CALIBER
Quickstart
Strategy

CALIBER v1.0.0 product requirements document

Product requirements and delivery evidence for CALIBER v1.0.0 as a bounded, enterprise-usable single-tenant product.

Decision-makerArchitectDeveloperOperatorStrategyINTERNAL
PrerequisitesRead docs/roadmap.md for sequencing, milestones, and capacity assumptions · Review the current GitHub Project #2 issue hierarchy and dated delivery evidence before approving a release
Reviewed 2026-08-17 · current main branch, GitHub Project #2, and the proposed v1.0.0 release

Status and decision

Status: draft for approval. This PRD defines the product contract that the roadmap delivers. It does not assert that CALIBER has reached v1.0.0 or that all visible alpha capabilities are supported.

The proposed v1.0.0 product is an enterprise-usable, single-tenant CALIBER deployment for one governed prompt-refinement journey. A named operator can take a flagged MLflow trace through verification, diagnosis, candidate creation, evaluation, authorized Apply, and auditable release, reconciliation, or rollback. The journey is available through the documented UI and a small, contract-tested API/SDK/CLI subset.

The product decision is deliberately narrower than the repository's alpha surface. It prioritizes an operable and recoverable path over broad lifecycle coverage or distributed scale.

Problem and users

Agent-development teams can observe a poor model or agent outcome but often cannot safely turn that evidence into a reviewed production change. The work is split among tracing, prompt authoring, evaluation, release tooling, and operator runbooks. Failures across a non-transactional provider boundary can leave an uncertain live state.

UserJob to be doneProduct outcome
AI/application developerTurn a flagged trace into a measurable candidate without losing the provenance.A repeatable diagnosis, refinement, and evaluation path.
Authorized operatorInspect evidence and explicitly apply or recover a change.Visible approval, release state, reconciliation, and rollback actions.
Platform operatorRun one tenant safely and restore it after a failure.A documented, observable, one-process topology with rehearsed recovery.
Engineering/product leadDecide whether the product is ready to support.Dated pilot, security, workload, and acceptance evidence—not feature counts.

Product goal and success measures

The goal is not to make every CALIBER asset family production-ready. The goal is to make one named prompt-refinement path supportable for one enterprise tenant.

Measurev1 success conditionEvidence owner
Supported journeyThe canonical journey completes from a clean, documented environment with expected evidence.#102–#107, #119
Release integrityExternal-effect failure, timeout, retry, duplicate Apply, rollback, and reconciliation are explicit and tested.#105, #113–#114, #152
Security boundaryUnsafe bootstrap, secrets/egress, authentication, and project isolation are refused or verified.#122–#125, #159
Operational readinessOne-process ownership, bounded workload, recovery, upgrade, and incident behavior have dated drill evidence.#130–#138, #157–#159
Release decisionNo unaccepted P0/P1 blocker remains; a human records the residual risks and go/no-go decision.#139–#142

Pilot PASS proves only the named commit, environment, fixture, and journey. It does not prove production readiness without the v1 security, recovery, and acceptance evidence.

Scope

Supported v1 journey

flagged MLflow trace
  -> human verification
  -> diagnosis
  -> candidate creation
  -> evaluation gate
  -> explicit authorized Apply
  -> durable release, reconciliation, or rollback evidence

The supported deployment is one tenant, one active CALIBER process, a durable PostgreSQL metadata store, a named MLflow tracking/provider profile, selected object storage when required by the canonical fixture, external TLS/ingress and network policy, least-privilege credentials, authenticated role-based users, logs/traces/metrics/readiness, backup/restore, and a human-operated upgrade and incident procedure.

Explicit non-goals

  • Multi-tenancy, SSO/SCIM, and cross-tenant lifecycle management.
  • HA, active-active or multi-worker operation, broker acknowledgement/replay, and multi-region disaster recovery.
  • Managed hosting, Kubernetes operators, or automatic TLS issuance.
  • General-purpose untrusted tool/MCP execution.
  • A uniform lifecycle or production-support claim for every visible asset family, provider, optional dependency, API, SDK, or CLI operation.
  • An unmeasured scale or performance promise.

These are post-v1 decisions, not implicit v1 commitments. They are represented by Epic #58 and its child work.

Product requirements

IDRequirementPriorityAcceptance outcome
PRD-01Define one supported MVP contract, persona, provider profile, and capability boundary.Release blockerSupported and unsupported combinations are explicit; no alpha surface is implied to be v1 support.
PRD-02Make the canonical workflow durable and observable across every stage.Release blockerThe seeded journey is reproducible through the supported UI and automation surfaces.
PRD-03Require an authorized Apply and preserve release state across the MLflow boundary.Release blockerThe operator can see the intended before/after target and either a settled or reconcilable outcome.
PRD-04Publish a deliberately small API/SDK/CLI contract and actionable failure diagnostics.Release blockerOnly documented operations are supported; cancellation, retry, restart, correlation, and readiness behavior are testable.
PRD-05Provide controlled-pilot release evidence before v1 hardening begins.Release blockerA clean checkout, documentation, deployment procedure, bug bash, and human controlled-pilot decision are recorded.
PRD-06Enforce the one-tenant, one-active-process security and topology boundary.Release blockerUnsafe bootstrap/configuration is refused; roles, project isolation, redaction, egress, and loop ownership are verified.
PRD-07Operate and recover the selected v1 envelope.Release blockerWorkload limits, backup/restore, rollback/reconciliation, upgrade boundary, incident procedure, and acceptance packet have dated evidence.
PRD-08Keep post-v1 breadth separate from release scope.RequiredBroker, identity, lifecycle breadth, hosting, and scale remain explicit decision work unless a new PRD revision approves them.

