Evidence & Performance
Independently reviewed results from recorded live integration tests and offline benchmarks. This page is read-only and performs no inference or extraction.
AMD Inference — S4E.15 DEMONSTRATED — RECORDED LIVE
Recorded RECORDED LIVE TEST — NOT RE-RUN BY THIS PAGE. Sixteen confirmed live AMD Token Factory requests across DeepSeek and Gemma.
- DeepSeek-V4-Flash
- 8 requests · 1754 tokens
- Google Gemma-4-31B-IT
- 8 requests · 2186 tokens
- Combined
- 16 requests · 3940 tokens
- LLM-first
- 12 requests · 2930 tokens
- BOUNDARY deterministic-first
- 4 requests · 1010 tokens
- Request reduction
- 66.7%
- Token reduction
- ~65.5%
- Authority
- NONE
AMD Inference — S4E.5 earlier pilot MEASURED
Recorded October 13, 2026. Earlier DeepSeek-only pilot — RECORDED LIVE TEST — NOT RE-RUN BY THIS PAGE.
- Model
- DeepSeek-V4-Flash
- Endpoint-reported identifier
- self-dploy/DeepSeek-V4-Flash
- Model identity classification
- unverified_mismatch
- Confirmed responses
- 8
- Strategy A (LLM-first)
- 6 requests · 1314 tokens
- Strategy B (BOUNDARY deterministic-first)
- 2 requests · 473 tokens
- Request reduction
- 66.7%
- Token reduction
- 64.0%
- Authority
- NONE
Controlled eight-request pilot across three synthetic scenarios (two repetitions). Strategy A (LLM-first) made six requests (1314 tokens); Strategy B (BOUNDARY deterministic-first) made two (473 tokens). Request reduction is measured; token reduction is measured for this scenario set. No per-request latency improvement and no measured cost savings are claimed.
Google Gemma DEMONSTRATED — RECORDED LIVE
Recorded RECORDED LIVE TEST — NOT RE-RUN BY THIS PAGE. Eight confirmed real inference requests through AMD Token Factory.
- Model
- Gemma-4-31B-IT
- Endpoint-reported identifier
- self-dploy/Gemma-4-31B-IT
- Model identity classification
- unverified_mismatch
- LLM-first
- 6 requests · 1614 tokens · 5/6 source-consistent
- BOUNDARY deterministic-first
- 2 requests · 572 tokens · 2/2 source-consistent
- Total
- 8 requests · 7/8 source-consistent
- Authority
- NONE
The single source-inconsistent Gemma record is the absent-purchase-order evaluator convention (the invoice amount versus the evaluator's convention of zero), not a measured interpretation error.
View the full four-path comparison and replay →
Gemma history
- The first October 13, 2026 Gemma smoke attempt produced code-fenced JSON that failed structured-output parsing.
- Gemma-specific fenced-JSON normalization followed by strict parsing addressed that problem.
- Subsequent live tests successfully demonstrated Gemma inference (eight confirmed requests).
- Later, Gemma disappeared from the authenticated AMD Token Factory model catalog, preventing additional requests. This is an observed availability limitation — not a claim that AMD permanently discontinued Gemma.
- AMD later restored Gemma to the authenticated catalog (owner-confirmed, 2026-10-16). A single post-restoration request then succeeded with a source-consistent interpretation of the INV-500 discrepancy. The restored model again returned code-fenced JSON, confirming the normalization correction remains necessary.
- The original Evolus-to-AMD integrated trial was interrupted by Gemma's outage: DeepSeek completed its two paths while Gemma was unavailable. A separate October 16 completion experiment then ran the Gemma LLM-first and BOUNDARY paths through the same Evolus document extraction: both succeeded and were source-consistent (the $281.77 discrepancy), with code-fence normalization applied and execution authority NONE.
- The exact underlying serving-model identity has not been independently verified (the endpoint reported "self-dploy/Gemma-4-31B-IT", classified "unverified_mismatch").
Gemma integration completion COMPLETED — RECORDED LIVE
Separate post-restoration completion experiment (gemma-completion-20261016). Recorded RECORDED LIVE TEST — NOT RE-RUN BY THIS PAGE.
- Evolus multipart extraction
- SUCCESS
- Same validated packet
- 7ba9bf9a3f7fec4c26ea828f401005cbf656ff0d7779057037fa1f5d860c8aa2
- Invoice
- INV-500
- Verified discrepancy
- $281.77
- Direct Gemma requests
- 2 (no retries)
- Direct DeepSeek requests
- 0
- Authority
- NONE
LLM_FIRST · SUCCESS · source_consistent · 315 tokens · 3891 ms · normalization applied · identity unverified_mismatch
BOUNDARY · SUCCESS · source_consistent · 315 tokens · 3516 ms · normalization applied · identity unverified_mismatch
The Evolus workflow's internal advisory agent activity is platform-internal and unverified, distinct from the two direct Gemma requests recorded here.
Evolus Document Extraction DEMONSTRATED
Recorded October 13, 2026. RECORDED LIVE TEST — NOT RE-RUN BY THIS PAGE.
- Document type
- INVOICE
- Invoice
- INV-2048
- Supplier
- SUP-9001
- PO reference
- PO-77
- Stated total
- USD 1,250.00
- Advisory analysis
- present
- Evidence status
- unverified
- Authority
- NONE
A genuine API-driven synthetic document extraction completed in a separate live integration test; it was not jointly executed with AMD by this demo.
Offline Routing Benchmark OBSERVED_OFFLINE
- Synthetic categories
- 6
- Offline observations
- 18
- Repetitions per category
- 3
Six synthetic routing scenarios, 18 offline observations (three repetitions per category). These are actual offline routing measurements.
LLM-First Comparison COUNTERFACTUAL
- Hypothetical call opportunities
- 6
- Call opportunities avoided
- 2 of 6
A hypothetical LLM-first comparison: two of six call opportunities avoided by deterministic-first routing. This is arithmetic — NOT measured token, latency, or cost savings.
Limitations
- The recorded S4E.15 comparison is a small development sample (six synthetic cases per model, one repetition, sixteen genuine AMD requests); results are not a prediction of general production performance.
- Synthetic cases only.
- No per-request latency improvement was measured.
- No measured financial savings; the pilot does not claim end-to-end Evolus-plus-AMD savings.
- Human authorization is simulated only; no authenticated human-authorization gate is implemented.
- No comprehensive security certification.