ClauseGraph connects source clause -> reviewed rule -> dependency -> permitted action -> integer-cent cash consequence. The existing baseline includes source review, nominal CP-SAT planning, selected-action decision traces, nonmutating side-by-side previews and bounded fixed-plan verification. Preserve those workflows; do not rebuild them as proposed features.
Synthetic acceptance values remain: baseline minimum/end -40000/50000 cents; approved shift5000/50000; denied shift-40000/50000; cancellation alone-82000/8000. Deferral is timing, not savings. Verification is exhaustive over declared bounds, not a probability or an unbounded guarantee. SAFE requires complete resolved coverage; a concrete counterexample proves UNSAFE; incomplete checks otherwise produce UNKNOWN. Future debt remains visible.
The user expanded the scope and rebalanced remaining work on 2026-09-20 to approximately 35% A / 35% B / 30% C. The A/B/C assignments and work board supersede the old freeze sequence and optional-only C role. Refreshed main at 2e76537 includes the complete cash-gap workflow/C11 demo, C10 kit, and multi-dimension uncertainty controls. C12's dedicated failure view is still required. Preserve these merged deliveries rather than restarting them.
Preserve merged consent/closeout preparation and the PR30 incomplete-source correctness fix. Native timeout/OCR503 are recorded; successful extraction and human spoken rehearsal remain outstanding. C owns demo preparation; A owns live/correctness evidence and B shared UI integration. Automated rehearsal and provider connectivity do not establish human delivery or extraction accuracy. External blockers stay explicit while independent local work continues.
Required feature order:
- Idea #2: explain the minimum cash change. Generalize the existing offline diagnostic to a private, nonmutating fixed-schedule workflow with proof-qualified hypothetical opening cash, limiting dates and evidence. Money cannot repair missing authorization or evidence.
- Idea #3: uncertainty explorer. Expose the supported amount/date/approval dimensions and multiple incomes, with case-count warnings and understandable failure exploration. Every cent counts; large or incomplete domains remain honestly unknown.
- Idea #1: find a resilient plan. Review a separate specification, then search for one fixed permitted schedule that survives the declared domain. Preserve the distinction from nominal optimization and fixed-plan verification; no adaptive policies or per-outcome reoptimization masquerading as one robust schedule.
- Idea #4: cross-document consequence walkthrough. Connect the existing graph, evidence, deterministic effects and cash comparison into an action-focused explanation, including cancellation debt acceleration and retained future obligations.
A owns engine/contracts/generated types; B leads UI state/forms/shared integration; C implements the uncertainty failure view and consequence walkthrough and delivers demo materials/tests/rehearsals. Merge A contracts before B/C consumption. Each bounded task gets its own handoff and fresh worktree. C10-C15 replaces the old limited review queue, with exact paths and shared-file releases in the existing dev-b lane. C10/C11 are delivered; C12-C15 remain ordered work. A's synthesis design is proposed for B review; synthesis and consequence walkthrough remain unimplemented. C validates its deliverables; A/B mutual review remains required. No successful live inference or human rehearsal is claimed.
Idea #5, document-change impact, is if time permits after the required stages. Idea #6, review by decision impact, has lower priority still. A/B/C assess remaining capacity and A records it before activation; preserve implementation, integration and human rehearsal time. Exact scope/acceptance is in the assignment document. Correlations, uncertain expenses and adaptive policies are not silently included in resilient synthesis.
Keep a tested synthetic demo checkpoint and local fallback after each stage. Rehearse the final source -> review -> plan -> failure -> cash diagnostic/resilient alternative -> consequence story; update demo claims only after the corresponding implementation is merged and checked. Source support, confidence, human review, conditions and third-party approval remain separate. An empty review queue is not proof of safety.
Defer bank aggregation, automated payments/cancellations/applications/messages, broad benefits discovery, generalized recurrence, new providers and infrastructure rewrites. Public real-data deployment needs separate privacy/recovery/operational work. No paid provisioning or provider-accuracy claims from offline fixtures.
Current delivery/checks are in RESUME.md; verification semantics and integration history remain historical evidence. Shared docs are A-owned; each developer records only its own task handoff.