Experiment
How identical synthetic invoice opportunities behave under two models and two strategies. This is an evidence visualization — no live inference is performed.
Measured two-model comparison
S4E.15 — six synthetic development cases, one repetition. Historical AMD inference; results are not a prediction of typical performance.
| Model | Strategy | Requests | Tokens | Source-consistent |
|---|---|---|---|---|
| DeepSeek-V4-Flash | LLM_FIRST | 6 | 1316 | 4/6 |
| DeepSeek-V4-Flash | BOUNDARY | 2 | 438 | 2/2 |
| Gemma-4-31B-IT | LLM_FIRST | 6 | 1614 | 5/6 |
| Gemma-4-31B-IT | BOUNDARY | 2 | 572 | 2/2 |
Measured comparison: 12 LLM-first calls versus 4 BOUNDARY calls; 2930 versus 1010 tokens. That is 66.7% fewer requests and approximately 65.5% fewer tokens in this small development-set experiment.
Eight additional BOUNDARY zero-call cells are schedule-derived deterministic resolutions, not live model requests. Three source-consistency differences concern the missing-PO discrepancy convention. Exact model identities remain unverified.
Explore the recorded invoices
Original historical observations, not newly generated answers.
v2-DET-00 — deterministic_match — INV-200 (2 recorded model responses)
| Model | Strategy | Outcome | Source consistency |
|---|---|---|---|
| DeepSeek-V4-Flash | LLM_FIRST | SUCCESS | source_consistent |
| Gemma-4-31B-IT | LLM_FIRST | SUCCESS | source_consistent |
v2-DET-01 — deterministic_match — INV-201 (2 recorded model responses)
| Model | Strategy | Outcome | Source consistency |
|---|---|---|---|
| DeepSeek-V4-Flash | LLM_FIRST | SUCCESS | source_consistent |
| Gemma-4-31B-IT | LLM_FIRST | SUCCESS | source_consistent |
v2-DISC-00 — financial_discrepancy — INV-500 (4 recorded model responses)
| Model | Strategy | Outcome | Source consistency |
|---|---|---|---|
| DeepSeek-V4-Flash | LLM_FIRST | SUCCESS | source_consistent |
| DeepSeek-V4-Flash | BOUNDARY | SUCCESS | source_consistent |
| Gemma-4-31B-IT | LLM_FIRST | SUCCESS | source_consistent |
| Gemma-4-31B-IT | BOUNDARY | SUCCESS | source_consistent |
v2-DISC-01 — financial_discrepancy — INV-501 (4 recorded model responses)
| Model | Strategy | Outcome | Source consistency |
|---|---|---|---|
| DeepSeek-V4-Flash | LLM_FIRST | SUCCESS | source_consistent |
| DeepSeek-V4-Flash | BOUNDARY | SUCCESS | source_consistent |
| Gemma-4-31B-IT | LLM_FIRST | SUCCESS | source_consistent |
| Gemma-4-31B-IT | BOUNDARY | SUCCESS | source_consistent |
v2-NOPO-00 — missing_purchase_order — INV-600 (2 recorded model responses)
| Model | Strategy | Outcome | Source consistency |
|---|---|---|---|
| DeepSeek-V4-Flash | LLM_FIRST | SUCCESS | source_inconsistent |
| Gemma-4-31B-IT | LLM_FIRST | SUCCESS | source_inconsistent |
v2-NOPO-01 — missing_purchase_order — INV-601 (2 recorded model responses)
| Model | Strategy | Outcome | Source consistency |
|---|---|---|---|
| DeepSeek-V4-Flash | LLM_FIRST | SUCCESS | source_inconsistent |
| Gemma-4-31B-IT | LLM_FIRST | SUCCESS | source_consistent |
Large-scale offline verification
S4E.10 — workload-v1. Simulated model responses only.
| Observations | 720 |
|---|---|
| Synthetic cases | 60 |
| Independent integrity checks | 720/720 |
| Task successes | 480/720 |
| Task failures | 240 |
| Correct routes counted | 360 |
| Authority violations | 0 |
| Live model requests | 0 |
Malformed mock responses contribute to task failures. These results measure software behavior, not live Gemma accuracy. Workload-v1 is distinct from the later live workload.
Integrated Evolus-to-AMD trial (S4E.17)
Separate owner-confirmed live trial, not part of S4E.15. Genuine Evolus document extraction succeeded, and both DeepSeek inference paths (LLM-first and BOUNDARY) completed successfully. The four-path integrated run is partial.
| Evolus multipart extraction | SUCCESS |
|---|---|
| INV-500 validation | True |
| DeepSeek LLM-first | SUCCESS |
| DeepSeek BOUNDARY | SUCCESS |
| Gemma LLM-first | UNAVAILABLE |
| Gemma BOUNDARY | Not executed |
| Execution authority | NONE |
The ORIGINAL S4E.17 integrated run was interrupted when Gemma became unavailable from the authenticated AMD Token Factory catalog. This is a later catalog change, not a failure of the earlier eight successful S4E.15 Gemma requests, which remain separately preserved above. The original four-path execution is therefore partial.
Gemma integration completion COMPLETED — RECORDED LIVE
A separate post-restoration completion experiment (gemma-completion-20261016) ran both Gemma paths through the same validated Evolus extraction.
| Evolus multipart extraction | SUCCESS |
|---|---|
| Gemma LLM_FIRST | SUCCESS · source_consistent · 315 tokens · 3891 ms |
| Gemma BOUNDARY | SUCCESS · source_consistent · 315 tokens · 3516 ms |
| Verified discrepancy | $281.77 |
| Direct Gemma requests | 2 (no retries) |
| Execution authority | NONE |
The Evolus workflow's internal advisory agent activity is platform-internal and unverified, distinct from the two direct Gemma requests recorded here.
Routing journey ILLUSTRATIVE
Three synthetic routing outcomes through the accepted pipeline. Per-case traces are on the linked pages.
Provenance & limitations
- S4E.10: 720 mock/offline observations; no live inference.
- S4E.15: 16 historical live AMD requests across DeepSeek and Gemma.
- Eight zero-call comparison cells were schedule-derived.
- Three absent-PO convention differences remain disclosed.
- S4E.17: genuine integrated extraction and DeepSeek inference; Gemma later unavailable, original four-path run partial.
- October 16 completion experiment: both Gemma paths (LLM-first and BOUNDARY) succeeded through the same Evolus extraction — a separate record from S4E.15.
- Model serving identifiers remain independently unverified.
- S4E.15: sixteen genuine AMD requests across six synthetic development cases per model (one repetition).
- Financial approval is simulated; no authenticated human authorization or real payment is performed.
Additional records: Evidence & Performance.