Status: normative execution overlay for the 425-stop architectural roadmap.
This document converts the consolidated implementation audit into a finite delivery strategy. It does not add release stops, weaken an invariant, or make an unimplemented capability Supported. The detailed roadmap remains the requirements and evidence catalog; this plan determines which capabilities form the selected production path, in what integration order they are proven, and which planning-only or optional profiles may remain unimplemented.
- The owner-mandated repository-wide prohibition on third-party Cargo code remains in force. N0, N1, and Hosted crates remain dependency-free until a separate explicit owner decision changes policy. An implementation-admission record is not an exception.
- No TLS, cryptographic, database, identity, MIME, Wasm, UI, or observability protocol is improvised to evade that prohibition. If a safe project-owned implementation cannot be evidenced, the affected Hosted capability and production claim are blocked or Unsupported.
- The 425 exact stops are a planning/evidence superset, not a requirement to
implement every optional profile serially before
1.0.0. - A capability becomes Supported only at its own exact stop with its complete transitive prerequisites and exact-commit pentest. Early spikes, integration branches, fakes, or prototypes grant no support claim.
- A skipped or deferred lower-numbered stop is never implemented later under its old release identity. Future implementation requires a new successor version and preserves the old Unsupported/Specified evidence.
- No further gap-analysis finding creates a release number unless it proves a genuinely independent missing authority owner or production-critical implementation boundary. Acceptance details go into the existing owner.
- Confirmation-only reviews close without roadmap expansion. Once an admitted package exists, concrete implementation, model, adapter, failure-injection, recovery, or pentest evidence is the source for corrective scope. A split is justified only when that evidence proves a separately shippable owner, prerequisite, migration, rollback/refusal, or pentest boundary.
- Every review is reconciled against the actual current commit and workspace, not the revision or file inventory asserted by the review. Findings from a stale snapshot must reproduce on the current tree before they can alter a version, dependency, evidence state, or support boundary.
| Finding | Disposition |
|---|---|
| G-01 Hosted dependency contradiction | Retain the owner-selected no-third-party policy. Execute 0.38.0 as a time-bounded feasibility/no-go owner, not a mandate to reimplement mature TLS, cryptography, PostgreSQL, OIDC, or hostile-input stacks. Missing safe dependency-free technology immediately marks the capability Unsupported and blocks later Hosted implementation plus the 0.398.0 production gate. |
| G-02 425-stop serial production path | Replace serial-all-stops interpretation with the selected mandatory-profile dependency closure and the capability increments below. Optional/deferred stops do not block the selected core claim. |
| G-03 logical Phase C/retirement monolith | Enforce the authority decomposition below. Dynamic built-in-domain retirement/reinstall is not in the default v1 profile; tenant deletion and safe plugin uninstall remain separate mandatory lifecycles. |
| G-04 oversized milestone prose | Treat phase ledgers as context only. Before code, each owner extracts a focused protocol package and executable model. No coordinator may absorb child state machines merely to avoid another file or crate. |
| G-05 late persistent authenticated slice | Preserve numeric ownership. 0.119.0–0.121.0 provide a PostgreSQL-backed transport/domain test slice with static identity facts only. The first production-authenticated incident proof occurs after 0.225.0, consumes rather than duplicates every earlier owner, and cannot back-promote the test slice. |
| G-06 premature five-store/migration scope | PostgreSQL is the reference production backend; SQLite is a bounded local/single-node profile. MySQL, MongoDB, SurrealDB, and online arbitrary-backend migration remain experimental/deferred unless separately promoted by evidence. |
| G-07 late technology choices | The Increment 0 dossier and bounded spikes complete before affected Hosted API, schema, or adapter contracts freeze. Final 0.371.0–0.398.0 decisions revalidate; they do not make the first selection. |
| G-08 prose-only protocols | Retain 0.44.0 and 0.56.0; every critical later state machine must register an executable finite model and adapter correspondence before its own exit. |
| G-09 qualitative performance/recovery | Every implemented deployment profile carries a versioned quantitative workload envelope from the first vertical slice onward. “Bounded” without numeric values is not passing evidence. |
| G-10 unclear first topology/TCB/UI stack | Increment 0 freezes the first single-node process/TCB boundaries and the exact HTTP/TLS/identity/UI technology disposition. HA remains a later extension of an operated profile. |
| G-11 broad product boundary | The v1 capability contract and explicit non-goals below govern API, UI, installer, documentation, and connector claims. |
