Citadel is building the evidence layer between an optimization policy and its claim. It controls what an agent operation may run, records what actually ran, and asks a deterministic verifier outside the routed model whether the work counts. A cheaper failure is not a saving. An unknown is not zero.
Prompt routing is only one decision. Real agent economics also depend on topology, decomposition, retries, tools, timeouts, local versus hosted execution, verification, and recovery. Citadel turns those choices into a bounded contract whose result can be reproduced and rejected.
A selected executor cannot become the winner because it sounded confident, reported a low number, or produced a patch. Requested and observed runtime facts are reconciled, required artifacts are checked, and a model-external repository verifier decides completion.
Across a preregistered multi-stack benchmark, the funded controller targets:
Frontier must first verify at least 80% overall and 70% in every frozen task stratum, or the comparison is baseline-invalid.
Each rung answers a different question. Implementation evidence shows the mechanism exists. Retrospective evidence calibrates it. Prospective evidence shows the public runtime seam works. Six comparative studies show the progression from timeout sensitivity and escalation regressions to a passed, narrowly calibrated hybrid support envelope.
/do, repository state, campaigns, fleets, recovery, evidence, and handoffs work around the coding agent a developer already uses.
Ten frozen scenarios across three repositories and four economic policies preserve 33 verified, 51 failed, and 36 unknown outcomes.
VerifiedA thin adapter controlled a pinned recursive solver module by module. Its 24-cell diagnostic passed the evidence gate and failed the efficiency hypothesis.
VerifiedOne preregistered Claude Code operation on a fresh public clone matched model and topology, changed the required artifact, and passed a deterministic verifier outside the model.
VerifiedAcross 12 tasks × 2 policies × 3 timing repetitions, adaptive recorded 27/36 verified cells versus 24/36. The frozen aggregate used 9.9% less GPU energy, but excluding one same-route timeout pair reverses the comparison to 3.5% more.
Verified negative gateA separately frozen 72-cell follow-up matched 24/36 baseline cell completion, but 12 escalations caused 15.7% more GPU energy and 16.4% more modeled GPU cost.
Verified regressionSix artifact-producing fixture tasks, two policies, and two timing repetitions produced 24 signed cells. Both policies verified 6/12; zero false passes and path violations survived replay, while the 7.1% energy reduction missed the frozen 20% gate.
Integrity passed; economics failedClaude and Citadel each verified 12/12 fresh tasks. Citadel avoided four Claude calls and reduced comparison cost 28.4%, missing the frozen 30% gate.
Quality passed; economics failedOn twelve new tasks, both policies verified 12/12. Citadel used local 3B eight times, recovered once, reduced Claude calls from twelve to five, and reduced comparison cost 38.7%.
Every frozen gate passedTwenty-four distinct repositories produced eight calibration and sixteen untouched evaluation tasks. Direct Claude verified 2/16 and the sealed controller 3/16 at 1.26% lower comparison cost. The baseline was too weak for a general result.
Evidence complete; baseline invalidThe grant does not fund a promise to make Citadel look intelligent. It funds a public research program whose method, artifacts, cost lenses, and negative results remain inspectable whether the final performance target passes or fails.
Choose models, topology, decomposition, retries, tools, and stopping from calibrated outcome evidence instead of prompt difficulty alone.
Separate reported, derived, market-equivalent, marginal, and end-to-end cost. Include local hardware and energy without converting missing evidence to zero.
Exercise the same operation contract across multiple open and proprietary agent stacks without replacing their native planning or execution logic.
Extend the three local studies to multiple stacks, repositories, model families, hardware profiles, and tool routes. Freeze every task, baseline, policy, verifier, stopping rule, and economic target before execution.
Citadel has a positive synthetic result and a later outside-authored diagnostic that invalidated its baseline. The controller can be held accountable when a policy or comparison fails. It does not yet have the complete-cost, strong-baseline, multi-stack evidence needed for a general claim.
Sentient's product request asks for a layer that can optimize agents across models and workflows. Citadel contributes the operation-level control and proof discipline needed to tell a genuine economic win from a cheaper failed attempt. The first ROMA adapter makes that fit concrete rather than hypothetical.