CALIBER
Quickstart
Architecture

CALIBER — The Refinement Loop

The canonical prompt-refinement path — verify, diagnose, optimize, evaluate, review/apply, and durable release — plus the evidence and recovery boundaries that differ across other asset families. Read this first.

ArchitectDeveloperEvaluatorConceptGA
PrerequisitesLayered architecture overview
Reviewed 2026-08-10 · current main branch docs contract

The one idea that ties the rest of these docs together. Read this before the reference pages: it explains why the building blocks exist and how they connect.

At a glance

Every other topic in this documentation — prompts, tools, skills, MCP servers, workflows, knowledge bases, test sets, evaluation, calibration, governance, and the Aria copilot — supports one product motion: turn a flagged production response into a measured candidate with an attributable deployment decision. The concrete prompt-refinement path below has six numbered stages and two human decisions: verification and review/apply. The seven-term Signal → Evidence → Candidate → Measurement → Decision → Release → Trace chain in the layered architecture is an abstract lifecycle map, not a second seven-stage worker pipeline: the incoming trace and feedback supply its signal, evidence assembly is folded into the transition from Verify toward Diagnose, and the next trace closes the loop after stage 6. Other asset families implement subsets of the lifecycle and must not be assumed to share its aliases, gates, or rollback semantics.

flowchart LR
    trace["Production trace<br/>+ flagged feedback"]:::src
    verify["① Verify<br/>human · 1 click"]:::human
    diagnose["② Diagnose<br/>LLM root-cause"]:::auto
    optimize["③ Optimize<br/>policy-selected optimizer"]:::auto
    evaluate["④ Evaluate<br/>judges + per-dimension gate"]:::auto
    approve["⑤ Apply decision<br/>operator · diff + eval"]:::human
    promote["⑥ Promote<br/>durable alias release"]:::ship

    trace --> verify --> diagnose --> optimize --> evaluate --> approve --> promote
    promote -.->|"code keeps loading @prod — no change"| trace

    classDef src fill:#f1f5f9,stroke:#64748b,color:#0f172a;
    classDef human fill:#fef3c7,stroke:#d97706,color:#78350f;
    classDef auto fill:#dbeafe,stroke:#2563eb,color:#1e3a8f;
    classDef ship fill:#dcfce7,stroke:#16a34a,color:#14532d;

🟡 human decision · 🔵 automated · 🟢 shipped. The dashed edge closes the loop: your application keeps loading @prod, and the next call transparently gets the new version.

The six stages

StageWhat happensReference
① VerifyA human confirms the flagged trace is actionable — one click.Platform
② DiagnoseAn LLM identifies the root cause from the trace and its evidence.Calibration
③ OptimizeA policy-selected optimizer proposes a fix. A manual pin or agent override wins; diagnosis heuristics choose among the remaining live paths.Calibration
④ EvaluateThe candidate is scored against a pinned test set with per-dimension regression checks. A pass advances the job to candidate_ready; it does not promote automatically, and registry gate verdicts elsewhere remain advisory.Evaluation · Test sets
⑤ Apply decisionAn operator-scoped actor (operator or admin; approver is a sibling scope) reviews the diff, evaluation comparison, and root-cause summary, then either invokes Apply or leaves the candidate unapplied. This is not a separate vote/quorum/reject API.Prompts
⑥ PromoteOn the canonical prompt path, Apply first commits an idempotent release operation containing the exact outgoing and target versions, then rotates the live alias and settles the operation. Ambiguous provider outcomes remain visible for operator reconciliation; rollback uses the recorded target. Other assets retain their own release semantics.Prompts

Why it matters

Most tools own only one arc — observability, visual building, or evaluation. CALIBER's bet is integrating the full prompt-refinement circuit with a broader asset inventory, self-hosted and MLflow-integrated. Its differentiator is that connected evidence/review/audit path, not a claim that every asset shares one unbypassable lifecycle. See how CALIBER compares to the alternatives and where it is headed.

How the rest of the docs map to the loop

The loop runs over the artifacts the following sections document:

New here? Start with the Platform overview for the embedded and standalone topologies, then this page for what the canonical refinement path does, then dip into whichever artifact you're working with.

CALIBER : Contextual Adaptive Lifecycle for Intelligent Build, Evaluation, and Refinement — this page is generated from the authoritative Markdown sources in docs/ and the repository-level ARCHITECTURE.md.