| G-12 placeholder foundation types | Existing 0.2.0 and 0.3.0 are hard gates before envelopes, durable schemas, or tenant-bearing authority code. |
| G-13 selected closure remained prose | Make SelectedProfileManifestV1 an exact 425-row machine-checked artifact owned by reopened 0.1.0; use zero-or-more capability edges, graph-wide cycle/selection validation, claim-specific closures and work-package exact refinement rather than calling a previous-mandatory serial chain a transitive dependency graph. Generation 2 is the current non-weakening successor. |
| G-14 small-pass/model admission remained prose | Add ImplementationWorkPackageV1, the concrete 0.2.0 package, a reviewed minimum critical-model set, ExecutableModelBindingV1, and project-owned validation before implementation begins. Every entering package decides criticality with rationale; a list is never called complete merely because its existing rows have bindings. |
| G-15 PostgreSQL/SQLite order contradicted the selected profile | Assign PostgreSQL to 0.57.0 and bounded SQLite to 0.58.0; keep PostgreSQL mandatory and SQLite separately selected. |
| G-16 mandatory retirement certification consumed refused implementations | Keep the advanced 0.100.0–0.101.0 and 0.156.0–0.210.0 program Unsupported. Redefine default 0.414.0 as refusal, installed/disabled-state, retained-history, tenant-deletion/plugin-uninstall separation and no-exposed-retirement-path certification. A future selected successor owns destructive retirement certification. |
| G-17 delivery and optional-profile closure were ambiguous | Give rows explicit required_for_claims, allow multiple delivery slices, share search/knowledge 0.286.0–0.300.0 across both checkpoints, place 0.102.0–0.118.0 in service foundations, and make 0.424.0/1.0.0 depend on all promoted optional profiles. |
| G-18 package/model checking was predominantly syntactic | Require exact package prerequisites, path roots/allowlists, base commit, test IDs and criticality rationale; validate graph agreement, registered control IDs, unique transition/test IDs, and resolvable owners once implementation begins. Reviewer judgment remains required for conceptual cohesion. |
| G-19 stale companion-version report | The companion file is absent and stays absent. Do not recreate a second version authority from a stale untracked-file observation. |
| G-20 PackageExact could silently weaken the reviewed minimum | Preserve declared_minimum_dependencies independently beside effective package edges. PackageExact is a superset unless an approved DependencySupersessionV1 binds the removed edge, affected controls, replacement, reviewer, digest and tests. |
| G-21 acyclic dependencies could point to future implementation stops | Retain graph-wide cycle detection and also reject every same-version or numerically later prerequisite. Later integration is a retest owner or successor stop, never time-travel in an earlier work package. |
| G-22 registered tests were syntactic rather than executable | Resolve every test through the symbol inventory and an executable project-owned list/run script. ExecutableTestEvidenceV1 binds implemented source, runner, toolchain, candidate commit, result and evidence digests, and the implementation gate runs the registered test. |
| G-23 repeated stale companion-file report | The reported path is absent in the reviewed workspace. Preserve that absence and the one canonical contiguous-minor authority; no deletion or new version is needed. |
G-24 0.2.0 could not change the files required by its own handoff |
Advance its planned package to generation 2 and admit exact ID/time/error sources, workspace version/lock/toolchain/SBOM truth, Phase A/package status, selected manifest, symbol/test/evidence registries, exact test/release runners, bounded evidence and pentest artifacts, threat/security records, README, release notes and changelog. WorkPackageClosureV1 rejects an omitted transition path before entry. |
G-25 0.2.0 package ownership covered only IDs while the milestone promised time and errors |
Keep one coherent existing 0.2.0 foundation-primitives release, but require all three primitive no_std crates, their thin vitheim facade and thirteen exact ID/time/error/facade test intentions. Replace the unrelated future VIT-INV-001 reference with the bounded phase-a-foundation-primitives owner. The closure checker rejects a missing crate, test family, verification owner, or authority owner. |
| G-26 work-package closure accepted undeclared additions, generation-ahead packages and orphan packages/rows | Correct the reopened 0.1.0 admission control rather than add a product version. Advance WorkPackageClosureV1 to generation 2: package generation is equal, normalized crate/path/test/verification sets are exact, authority is equal, and the package↔closure relation is one-to-one in both directions. Negative fixtures independently add crates, exact paths, globs, tests and verification owners, substitute a test, drift generation and create both orphan directions. |
The planning-control baseline remains owned by reopened 0.1.0; passing those
checks proves roadmap integrity, not future product implementation. The sole
next admitted product package is 0.2.0. It remains planned until one entry
commit freezes its admission base, exact path boundary, PackageExact
prerequisites, source-symbol owners, and registered executable test IDs.
