Typical letter path (entry → Data Format code → Receptionist GPT5 → Validate Contacts code → Flash condition → Primary Classification GPT5 → GPT4_1 condition → one specialist GPT5 → one Variables code): 4 code/entry ×$0.025 = $0.10 + Flash condition $0.10 + GPT4_1 condition $0.30 + 3 GPT5 judgments ×$0.20 = $0.60 → $1.10/letter at the credit-rail card ($1.00 with GPT4_1 folded to mid). All-branches sum: 11×$0.025 = $0.275 + 8 GPT5×$0.20 = $1.60 + $0.30 + $0.10 = $2.275. Measured figure: ~$0.23/task at-cost billed on a private instance (the finance deck's 'complex regulated agent' anchor). $0.23 is 4.3-4.8x below the typical-path rate-card arithmetic — squarely inside the catalog meta note's 'at-cost lands 5-10x lower' band, so the gap is the billing rail (at-cost enterprise vs credit-rail ceiling), not a tier assumption; no model downgrade is needed to close it.

Measured: ~$0.23 per task at-cost, billed on a private instance — the 'complex regulated agent' anchor in the finance deck appendix, named as BID in the rate canon (at-cost billed)

Models seen in the graph: GPT5 on all 8 classifier judgments (Receptionist, Primary Classification, Schreiben InsoW, Schreiben Gericht, Restschuldbefreiung, MRH, SBP/Nullplan, FOA); GPT4_1 and GEMINI_3_FLASH on the two condition nodes; GEMINI_3_FLASH on all 10 CodeExecutors and the entry

Graph read: 04-agents/bid-coburg/prd-inso-agent-1/agent.json (21 nodes, all edges traced: entry → Data Format → Receptionist → Validate Contacts → condition → Primary Classification → condition fan-out to six specialist branches).

• Two-stage classification with a fan-out: Receptionist (sender + contacts) → Primary Classification into letter families → a condition routes to exactly one of six specialist classifiers (administrator, court, discharge, mandate/dormancy, SBP/zero-plan, FOA) — only one specialist bills per letter.
• Code brackets every AI step: Data Format before the Receptionist, Validate Contacts Array after it, and every branch terminates in a cheap Variables CodeExecutor that sets the routing variables — classification and routing are cleanly separated.
• No external system writes anywhere in the graph — it is a pure classify-and-route front-end (email intake trigger, prompts + Code Executor only); the FOA/BID-system hand-off happens downstream.
• Uniform model pin: all eight legal judgments on GPT5, conditions on cheaper models — the regulated-domain judgments get one consistent tier rather than per-node tuning.

• The repo's own metadata misled a reviewer: a draft called bid-coburg/ a 'dead archaeological copy' while the 237-task PRD_INSO_Agent - 1 lives exactly there — verify against the probe, not folder impressions (03-projects/30-module-library/04-outputs/four-line-module-summary.md §5).
• The circulated 85-94% accuracy figures are third-hand with no first-party measurement — don't quote them to a client (30-module-library/04-outputs/four-line-module-summary.md).
• $0.23/task is the at-cost anchor, an order of magnitude under credit-rail arithmetic — always quote the rail the contract actually uses (03-projects/14-agent-modularization/01-planning/two-axes.md line 244; solution-intelligence/content/agent-catalog.json meta note).
• The sibling LS Agent contributes the reusable pattern for unknown inbound types: sample-based taxonomy discovery → client confirms categories → triage — worth copying when a client can't enumerate their letter types (03-projects/alloyed-cx/04-outputs/production-cx-table-zh.md §D2).