Glossary

Glossary

Plain definitions first, then the precise one where it differs. Terms are grouped so related ideas sit together.

The four words

Identity. Who is acting: a person, an organization, a system or an agent, as a record the platform knows, not a name typed into a request. On the platform every credential is bound to one. In the language it is the ID[...] clause.

Boundary. What an actor is allowed to do, declared before the work and enforced during it. Platform: scopes, purpose profiles and policies. Runtime: the grant lattice on the actor's graph.

Intent. The declared reason for an action, stored with it. Platform: the purpose on a projection and the outcome on an interaction. Runtime: the Intent[...] clause.

Trace. The record of what happened, written as it happens into a structure that detects edits. Platform: the per-tenant audit chain. Runtime: the journal.

Authority

Granted. Knowledge that entered the graph through a grant signed by a registered grantor. It is authority.

Observed: governing. Knowledge attested by a deterministic tool, anchored and replayable. A candidate for promotion by a human; not authority yet.

Observed: proposed. Something a model said. Never authority on its own.

Seed, proposed, certified. The platform's status ladder on facts, evidence and relationships. Only certified edges appear in projections.

Attested, proposed. The status ladder on obligations. An attested obligation was confirmed by a person; a proposed one is a suggestion and never a due date.

Weakest link. When a trace mixes layers, the lowest layer present governs the whole trace.

Projections

Purpose profile. A named reason for asking, with the node types it permits and a sensitivity ceiling. Four ship with every tenant: identity-introduction, directory-lookup, scientific-research, regulatory-filing.

Projection. The slice of an identity returned for a purpose, after the four gates, with a receipt.

The four gates. Purpose, sensitivity, consent, jurisdiction, tested in that order; the first failure is the reason.

Sensitivity. public, internal, confidential, restricted, on every node, with defaults by type.

Consent record. The stored fact that this identity may be projected under this purpose, by whom it was granted, for what scope and until when.

Transfer policy. Per tenant, which sensitivities may be served to callers in which jurisdictions.

Explain. The projection endpoint that also returns every node considered and the result of each gate.

Withheld. The count, by node type, of what a projection did not return.

Node budget. The maximum number of nodes a projection returns (default 50, at most 500). Beyond it the response is truncated.

Records and proof

Governance receipt. A signed statement of who asked, what was decided, which policies and evidence were involved, and a hash of what was returned. Verified with the public key at auth.id-me.ai.

Audit chain. The per-tenant sequence of events where each entry's hash includes the previous entry's hash. The database refuses to update or delete entries.

Interaction. What an assistant did with a projection: completed, refused, proceeded on truncated or empty context, or escalated. Recorded into the audit chain.

Journal. The runtime's hash chained, file backed trace. Refuses to load if altered.

Merkle inclusion proof. Proof that one journal entry belongs to a journal with a published root, without revealing the rest.

Runtime digest. The SHA-256 of the runtime's own source manifest, carried inside every runtime receipt, so the build that produced a receipt is known.

Tombstone. What remains when a learned record is forgotten: its id, kind, the time, and a SHA-256 of the record as it was. The content is gone; the act is provable. Append-only.

Learning state. Whether new context, facts and relationships may be recorded about an identity. Paused means refused with 409; what is known is still served.

Fingerprint. The SHA-256 over the canonical form of an identity's whole exported record. Any change to the record changes it; the export says how to recompute it.

Lineage. The list of audit chain events about an identity, each with its sequence and hash, carried in the export.

Attestation. A signature over a statement by a key. Classed as ai (no signature), service (the service's key) or human (a registered person's key over this exact statement).

Credentials and access

Tenant. One customer's isolated space: identities, purposes, policies, keys, audit chain. Every request names it.

API key. igl_live_ or igl_test_, shown once, stored hashed, bound to a tenant and an identity, carrying scopes.

Scopes. igl.read, igl.write, igl.execute, igl.admin, and igl.platform for the single key that administers tenants.

Token. A short-lived EdDSA JWT obtained from the key through the OAuth 2.0 client credentials grant, usually with narrower scope.

Platform key. The one igl.platform credential, created by the bootstrap call, used to issue customer tenants and keys. Rotated by revoking it and bootstrapping again.

Idempotency key. A header on mutations that makes retries safe: same key and body replay, same key and different body conflict.

Doing things

Policy. An effect (allow, deny, conditional, escalate) and a rule over actions, resource types, conditions and scope. Evaluation fails closed; deny wins.

Agent. A registered capability with a declared purpose and scopes, invoked to produce runs.

Run. One execution of an agent or workflow, with its policy decisions, receipts, state and outcome.

Workflow. A versioned sequence of steps (query, research, agent, policy check, tool, transform, outcome, human approval), each pointing to its successor on success and on failure.

Outcome. What happened after a decision, recorded with structured values and later measurements.

Webhook. An HTTPS endpoint a tenant registers for graphUpdated, outcomeRecorded and approvalRequired, with HMAC signed deliveries.

The language

Statement. One governed unit of work: identity and boundary, intent, context, compute, output, and what to do on failure.

UDM, AI, IOS. The three compute namespaces: deterministic resolution, model inference with a pinned model, and the trace and attestation subsystem. They are never confused.

Pinned model. An inference step must name its model and parameters. No inference is asserted without one.

Two-phase commit. Nothing material is produced until the trace is durable.

Confidence decay and depth cap. Unanchored inference loses force each time it is reused and may travel only so many derivations deep. These make the recursive loop convergent.

BOUNDARY graph and FOOTPRINT graph. A company's prescriptive graph (what is granted) and an individual's descriptive graph (what has been granted and what has merely been observed).

Decode bridge. The part of the runtime that compiles the granted boundary into a per-token mask, so a model can only emit values inside the boundary.

MCP. Model Context Protocol, the standard by which AI assistants call tools. IGL serves its operations as MCP tools on both the hosted platform and the local service.