Experience and operating constraints

The user experience must make the governing decisions visible, not inferred:

  • Verification and Apply are human decisions; evaluation evidence informs them but does not silently change the live target.
  • Every external release effect has a durable state and an operator-visible next action. Indeterminate effects must be labelled reconcile_required, not reported as successful.
  • The API/SDK/CLI contract is intentionally smaller than the current alpha API surface. Unsupported operations must fail clearly rather than degrade.
  • The hardened deployment runs one active CALIBER process. Starting a second process cannot silently produce duplicate scheduler, reconciler, or worker effects.
  • The release decision is human-owned. Automation, an agent, or a passing test suite cannot approve or merge a production decision.

Delivery traceability

The GitHub Project is the execution system; this table is the requirements trace. Individual issue bodies remain the source for implementation details, test commands, dependencies, and completion evidence.

PRD requirementPrimary epic/workstreamDelivery issuesExit evidence
PRD-01#78, #56#79–#101, #153–#154Capability matrix, pilot evidence/triage rules, MVP contract, provider matrix
PRD-02#56 / #60#102–#104, #107Seeded fixture; durable stage evidence; browser and API evidence; single-environment contract
PRD-03#56 / #61#105–#106, #152Release state machine; approval-gate test; failure-injection and reconciliation evidence
PRD-04#56 / #62–#64#108–#116, #151Operation matrix; SDK/CLI contract tests; failure taxonomy; recovery, correlation, and readiness evidence; topology decision
PRD-05#56 / #65–#66, #76#117–#121, #129Clean-checkout/package and deployment evidence; documentation parity; bug bash; human v0.1 decision
PRD-06#57 / #67–#68#122–#125, #130, #157, #159Rejected unsafe defaults; isolation/redaction tests; loop guard; hardened deployment and retention profile
PRD-07#57 / #69–#72, #77#131–#142, #158Measured envelope; backpressure; recovery objectives; restore, rollback, incident, regression, acceptance, upgrade, and v1 decision evidence
PRD-08#58 / #73–#75#126–#128, #143–#149Decision records or bounded prototypes after #142; no v1 release dependency

Release gates and timeline

The roadmap owns dates and capacity. The PRD gates those milestones as follows:

GatePlanned milestoneProduct decision
M1 — scope freezeSeptember 2026Approve PRD-01 using pilot and provider-profile evidence; choose the canonical fixture.
M2–M3 — foundation and contractsOctober–November 2026Prove PRD-02 through PRD-04, including durable release and recovery semantics.
M4 — controlled pilotDecember 2026Decide whether evidence is sufficient to begin v1 hardening. This is not the v1 release.
M5 — hardeningJanuary 2027Prove the security, one-process, topology, retention, and workload prerequisites for PRD-06 and PRD-07.
M6 — v1 decisionFebruary 2027Use recovery, acceptance, workload, and residual-risk evidence for the human v1.0.0 go/no-go decision.
Post-v1 reviewMarch 2027Select deferred breadth only from v1 evidence and demand.

Risks, assumptions, and open decisions

ItemStatusRequired handling
Current repository maturityObserved alpha/current-main stateDo not treat UI pages, routes, or unit tests alone as support evidence.
Provider compatibilityPlanned contract#154 selects and verifies one real provider path and one deterministic test path.
Worker/event topologyOpen decision#151 must select the supported model before the v1 topology is certified.
External release consistencyKnown non-transactional boundary#152 and #137 must prove settlement, rollback, or operator reconciliation.
Performance envelopeUnknown until measured#131 and #134 must publish the supported workload; #132 is conditional on measurement.
Deployment/upgradeMissing v1 artifact#157 and #158 must provide an operator-owned reference deployment and observed rollback boundary.
Two-week alpha checkpointProposal executed, Days 1–9docs/two-week-alpha-plan.md's Days 1–9 ran against current main as evidence feeding M1's discovery exit criterion; Day 10 (the release tag) is left to the repo owner. It does not change this PRD's supported v1 journey, scope, or milestones.

Approval and change control

Approval requires the product lead to accept this boundary and record the decision in #153. Any material change to the supported journey, topology, provider profile, or non-goal must update this PRD, the roadmap, the Project hierarchy, and the affected issue acceptance criteria in the same pull request.

This PRD is superseded only by a versioned revision that identifies the changed requirements and the evidence required to support them.

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.