What Stage 4 looks like in numbers
The Shape of the Orchestration
3
Data streams
Development raise targets · Finance historical-and-projected · Pathstone market projections
11
Opus agents
fpa-lead · architect · data-engineer · quant · insight-hunter · analyst · strategist · modeler · viz · storyteller · frontend
13
Expertise references
Damodaran · Tufte · Knaflic · Hyndman · McKinsey · Bridgespan · Few · NACUBO · AFP · BoardSource · CFA IPS · NFF · NACUBO Spending
4
Verification layers
L1 source-cell provenance · L2 parser hard-gate · L3 runtime cross-foots · L4 defaults alignment
Why this is Stage 4 and not Stage 3
What Stage 4 Carries That Stage 3 Cannot
Artifact C ran the thought experiment: compress financial_modeling down to a Stage 3 Project. That exercise produced a workable Project but lost the discipline that makes the tool defensible to KPMG, the Investment Committee, the Finance Committee, the Audit Committee, and the Board. The Stage 4 architecture carries that discipline specifically:
- Per-cell provenance. Every value traces to a
{sheet, cell, value, last_seen}tag at parser intake. Untagged values are dropped at L1. - Parser hard-gate. The L2 gate halts intake if anchors, sheet names, or row labels have drifted from the contract. Structural drift cannot reach the math layer.
- Runtime cross-foots. L3 reconciliation against the workbook's own totals inside
fpa-quant. Math mismatches halt the lane. - Defaults alignment. L4 reconciliation against the IC-approved set before any surface renders. The Lead consolidates and decides whether to publish.
- Explicit handoff state. Each agent writes its output to a state file the next agent reads. Lead routes; Lead never overwrites. The audit trail across roles is itself an artifact.
- Multi-cadence reconciliation. Three refresh cadences (weekly Development, per-cycle Finance, quarterly Pathstone) all reconcile inside the same conversation. The L2-L4 gates run on every build.
A Stage 3 Project carries one workflow surface. A Stage 4 Orchestration carries many roles with verification gates between them. The cost difference is real, and so is the gain — but only when the work requires it.
The framework, applied end to end
The Same Eight Questions, Answered at Scale
The H3 walkthrough page walks all eight Week 7 questions against financial_modeling specifically:
- Q1 · Which container? Why three data streams forced the Stage 3 → Stage 4 promotion.
- Q2 · Do you have the expertise? The model identity (visibility into costs that exist, what we are committed to, what we can control), feeding three strategic surfaces (Development, Finance, Operating).
- Q3 · One role or many? The eleven Opus agents, each with one job and an explicit handoff.
- Q4 · Instructions or files? CLAUDE.md as the durable spec, the references/expertise/ folder as the world the spec applies to.
- Q5 · What skills chain in? Plugin layer (
finance:*,data:*), BRAIN-shared skills (preston-writing, funder-framing, dccap-brand, dc-cap-org-intelligence, executive-summary-formatter, data-interpreter, program-budget-intake), and the 13 references/expertise sources. - Q6 · Tools and permissions? Deliberately small. File read, scoped file write, scoped code execution. No web at runtime. No email. No external API calls during a model run.
- Q7 · How do you know good? The four verification layers run automatically. Each agent ships with cosmetic / costly / catastrophic failure modes named in its own instructions.
- Q8 · How does it stay alive? ADRs revisit quarterly. Reference updates land between runs, in the open. The audit trail across builds is itself the iteration cadence.