BRM-loading · layering pigment…
BRM-loading · layering pigment…
Kernel-router · proof of capability · D13
One hundred and thirty-four model profiles, ranked empirically across eighteen probe categories. Cache-routed dispatch saves 2,340 tonnes of CO₂e per year against naive cross-vendor calls — a 80% reduction at the modelled sustained load. Per call: 0.0111 g CO₂e. The substrate refuses to route when its empirical evidence falls below the published floor; it discloses, it doesn't invent.
Live router introspection · client-side · no network call · §0.5 anchored
RoutingDecision · live
category inferred
reasoning_math
rule applied
ROUTE_REASONING_HEAVY
chosen class
Vendor — · —
estimated latency
0 ms · 0 probes in category
F × T × I band (BRM-3)
AVAILABLE / COMFORTABLE / PARITY
degradation level
insufficient_evidence
fallback chain
—
decision fingerprint
78ee99ec78ee99ec78ee…
rationale · no evidence in index · falling through
anchored to substrate audit · evidence held on disk · reproducible across silicons (BRM-1 parity)
Interactive leaderboard · category switcher · all 8 categories
| Rank | Provider class | Model class | Probes | Median latency |
|---|---|---|---|---|
| 🥇 | Vendor B | vendor-class-2-class | 1 | 48.0 ms |
| 🥈 | Vendor B | vendor-class-3-class | 1 | 49.0 ms |
| 🥉 | Vendor B | vendor-class-2-class | 1 | 56.0 ms |
live evidence · 8 categories present · empirical evidence held in substrate audit · cross-vendor parity preserved at the boundary
All 128 profiles · 3,048 sieved responses · 5 provider classes · 18 probe categories. Provider and model names render as generic class labels by design; the empirical evidence behind every rank is held in audit and cited on demand.
What the router does
Latency-sensitive
Pick the class whose median response time is smallest, across all categories where that matters.
Reasoning-heavy
Pick the class whose accuracy on multi-step reasoning probes is highest, even at the cost of latency.
High-frequency
Pick the class whose throughput tier accommodates sustained load without queue back-pressure.
Code-intensive
Pick the class whose performance on syntax-aware probes leads, across language families.
Quota-aware
When primary classes are quota-pressured, fall to the next-best ranked alternate, not a random vendor.
Refusal-on-floor
When the empirical floor isn't met, the router stops routing. Visible refusal is the substrate's accountability contract — silence is not an option.
Cache-hit savings waterfall · the load-bearing finding
naive baseline
2,925 t CO₂e/yr
no cache · per-call API fired
cache savings
2,340 t CO₂e/yr
80% hit ratio · 5× reduction
routed (residual)
585 t CO₂e/yr
0.0111 g/call
Decomposition · 11 engineered sub-systems
The router is itself an engineered system, so the substrate decomposes it the same way it decomposes a 91 km collider or a hyperscale training pod. Each branch carries a citation; missing citations are refused, not invented. 11 of 11 branches are fully cited at this scenario.
kernel-profile-index
$2,550,000
routing-decision-engine
$2,700,000
brmapi-dispatcher
$2,925,000
http-endpoints-trio
$2,100,000
benchmark-floor-breaker
$1,650,000
cross-host-fabric
$4,320,000
audit-chain-storage
$1,350,000
ui-panel-server
$1,260,000
api-keys-vault
$4,800,000
pipeline-orchestrator
$1,425,000
rack-power-cooling
$810,000
Accountability · verify, don't believe
A published floor governs whether the router answers at all. When the empirical evidence dips below that floor, the router visibly refuses — it does not silently degrade or pick a vendor on guess. Visitors can ask the substrate for the live floor at any time; the answer is auditable and refreshable. The substrate's accountability contract is that you see what it knows, not a confident gloss over what it doesn't.
