Citadel binds an operation's declared route to what actually ran, its measured and modeled cost lenses, and whether a deterministic verifier outside the routed model accepted the outcome. Failed policies stay failed. Unknown cost stays unknown.
Local studies exposed timeout, escalation, and baseline-validity defects. A calibrated synthetic support envelope passed. The outside-authored follow-up then showed that the retrieval/edit substrate and strong baseline were not ready.
Across 12 tasks × 2 policies × 3 timing repetitions, adaptive recorded 27/36 verified cells versus 24/36. Its frozen aggregate missed both 30% gates, and excluding one matched 60-second baseline timeout reverses all three economic comparisons.
Twelve new exact instances, mostly from task templates already seen in v1, routed work to 1.5B, 3B, or 7B. Model-external verification matched the baseline cell completion at higher measured GPU energy and modeled GPU cost.
Six fixture repositories × 2 policies × 2 timing repetitions. Both policies verified 6/12 cells. Citadel used 7.1% less measured GPU energy, below the frozen 20% gate, and 13.2% more tokens.
Both policies verified 12/12 fresh tasks. The risk-only policy avoided four Claude calls but made four unsupported local attempts, reducing comparison cost 28.4% against a frozen 30% gate.
Twelve new tasks, a valid Claude Sonnet 5 baseline, and a preregistered support envelope. Citadel used eight local attempts, recovered once, reduced Claude calls from twelve to five, and reduced comparison cost 38.7%.
Twenty-four distinct repositories supplied eight calibration and sixteen untouched evaluation tasks. Routes were published before calls. Qwen verified 1/16, direct Claude 2/16, and the Qwen-first controller 3/16 at 1.26% lower comparison cost. The baseline was too weak for a general result.
Each artifact answers a different question. The state at right is the gate's actual outcome, not a maturity badge.
All 120 signed cells preserve 33 verified, 51 failed, and 36 unknown outcomes; 84 cells reached a model and 36 remained setup-unknown.
Integrity passed 02 / ROMAA pinned recursive stack consumed the operation contract; control evidence passed and the efficiency hypothesis failed.
Policy failed 03 / RUNTIMERequested and observed model/topology matched, the public clone changed as required, and a deterministic repository verifier outside the model passed.
Integration passed 04 / LOCAL V112 tasks × 2 policies × 3 timing repetitions; 27/36 versus 24/36 verified cells; frozen economic gates failed and timeout sensitivity reversed the economic direction.
Gate failed 05 / LOCAL V212 exact task instances × 2 policies × 3 timing repetitions; matched baseline cell completion; verifier escalation made measured GPU economics worse.
Policy regressed 06 / REPOSITORYSix artifact-producing tasks × 2 policies × 2 timing repetitions; both policies verified 6/12 cells; integrity gates passed and economic gates failed.
Gate failed 07 / HYBRID CALBoth policies verified 12/12 fresh tasks; four Claude calls were avoided; 28.4% comparison-cost reduction missed the frozen 30% gate.
Gate missed 08 / HYBRID V2Both policies verified 12/12 new tasks; Citadel reduced Claude calls from twelve to five and comparison cost 38.7% with every gate passed.
All gates passed 09 / ONBOARDFive unattended engineering stages completed from a clean local clone in 28.17 seconds; the nested doctor command exited zero but reported semantic health as unknown.
5 stages completed 10 / PUBLIC HOLDOUT24 distinct repositories, 16 sealed evaluation routes, 32 official verdicts; direct Claude 2/16 and controller 3/16. Baseline validity failed for a general claim.
Baseline invalidCitadel is not presented as a universally best-in-class model router. It now demonstrates a narrower, useful result: an operation policy can be controlled, observed, graded outside the routed model, rejected when it fails, and accepted when it preserves verified outcomes while clearing a frozen economic gate.
This command runs seventeen offline checks across the retained failed calibrations, the passed hybrid result, the public holdout, routes, exact-answer verdicts, model identity, cost derivations, source bindings, artifact digests, receipt chains, Ed25519 signatures, public claims, the application package, and the site story.
Evaluator guidenpm run grant:verifyFunding expands the controller from frozen local pilots to multiple open agent stacks, task categories, model families, hardware profiles, tool routes, and complete cost lenses. The method and negative results remain public either way. Frontier must first verify at least 80% overall and 70% in every frozen task stratum, or the comparison is invalid.