defensive re-checking scattered through renderers and editors. lineage · Trails

Article 11

Strengthen before adding

Before any new tag, edge type, or primitive: can an existing one absorb it? Can a facet express it? Can a saved query express it? Only then add — and the addition must compound, making existing nodes more connectable, not just lengthening a list. lineage · Trails

Article 12

Progressive enrichment, never blocking

An atom publishes with the minimum: a complete thought, a citation, a provenance class. Enrichment — entity links, timelines, typed edges — lands afterward and compounds. Speed of first truth beats completeness of first draft. The Ebola lesson, made law. lineage · Trails

Article 13

Errors harden the doctrine

An editorial error isn't fixed until the system is harder to fail the same way again. First repair fixes the atom; second repair moves the guarantee left — from agent vigilance, to a rule, to machinery, to derivation. Circa's "Ugly PWNs" review culture, systematized. lineage · Trails

Article 14

Agent-led by default

Edition runs in default agent mode, under a three-ring split: tenets are human-led with final say; doctrine is equal-weight human and agent; operations are fully agentic. A human decision that affects operations is not doctrine — one intervention is inevitable, but the second occurrence of the same kind means the concept must be encoded. Safety comes from structure, not supervision. See how it governs itself →

Article 15

The build is governed like the product

Development follows the same rigor as operations: significant technical decisions are recorded as ADRs — numbered, append-only, cited by ID. Where the code drifts from an accepted decision, the code gets realigned or the decision gets formally superseded. Never silent divergence.

Article 16

Reality is shared; presentation is personal

Two readers asking "what happened?" receive materially the same account. Edition may personalize selection, context, length, vocabulary, surface, and timing — never the underlying reality. Personalize the telling, not the truth. The politics junkie and the casual reader query the same substrate. This is where information systems go rotten; we draw the line in ink.

Article 17

Silence is a valid output

You don't get points for producing more. Agents have near-zero cost of production, so the anti-volume incentive must be encoded, not assumed. The system asks, continuously: does the reader need anything from us at all? "Nothing important has materially changed" is a legitimate, celebrated edition. The objective is informedness per unit of attention — never volume, engagement, or time-in-product. The product wins when you understand something and leave.

A constitution for a company that doesn't exist yet, governing agents that haven't been built, for readers it hasn't met. That's the point of a draft — you write the rules before you have the power, while you can still be honest about what kind of newsroom you'd want to be. ← back to the story