Skip to content

Spillway Progress

Internal planning projection. Canonical sources: docs/EXECUTION_MAP.md, docs/PROJECT_STATUS.md, and current GitHub issues.

Last planning evidence: 2026-09-13
Latest verified Candidate: 202609121805 on both macOS and iOS, from d12a65b — Confirmed, promoted to Spillway Testers 2026-09-13. E1 and E2b physically validated on the Mac.

Current priority

Priority Workstream State Next meaningful step
P0 #574 Teach Spillway (user-generated grounding) E1/E2 validated and closed: machine assertion → presentation identity → explicit human response → journal projection, proven on real devices; new primitives carry a presumption against them Slice A: free-text teaching captured before AI interpretation, async derived interpretation, Edit/Undo, Recent Teachings; no autonomous action from inferred scope
P0 #573 comparative priority calibration designed small ranking sets as durable observations; forward-stagewise sparse adjustments over existing priors
P0 iPhone / cross-device credibility iOS is plausibly independently usable; ordinary Mail, Triage/Action, attachments, and category-suggestion sync have physical evidence; extracted tasks do not reach iOS (#126/#19) targeted restart/offline/install-over checks and workflow-disposition assertion/withdrawal validation
P1 Knowledge measured-value loop K1A–D, K2A/B, K3-A merged; reversible/grounded evidence representation exists validated E1/E2 path, #443 as needed, bounded context, then K3-C-style quality/coverage/abstention/compute measurement
P1 First Task/Work integration durable decision substrate exists and E2b proves the response pattern on one Mac surface; Work projection is disconnected; Work Ledger folds into Task/Work View reuse the E2b pattern for authored decisions; #126 iOS task transport/display; #572 Dates as Now/Soon · Plan Now · Important Ahead · Recently Passed
P1 #525 observation persistence cost safe amortized-retention half merged; remaining cost is dominated by whole-store persistence choose a durability-preserving storage evolution when warranted
P1 System interoperability designed #423 compose/handoff → #523 temporal handoff → #530 inbound iOS Share Extension

Key current facts

  • iOS crossed from companion/testing surface to a plausibly independent daily-use client; this is not a claim of full parity or complete physical acceptance.
  • Email.disposition preserves authorship/provenance and crosses devices as exact assertions plus identity-targeted withdrawals.
  • Cross-device transport is iCloud Drive ubiquity-file sync, not CloudKit.
  • Production sync retains journals and re-reduces them. Unknown operation kinds and newer reducer versions are skipped per operation; MaterializedStore quarantine/suffix-blocking is not the deployed path.

  • 554 completed the durable-vocabulary sweep: OperationKind, EvidenceSource, and AttentionMark are registry-covered, while reducer-version checks remain the actual applicability gate.

  • Assertion, rejection, and withdrawal are distinct. Correlation is not warrant. Missing metadata cannot manufacture human independence or grounding.
  • Recursion permission is distinct from inference strength. Machine/recursive/synced evidence may strengthen a permitted inference but cannot bootstrap deeper recursion permission.
  • One production human-decision → semantic-evidence path exists and is physically validated (E2b): Confirm/Reject task on the Mac is journaled and rendered as evidence (including .rejects) at read time. No .rejects row is stored in SemanticEvidenceStore, no consumer acts on it, and the journal has no transport.
  • Stamped responses carry exact respondingTo (correlation, never requiredSupport warrant) and project as solicitedHuman; unstamped ones stay correlationUnresolvedHuman. The real corpus holds a small amount of exactly-correlated human grounding; most of it remains ungrounded.
  • iOS Prioritize is attention/work intent, not task confirmation; a reply is behavioral evidence, not completion.

  • 525's retention scan has been amortized; remaining observation cost is dominated by whole-store persistence.

  • WorkComparisonRunner is observational comparison infrastructure, not a semantic-evidence or human-decision consumer.
  • Inbound Share-to-Spillway from Safari/other iOS apps remains unimplemented even though outbound attachment sharing and file-based sending work.

  • 486 remains a release-tooling hazard: stale local release state is not proof of a new upload.

Projected order

E1 / E2 — validated on 202609121805                              ✓
        ↓
#574 Teach Spillway → #573 comparative calibration (user-generated grounding)
        ↓
first broad conservative semantic consumer
        ↓
Knowledge measurement / calibration
        ↓
Task/Work + Dates integration (#19, #126, #572)
        ↓
system interoperability

physical sync validation, observation accumulation, and reliability work continue in parallel

Architecture lessons

  • Ambiguous human signals are evidence, not automatic semantic judgments.
  • Human, model, deterministic, synchronized, and recursively derived authority remain structurally distinguishable.
  • Synchronization is transport, not corroboration.
  • Durable identity does not imply unchanged semantics or applicability.
  • Derived evidence cannot manufacture the authority needed to justify deeper derivation.
  • Observations are irreproducible; declarations need real enforcement tests where feasible.
  • A hardcoded list over an open vocabulary is not proof of future exclusion.
  • OS handoff should use canonical intermediate representations rather than bespoke provider pipelines.
  • Candidate physical testing establishes facts automation cannot.
  • After #554, new semantic substrate needs a concrete failure mode; integration and empirical calibration are the default next phase.
  • E1/E2 physical validation exercised the whole chain on real data; the next primitive has to be demonstrated as a concrete gap by a real consumer.

Maintenance contract

This page is a display projection only. Keep it aligned with current repository evidence rather than copying stale planning text forward.

Documentation provenance: Iterative human–AI construction. See Documentation Provenance.