Skip to content

IGL 1.6.0

Latest

Choose a tag to compare

@cosudm cosudm released this 16 Oct 13:40
· 41 commits to main since this release
Immutable release. Only release title and notes can be modified.
23605ab

Contract 0.3.0. The Subject service: the controls a person holds over what
is learned about them, the rules Ricorsa is built on, as platform operations.

Added

  • GET and POST /identities/{id}/learning: pause and resume learning about
    an identity. While paused, context, facts and relationships that would add
    to its record are refused with 409 learning_paused; what is known is still
    served under its purposes
  • POST /identities/{id}/forget: each target (fact, relationship or context
    entry) is emptied in place and marked forgotten, leaving every projection,
    search, path, list and intelligence context at once. A tombstone and the
    audit chain keep the act, the id and a SHA-256 of the record as it was,
    never the content. Tombstones are append-only at the schema
  • POST /identities/{id}/reset: forgets everything learned in one governed
    action; the identity, its declared attributes, obligations and consents
    remain. The identity id is repeated in the body as confirmation
  • GET /identities/{id}/export: the whole record with a fingerprint over its
    canonical form, the sentence that says how to recompute it, the lineage of
    every chain event about the identity, and a receipt that signs the
    fingerprint and names the chain head
  • Every control is policy-evaluated (allowed by default, a policy can deny or
    escalate), receipted and chained. Conformance test covers all five; the
    suite exercises all 51 operations
  • docs/RICORSA.md: how Ricorsa's identity graph, regions, provenance chain,
    recursive learning loop and Discover map to the platform, and which rules
    the platform now guarantees
  • The served contract has one stable address, /igl-openapi.yaml; the
    versioned names redirect to it

Changed

  • GET /capabilities lists subject-controls; the seven MCP tools are
    unchanged on purpose: a model must not be able to change what is learned
    about a person