Source revision: 1ceae52c7896a59ac74a4429d410f3218f0b91c6. The original
partial review below has now been completed by the
forty-four-slide Iris guide and the
depth review. This establishes complete
teaching of the two pinned Iris pages, not memory safety, consumer UI or
deployed-service acceptance.
Learning sequence and reviewed units#
The Iris guide now separates identity, scope, recall, relevance scoring, consent, inspection and a worked write/recall comparison. Two focused slides were added for scoring and inspection; the count follows those learning tasks, not a fixed guide-size target. The shared substrate explanation remains unchanged. The Eve memory-consent slide is corrected in both Iris and the Eve context guide so the overview does not repeat the disproven guarantee.
Six reviewed assignments bind the
assembled explanations, including inherited chapter/release, to complete units
in V1/architecture/substrate-iris.md:
| Unit | Lines | Teaching home |
|---|---|---|
| Scope: eight kinds, not five | 132–164 | iris-scope-model |
| Consent on the entry vs. the ledger | 192–210 | iris-consent-and-inspection |
| Per-surface budgets | 295–312 | iris-recall-resolution |
| The ranker and its honest approximation | 313–336 | iris-ranker-factors |
| The audit envelope | 337–350 | iris-recall-resolution |
| Conflict resolution — provenance beats recency | 375–395 | iris-worked |
The full architecture and Iris feature pages were reviewed. Implementation review covered the complete pinned entry/primitives schemas, recall pipeline, in-memory store, write-conflict helper, consent ledger and inspection state machine, plus the adapter scope declaration. The privacy feature introduction and consent/withdrawal discussion establish its stated pure-core/runtime boundary; this is not a full privacy implementation review.
Corrections backed by executable code#
tools/presentations/tests/pinned-iris-memory.test.mjs transpiles and executes
the original pinned modules. The entry schema uses the actual pinned primitives
with the installed contracts package's Zod dependency. Recall calls the actual
sensitive-category predicate. Write-plan tests guard unused audit/redaction
dependencies with throwing replacements; they do not claim that append-audit
integration was exercised. Controlled dates, candidate records, requests and
sink acknowledgements keep these examples deterministic. Transpilation is not a
TypeScript project typecheck or a deployed authorization test.
Write policy is not recall deduplication#
The write helper compares adapter records using explicit metadata keys or a
short text prefix. Matching different-content candidates are ordered by
createdAt, not assertedAt. Metadata source-kind overrides precede consumer
and ordinary source inference; integrations must validate their provenance.
Prior user-stated/confirmed facts are protected from inferred writes, including
admin-copilot inference. Other conflicts need a response even when their rule is
named most-recent-wins. accept_new and merge allow a plan; neither
persists data in this function.
Recall independently groups canonical entries by category, actor handle and
normalized subject, then picks the newer updatedAt, regardless of origin or
score. Equal times retain the first ranked candidate. Different wording can
produce different keys; this is not semantic contradiction detection. The
architecture's claim that the provenance ladder also runs during recall is
contradicted by its pinned implementation. The short-versus-long preference
example executes both paths and demonstrates opposite winners.
Relevance, reachability and audit#
- Scoring blends recency 0.40, lexical Jaccard 0.35, capped reference frequency
0.15 and origin 0.10. Thirty days is the recency e-folding time, not the
half-life stated in the prose: the factor is
exp(-1)at thirty days and halves at about 20.8 days. A last reference supersedes the update timestamp for this calculation; nonpositive age scores one. - All eight canonical scope shapes and eleven adapter names are distinguished.
Pose-to-scene recall does not check room ID. Crisis scope alone is not
crisisFrame; tenant entries are excluded from crisis scope by an earlier branch. Tenant-null entries can pass the separate tenant check; matching scope/user constraints still apply. Top-level entry user ID is not checked against scope user ID by either the entry schema or recall. - Gates precede scoring: tenant, scope, crisis, suppression/lifecycle, then
sensitive consent/context. Sensitive multi-actor entries accept either their
category consent or
multi-actor-sensitive, with relationship/safety/crisis context instead of the category's ordinary context mapping. suppression.endsAtgoverns suppression;startedAtis ignored. Recall does not inspect entryexpiresAt, entry consentrevokedAt, the consent ledger or a revision store. Current grants and eligible revisions must be derived by its caller.- Budgets are 12/3/999/100 for assistant/shell/notebook/admin. Rationale is marked visible only for notebook/admin; it is a template, not a factor-level explanation. Missing surface budgets fall back to shell.
