Eve library

Iris · One sentence, and what decides whether it returns · receipt · 12 September 2026

Guide

Revision 1

On this page

Chapter five of Shared systems, delivered in the Eve edition on 12 September 2026. Forty-four inherited slides become a fifty-two-slide chapter under the iris id. The brief is authoring/iris-chapter.md. Machine-readable state is in state.json.

What was delivered#

Fifty-two slides in the guide: forty-two inherited slides rewritten in place, two carried unchanged, and eight new — a cover, six section dividers and a close. Nothing merges away.

The two carried slides already live in Eve-edition chapters: substrate-responsibilities moved into Sophia earlier today, and eve-memory-consent is in the Eve context chapter. The guide lists them; this chapter authors neither.

The forty-four arrived across five manuscripts. content/73-v1-iris-depth-a.json and content/74-v1-iris-depth-b.json held nothing but Iris slides and retire. content/06-evidence-memory-runtime.json loses six slides and content/23-shared-deepdives.json three; both keep the rest.

Evidence#

No capture. Iris has a member-facing memory centre in the plan and the pinned feature page marks it partial, so a capture would either show an unfinished surface or imply a finished one. Every claim is executed or read at the pin.

  • The entry contract, executed here. Twenty-one fields. Twenty-nine categories, six origins, five lifecycles, eight scope kinds. The schema refuses an unknown origin, an unknown lifecycle, an unknown category, a confidence above one, an empty body, a body over four thousand characters and a negative reference count — each at its own field. The body is not trimmed, so a whitespace-only string satisfies the minimum.
  • The retention envelope, executed here. Session raw thirty days and its summary ninety; scene fourteen and ninety; pose seven and thirty; operator copilot fourteen; tenant three hundred and sixty-five. In every scope with both, the summary outlives the raw detail.
  • The recall ranker, executed here. Four weights — recency 0.4, token overlap 0.35, reference count 0.15, origin 0.1 — summing to exactly one. Origin runs from 1.0 for what the member stated down to 0.4 for an operator copilot, with model inference at 0.5: what the member said outranks what the model inferred by two to one, and is worth a tenth of the score.
  • The decay constant is misnamed. The ranker computes exp(−age ÷ 30 days) and the constant in that expression is called halfLifeMs. It is the exponential time constant, not the half-life: at thirty days the factor is e⁻¹, about 0.368, and the actual half-life is about 20.8 days. The code is correct; a description copied from the constant's name is wrong by nine days, and the pinned prose repeats the error.
  • Surface budgets, executed here. Assistant twelve, shell three, notebook nine hundred and ninety-nine, admin one hundred; an unknown surface falls back to three. Rationale is hidden on the assistant and shell and shown on the notebook and admin — the two surfaces that return the most are the two that explain themselves.
  • Privacy suppression, executed here. Eleven sensitive categories, eleven recall contexts, and a policy per category naming the contexts it permits. No category permits general — so a member who granted consent for everything still hears nothing sensitive in an ordinary conversation. Crisis is permitted by eight of eleven. The storage gate refuses unless consent is granted: pending, denied, revoked and absent are all refused identically, and a memory touching two categories needs both.
  • Four categories reach no rule at all. Executed here by running both modules against each other: the contract marks nineteen categories sensitive, the suppression module declares policies for eleven, and an alias table carries four more (medical → physical health, sexual → sexuality, financial → financial distress, legal → legal jeopardy). Fifteen of nineteen are held back when nothing is consented in an ordinary conversation. Pose alignment, biometric, spiritual and safety pass through — no policy, no alias, no warning, no throw.
  • Redaction, executed here. Four PII kinds — email, phone, social security number, credit card — replaced with deterministic tokens. The same value yields the same token case-insensitively, so two systems can join on a reference without either holding the value.
  • Read at the pin. The store is in memory and does not call the schema parser; an update links a revision, appends audit, and rewrites the prior revision's lifecycle to superseded; a tombstone retains the body and the history. The consent fingerprint is a sixty-four-bit non-cryptographic hash that can be recomputed after rewriting a bound field. The inspection gate checks values the caller supplied and does not authenticate the operator, fetch the consent snapshot, verify the incident reference or confirm that the supplied entries belong to the target.

What this chapter's evidence tops out at#

Executing a pinned module proves what that code does with the inputs it was given. Reading a store proves what an author wrote. Neither establishes that any memory was durably written, recalled in production, or shown to a member. The four pass-through categories are a gap in the modules as they stand, not an observed leak: nothing here shows such a memory was ever written or returned, and the storage gate sits in front of the recall filter.

Checks#

Check Result
Build 1,682 slides, 701 sources; freshness check clean
Python suite 490 tests pass
Node suite 508 tests pass
Browser sweep 52 slides × 1440, 1280 with audio chrome, 390 and print — 0 findings
PDF 52 pages, exact authored visual text; the chapter has no table
Narration 52 tracks, 78.1 minutes; all 1,682 library tracks verified
Bindings 49 generated units; 12 reviewed bindings re-reviewed, 0 stale left

One sweep finding was fixed: a five-node record-anatomy ran 19px over at 1280 and its field details were shortened.

