The Cascade — Ranked by Asymmetric Leverage

Surfaced from factory forensic audit, queue archaeology, and 3 deep-dive sessions. Apr 9, 2026 CT.

Ralph Outcome Enforcement

BUILDING NOW

90% of completed tasks (304/339) have NULL outcome. 100% of failed tasks (158/158) have no outcome. The factory is blind — we can't tell what Ralph built or why things failed.

Fix: Add outcome parameter to complete_task() in pulse.sh. Wire meaningful outcome text from every call site in ralph.sh — branch name, files changed, test results, failure reason.

→ Step 1: Every task gets a written outcome on completion

→ Step 2: Dashboard can show WHAT was built, not just that something finished

→ Step 2: Verify cards get real context instead of guessing

→ Step 3: Failure patterns become queryable — auto-detect poison pills, model hallucinations, recurring blockers

→ Step 3: Morning digest can summarize overnight work with specifics ("merged 3 branches: X, Y, Z") not vague counts

~30 min build → fixes all future tasks permanently Compounds: every task from now on carries proof Principle 8: No Claim Without Proof

Fixes: Factory Defect #1 (Blind Factory), #4 (Silent Failures) — full forensic report

Supabase Management Token Refresh

JASON-DEP

The Supabase access token (from March 10) returns HTTP 401. This blocks run-sql.sh for DDL operations and the Neural Registry SQL migration (4 of 7 tables missing).

Fix: Jason generates new token at supabase.com/dashboard/account/tokens → paste into /opt/forge/config/supabase-access-token.

→ Step 1: run-sql.sh works again for DDL migrations

→ Step 2: Neural Registry migration runs — 4 remaining tables created

→ Step 2: Any future schema changes can be automated (no manual Supabase UI)

→ Step 3: Neural Registry enables event-sourced replay, goal tracking, librarian queue

2 min Jason action → unblocks all SQL DDL Token expired ~30 days ago silently

Verify Coverage — Close the Feedback Loop

READY

56 verification checklist entries exist (35 verified, 15 pending, 6 needs_work). But 280+ completed tasks have NO verify card at all. The human verification loop exists but covers <20% of work.

Fix: Backfill verify cards for high-value completed tasks. Fix generate-test-guide.sh model selection (groq-llama-8b hallucinates commands — use gemini-2.5-flash or Claude Sonnet).

→ Step 1: Every completed task gets a verify card with real test steps

→ Step 2: Jason's verification queue fills with actionable items — "test this, click that, expect this"

→ Step 3: Needs_work feedback creates rework cards with human context — Ralph gets better instructions on retry

→ Step 3: Pattern emerges: which types of tasks fail verification most → feed into task decomposition

Depends on Domino 1 (outcomes make better test guides) System exists — coverage is the gap

IHub Phase 1 — Knowledge Flywheel Foundation

READY

IHub Phase 1.5 (flush.py) is built and verified — extracts structured knowledge from sessions via Groq. Phase 1.6 (stenographer replacement) committed. The raw material pipeline works. Next: compile daily extractions into searchable knowledge base.

Fix: Complete IHub Phase 2 — compiler reads daily JSONL, deduplicates, indexes into searchable wiki. 7 Ralph cards queued (#766-772).

→ Step 1: Session knowledge auto-extracts and compiles daily

→ Step 2: Any session can search "what did we decide about X?" and get real answers

→ Step 3: Knowledge compounds — every session makes future sessions smarter

→ Step 3: This IS the Knowledge Flywheel thesis — Forge becomes the proof

7 Ralph cards ready to execute Flywheel effect: value grows with every session Foundation for MasteryOS knowledge layer

Parked — Important But Not Dominoes Right Now

235 Unmerged Branches PARKED

Factory Defect #2. Completed tasks with unmerged code. Need triage script to categorize: merge-ready vs stale vs conflict. Not urgent — code isn't lost, just undeployed. Unblocked by Domino 1 (outcomes tell us which branches matter).

Worktree Isolation NEEDS /ALIGN

Other session (98de5ef8) designed in-conversation only — no DESIGN.md written (Rule 0b violation). Critical for multi-session safety but needs proper /align → /design → /plan. Lifecycle scripts (lifecycle-gate.sh, session-reaper.sh) committed at 3c788b707 on ralph branch, will reach main on next deploy.sh merge.

PRISM (Codebase Memory) PARKED

Valuable concept — codebase-wide memory that persists across agents. No code exists yet, no Ralph cards. A tile until IHub proves the knowledge flywheel pattern works. Then PRISM becomes the code-specific instantiation.

Expert Knowledge Graph PARKED

PRD "Ready to Build" but test data needs Brian's expert docs. Infrastructure (KFS-Standalone at github.com/jdmac-msp/KFS) is built. Park until Brian material arrives — can queue build cards for skeleton without test data.

SESSION-QUEUE Cleanup PARKED

7+ items confirmed stale/done. Clean during next /wrap-up. Items aging 15-30 days. Low blast radius.

Reconciliation Timer FIXED

Was INACTIVE. Found correct name (forge-reconciliation.timer, not forge-reconcile). Installed and enabled via privileged-exec this session. Now runs on schedule.

Source sessions: 346754ca (this session, forge user) + 98de5ef8 (jason user, closing handoff)

Related: Factory Forensic Report

Queue cleanup this session: 29 → 19 pending tasks (5 perceive completed, 5 duplicates cancelled). Backup at /opt/forge/backups/queue/pending-snapshot-20260409.json

Generated: Apr 9, 2026 ~15:30 CT | Session 346754ca