CEP-R REPORTING
Model-written report prose is checked against immutable deterministic facts.
INFRASTRUCTURE PASS MODEL QUALIFICATION OPEN7 retained cases · model calls 0 · solver runs 0
AERO / CONTEXT EFFICIENCY · CEP-01
INFRASTRUCTURE PASS · MODEL QUALIFICATION OPENGive each model only the context its task needs.
Revision 1.5 qualifies the context and routing infrastructure across six separate tracks. The engineering stack remains the authority; model qualification is intentionally still open.
42 full-vs-progressive comparisons · 7 retained evidence states · 0 solver runs
What this result does—and does not—show
Demonstrated: six independently scoped qualification contracts, compact prompt construction, typed evidence expansion, capability-based routing, and deterministic claim checks over retained pass/fail/unknown states.
Not demonstrated: engineering-reasoning competence, CFD/FEA correctness, model reliability, or readiness for consequential decisions. The run is a public metadata fixture and contains no physical simulation or provider calls.
Six claims, six qualification tracks
CEP-S STATUS / TELEMETRY NARRATION
Status narration preserves authoritative run state and does not invent progress.
INFRASTRUCTURE PASS MODEL QUALIFICATION OPEN7 retained cases · model calls 0 · solver runs 0
CEP-E ENGINEERING EXPLANATION
Routine explanations can request bounded evidence and retain open gates.
INFRASTRUCTURE PASS MODEL QUALIFICATION OPEN7 retained cases · model calls 0 · solver runs 0
CEP-C CONTEXT COMPRESSION
Progressive context reduces payload size without changing decision-grade meaning.
INFRASTRUCTURE PASS MODEL QUALIFICATION OPEN7 retained cases · model calls 0 · solver runs 0
CEP-RTR ROUTING / ESCALATION
Tasks route by capability, consequence, and evidence rather than model name.
INFRASTRUCTURE PASS MODEL QUALIFICATION OPEN7 retained cases · model calls 0 · solver runs 0
CEP-INV INVESTIGATION ROUTING
Failed-analysis questions receive bounded evidence and an explicit escalation contract.
INFRASTRUCTURE PASS MODEL QUALIFICATION OPEN7 retained cases · model calls 0 · solver runs 0
A track-level infrastructure pass is not a universal “model passed Aero” verdict. Provider qualification requires held-out, repeated runs for each track.
Context measurement
The full-context comparator was constructed locally and never sent to a model. Across 42 fixture comparisons, the progressive view is an indicative 93.82% smaller (range 92.98–94.54%). These are serialization estimates, not tokenizer-equivalent or cost guarantees.
Full-context comparator
Synthetic measurement payload
Contains repeated history/log/metadata scaffolding; not transmitted.
Progressive context
Decision summary first
Typed Level 2 evidence expands only when the track needs it.
Historical local report-drafting sample
| Local profile | Disposition | Input tokens | Verifier findings | Latency |
|---|---|---|---|---|
| gemma3:4b | DRAFT REJECTED | 242 | unsupported causal claim | 6.67 s |
| qwen2.5:14b-instruct | DRAFT REJECTED | 230 | wrong gate state; unsupported causal claim | 14.92 s |
| qwen2.5:72b-instruct | DRAFT VERIFIED ON THIS SAMPLE | 230 | None detected on this sample | 148.66 s |
This older one-fixture provider run remains linked for provenance. Its 72B result was a single-sample passer only; it cannot establish a reliable model ranking or qualification.
Authority stays with evidence
- ClassifyTask, consequence, complexity, and context budget
- SummarizeDecision-grade facts and open gates by default
- ExpandTyped evidence retrieved only when requested
- RouteChoose by declared task capability, not a fixed model name
- VerifyCheck generated prose against authoritative facts
- EscalateReturn uncertainty or high-consequence decisions for human review
Deterministic software remains responsible for calculations, thresholds, hashes, gate states, and provenance. Models may explain or draft; they cannot promote numerical acceptability or design readiness.
What remains
The next meaningful test is provider-backed, repeated, held-out evaluation for each track. Any changed disposition, lost unknown, or unsupported claim remains a qualification failure; the benchmark is not tuned away.
Revision 1.5 fixture date: September 16, 2026 · Model calls: 0 · Physical solver executions: 0