ExergyNet measures the work required to retrieve, reason over, and act on persistent state, then redesigns the state-delivery path to reduce unnecessary computation without surrendering evidence integrity. Evidence is organized in three classes — computational efficiency, execution integrity, and state/evidence — each bounded to its tested envelope.
These results are not combined into one synthetic score. Each card states the measured result, the principle it supports, the test envelope, and the known limitation. This first class measures the cost of useful work: context utilization, corpus scaling, throughput, and hardware results. Execution-integrity and state/evidence results follow below. Every number is bounded to its tested hardware, corpus, and workload; H200 and A100 campaigns are reported independently and never blended.
11.3× more correct-task throughput than full-context replay at each method's highest tested SLA-compliant operating point on the documented H200 workload. The same evidence package reports a bounded 660-820 active-token xLMP prompt range as stored memory grew past the model context window.
87% to 94% symmetric correctness and 52,753.8× median paired speedup in the tested 1GB / 100-query enterprise population. The indexed path was faster on all 100 paired queries in that specific run, with 3.96× minimum observed speedup.
No single retrieval strategy won on every workload. LNES-86.2 selected among retrieval paths before expensive work and measured 1.886× vs always-legacy, 1.142× vs always-compact, 1.150× vs the first-generation router, 95.6% holdout route-selection accuracy, and 1.023× oracle gap.
Indexed retrieval is not universally faster. VMN-native validation showed aggregate median speedups below 1.0× at both 10MB and 100MB, while larger/filler-heavy subsets improved. That negative result led directly to adaptive deterministic routing.
The central scaling question: as the persistent corpus grows, does the evidence staged into active model context also grow? A ten-point ladder from 32K to 10M nominal tokens produced a linear fit with b = 2.79×10−6 (95% CI [−8.59×10−6, +1.42×10−5], includes zero; development data only). No material positive scaling trend was detected within the tested envelope. The 10M corpus point extended 312.5× beyond the 32K baseline.
A genuine sealed 5-run × 190-query holdout at the 10M corpus point confirmed K boundedness: 5/5 runs produced identical aggregate statistics and identical per-query staged-token counts (0/190 per-query mismatches), mean K = 895.96 versus 902.81 at 4M despite a 2.5× corpus increase. Correctness crossed the preregistered 40% floor at 39.5% (75/190 unique queries correct) — a retrieval-quality finding under adversarial density at 10M, not a system-level error: all 950 inference calls returned status=ok. Q1 (single-fact) held at 43.0% and Q2 (multi-evidence) reached 62.5%; Q3 (temporal authority) was impacted by the adversarial corpus design.
Execution integrity covers authority separation, transport/authenticated behavior, and rejection of unauthorized action. State & evidence covers durable, root-addressed state and its reconstruction. The checks below map to those two classes.
Retrieved memory can inform a decision. It cannot manufacture permission to execute one. Memory control, model reasoning, and execution authority remain separate planes.
The three classes above measure efficiency, integrity, and state. The frontier is measuring the verified delegated action end to end. The metrics below are proposed evaluation targets, not completed benchmarks — no result is claimed for them until measured under a stated envelope.
These public results publish the phenomenon. The forthcoming paper explains the architecture without exposing the private operating recipe.