One number was wrong and was corrected before this shipped. Ten slides, the brief and this receipt said the entry contract had twenty-two fields. Counting the pinned schema gives twenty-one — the probe had printed twenty-one all along and the prose did not match it. Every instance was corrected, and the rest of the chapter's numerals were re-checked against the probe transcript and the pin: retention, the promotion threshold, the four weights, the six origin weights, the four budgets and the unknown-surface fallback, the nineteen and eleven and four category counts, the seven inspection states, the eleven adapter names, the seven member controls and the seven-and-thirty day rights deadlines all hold.

The Node suite's parallel run intermittently fails pinned-tara-discovery when free memory is low; it passes serially and passes in parallel with memory available. That was diagnosed earlier today against a clean tree and is a property of this box, not of the change.

Binding work#

Thirty-four of this chapter's slides are cited in assignments/v1-iris-depth.json, which is generated. The regenerated file holds forty-nine units, the same count as before the rewrite. The assembler restored fifty-six heading citations and folded one duplicate path.

The hand-authored reviewed bindings are the part the generator does not touch, and the check caught twelve of them stale — six belonging to this chapter, four to Sophia, one to the discovery chapter and one to the architecture chapter that preceded them. Every one was re-reviewed here against the rewritten slide rather than re-fingerprinted, and four of them moved, because the rewrite had put the teaching on a different slide from the one the binding named:

  • Per-surface budgets and the audit envelope moved off iris-recall-resolution onto iris-surface-budgets and iris-audit-envelope, which is where the numbers and the fields now are. The architecture-page heading citations moved with them, and iris-recall-resolution now cites the recall-resolution section it actually teaches.
  • Scope: eight kinds, not five and consent on the entry vs. the ledger each gained a second slide, because the rewrite split each unit in two: scopes across iris-scope-model and iris-adapter-scopes, consent across iris-consent-entry and iris-consent-and-inspection.
  • In Sophia the grounding-state thresholds unit moved to sophia-thresholds, the slide the rewrite gave the thresholds to, and that slide gained the architecture-page citation it teaches from.
  • In the discovery chapter, curation-ownership was carrying the same heading citation twice; the duplicate was dropped.
  • In the architecture chapter, the partial-failure unit now cites partial-failure-contract alone: architecture-failure was rewritten into a persistence-identity slide that no longer draws on the communication-patterns page at all, so it could not honestly stand as evidence for it.

check-center-coverage.py --check now reports zero stale bindings across the library; reviewed units went from 359 to 371. The three earlier receipts whose state.json claimed a clean check carry a correction.

Tests updated#

Four, each with the reason written beside it: the Iris assignment test (slide count 44 → 52, and the sequence now bounded by the cover and close); the rollback guard (iris moved from APPROVED_ADDITION_GUIDES to REWRITTEN_CHAPTER_GUIDES); the tracked-chapter count, 1,199 → 1,251; and pinned-iris-memory.test.mjs, whose two pinned layout assertions were updated to the layouts the rewrite deliberately produces, after the four phrases it guards were restored to the rewritten slides — "not one-half" onto the ranker, "not an authenticated cryptographic signature" onto the consent slide, "does not authenticate" onto the inspection boundaries, and "independently" onto the worked case.

That guard already looks slides up wherever they live rather than in a fixed set of files, so the move between manuscripts needed no change — which is the shape the other pinned tests are being moved towards.

Treatment deviations#

Two at the chapter level:

  • No capture, for the reason the brief gives: the member-facing surface is marked partial in the pinned source, so capturing it would mislead either way.
  • Two sections run to eight slides rather than five to seven. Rights and inspection each needed their whole argument in one place, and splitting either would have separated a control from the evidence it owes.

Eight inherited slides were tables and none needed to stay one. Beyond that, eight slides depart from the layout the plan named. Five of them are one recurring judgement: the plan offered a card grid or a graph-diagram for an evidence-comparison, and where the slide compares exactly two things across aligned rows, a compare-panel is the layout that lines the rows up.

  • iris-reality-posture, iris-adapter-scopes, iris-write-policy, iris-suppression-gate, iris-evaluation-boundary: card grid → compare-panel. Each is two things held against each other — three evidence layers reduced to what is proved against what is left open, two scope vocabularies, the write ladder against recall's freshness rule, the storage gate against the recall gate, and module tests against release evidence. A card grid would set them side by side without aligning the rows that make the point.
  • iris-worked: card grid → case-study. One input followed through two modules is a case, not a set of cards.
  • iris-inspection-lifecycle: step-journey → state-machine. Seven named states with legal and illegal transitions between them is a machine; a journey would imply everything proceeds in one direction, which is exactly what the slide says it does not.
  • iris-inspection-boundaries: step-journey → card-grid. Four families of check, each with an owner, are not four steps in an order.

What this chapter does not establish#

No member data was used, no deployed service was reached and no durable store was written. Nothing here shows a finished member memory interface, deployed erasure, authenticated inspection or cross-device acceptance.