- Gate reason counts omit conflict and budget losers. Optional audit ingestion
is not awaited, so a successful result is not a persisted receipt. Admin
non-DSAR redaction removes
multiActor, not body text.
Revisions and erasure#
The schema has 29 categories, not approximately 28; 19 are sensitive. It validates a 1–4000-character body but accepts whitespace-only text and does not validate consent, revision topology or identity agreement across fields.
MemoryStore holds in-memory maps, not durable storage. Updates produce linked
revisions but also change the prior stored lifecycle to superseded; the old
record is not byte-for-byte immutable. External returned copies cannot mutate
the stored snapshot. A mutator can remove original provenance links, while the
audit chain is preserved separately. recordReference adds another revision.
summarize adds ancestor references without changing source-entry lifecycles.
Tombstones retain body/history and block subsequent updates to that entry in this store. They do not erase data or invalidate a previously copied active entry supplied to recall. Full erasure and non-revival require evidence from the persistence, replication, summaries and rights integrations.
Consent and inspection#
Consent append checks user, duplicate IDs and capture chronology, but does not call the full ledger validator. Invalid supersession/fingerprints can be appended and then rejected by explicit validation. Events/arrays are not durable immutable storage merely because their TypeScript fields are readonly.
The 64-bit FNV-1a event fingerprint binds selected fields; it excludes metadata and witness display labels and is not an authenticated signature. Rewriting a bound field and recomputing its fingerprint passes. Stored terms hashes are not automatically compared against live rendered terms. Active views choose the latest captured event per consent type, not per scope, and do not honor future effective/capture times. Only time-expired grants are downgraded.
The inspection state machine enforces transitions and a policy over supplied request fields. Consent snapshot IDs and signed incident references are checked for presence, not fetched or authenticated here. Target sensitivity and granted entries come from the caller; the gate does not validate each entry against target identity or recompute sensitivity. Non-DSAR metadata redaction leaves body text intact. Audit objects and a 72-hour pending-notice queue are local artifacts, not persisted audit records or delivered notices. No production privacy violation or authenticated exploit is inferred from these module-level counterexamples.
Completed source decomposition and remaining product verification#
The generated v1-iris-depth.json assignment now binds every meaningful unit
that was open in these two source pages. Keep the following product and
integration work open:
- Complete entry field, category, provenance, suppression and audit inventories; schema constraints and runtime refinements must be taught distinctly.
- Trace adapters, storage admission, retention envelopes, promotion thresholds, compaction and provenance through actual consuming routes. The write helper and canonical recall are not assumed to share one integrated pipeline.
- Teach privacy-suppression storage/recall helpers, revocation cascade, PII tokenization and multi-actor namespace/association forgetting in full.
- Cover all consent enums, validation errors, transport conversion, mutable aliasing boundaries, scope derivation and current-terms integration.
- Cover data-rights payloads, transitions, verification/deadline/retention behavior, result fingerprints, actual erasers, export artifacts and fan-out.
- Trace authenticated inspection callers, all policy conditions, snapshot selection, audit persistence, carveouts, notice delivery and replay.
- Inspect the partial customer memory UI and mobile equivalent against all intended edit/pause/forget/export flows, with actual screenshots and deep applicable browser/mobile automation. No UI-completeness assertion follows from these pure-module tests.
- Teach canonical continuation tokens, cross-device conflict/reconnect, scene/pose and shell bridge paths with end-to-end evidence.
- Reconcile every claimed evaluation/release gate with executable definitions and results. Module tests do not establish global leakage, deletion, deployed-service or full-library acceptance.
- Finish narration for the entire library, normal-speed listening, final all-slide/PDF/diagram review, whole-monorepo inventory expansion and actual Eve integration in the order required by the TODOs.
The design keeps one dominant comparison or sequence per slide. Detailed counterexamples stay in notes and cited source context. Browser playback and decoded-pause checks must not be reported as a human normal-speed listening review.