Four-path comparison

Identical intake snapshots flow to all four model/strategy paths, arranged as an intentional model-by-strategy 2×2 grid. BOUNDARY routes deterministically (bypassing the model for exact-match and missing-PO); LLM-first consults the model for every invoice. All results use offline simulated adapters; decision-call counts are not live AMD API requests.

Path panels (model × strategy)

DeepSeek-V4-Flash

LLM_FIRST shadow

6 decision calls 0 bypassed 1314 tokens

interpretation 6 / 6 no-call 0

DeepSeek-V4-Flash

BOUNDARY operational

2 decision calls 4 bypassed 438 tokens

interpretation 2 / 2 no-call 4

routing 6 / 6 correct

Gemma-4-31B-IT

LLM_FIRST shadow

6 decision calls 0 bypassed 1518 tokens

interpretation 6 / 6 no-call 0

Gemma-4-31B-IT

BOUNDARY shadow

2 decision calls 4 bypassed 506 tokens

interpretation 2 / 2 no-call 4

routing 6 / 6 correct

Per-invoice × per-path states

States derived from the recorded event stream (REQUEST = one model call; BYPASS = deterministic hold with zero decision-stage calls).

Invoice DeepSeek-V4-Flash/LLM_FIRSTDeepSeek-V4-Flash/BOUNDARYGemma-4-31B-IT/LLM_FIRSTGemma-4-31B-IT/BOUNDARY
INV-200 consulted DETERMINISTIC consulted DETERMINISTIC
INV-201 consulted DETERMINISTIC consulted DETERMINISTIC
INV-500 consulted INFERENCE consulted INFERENCE
INV-501 consulted INFERENCE consulted INFERENCE
INV-600 consulted HELD consulted HELD
INV-601 consulted HELD consulted HELD

Accumulated batch results

whole batch: 16 decision calls, 8 bypassed operational: 2 calls shadow: 3 paths operational outcome: 2 approved, 4 held, 0 blocked, 0 verification-failed