The 0.2.0 implementation then stays inside the single coherent
foundation-primitives boundary formed by crates/vitheim-id,
crates/vitheim-time, and crates/vitheim-error, exposed by the thin
crates/vitheim facade, plus its exact admitted
version, status, registry, runner, evidence, pentest, security and release-note
paths.
WorkPackageClosureV1 generation 2 binds that exact scope independently from
the package. Its normalized sets and generation must be equal, and neither a
package nor closure row may exist without exactly one counterpart. Before the
package becomes implemented, all thirteen registered test
IDs must execute through scripts/run_0_2_0_tests.sh and carry candidate-,
source-, runner-, toolchain-, result-, and artifact-digest evidence. The exact
implementation commit then receives the version-specific pentest; only after
its findings close may the next selected stop receive a work package.
This handoff adds no version. 0.38.0 still owns the first bounded Hosted
feasibility/no-go decision, while 0.398.0 still owns final selected Hosted
profile revalidation. Neither decision may be pulled into 0.2.0, assumed by
earlier code, or bypassed because planning checks pass.
The final closure confirmation against 8adf4ed found no remaining planning
gap. Its version disposition is explicit: reopened 0.1.0 continues to own
the admission controls, 0.2.0 remains one unsplit foundation-primitives
version in planned, and no later package opens. The next change in product
state is the existing 0.2.0 entry transaction: freeze its real base commit,
register the exact source owners and executable tests, and change the package
and Phase A statuses together. This confirmation is not implementation,
pentest, release, or Supported evidence.
Vitheim v1 is a service-operations platform with event-sourced domain authority. The selected core market promise includes ITSM, SecOps, assets/services, bounded workflow, and their policy-safe correlation.
The mandatory core profile contains:
- N0/N1 typed identities, checked time, stable errors, cumulative budgets, canonical codec/envelopes, pure aggregate decision/evolution, and the event kernel;
- PostgreSQL multi-user storage as the reference production profile;
- HTTP/TLS API, external OIDC resource/client integration, sessions, tenant lifecycle, RBAC/ABAC, field obligations/redaction, and immutable audit;
- incidents, requests/catalog, problems/known errors, changes, releases, tasks, approvals, assignment, attachments/quarantine, SLA, notifications, knowledge, and lexical search;
- assets/services, one admitted discovery profile, bitemporal provenance and reconciliation, service/relationship graph and bounded impact analysis;
- selected alert and vulnerability ingestion, normalization, deduplication/ suppression without evidence deletion, security incidents, remediation, forensic custody, and policy-safe cross-domain correlation;
- basic durable workflow with sequential steps, human task, timer, activity, bounded retry, and compensation;
- deterministic export/integrity verification, backup/restore, upgrade, observability, operator administration, and a minimal service-desk UI that is an ordinary API client.
SQLite is a separately evidenced bounded single-node/local profile. It is not an HA or hostile-host isolation claim.
The following are optional or post-v1 unless independently selected and fully evidenced:
- MySQL, MongoDB, and SurrealDB;
- online arbitrary cross-backend migration;
- SAML, SCIM, semantic/vector search, federation, and AI;
- Wasm storefront/vendor packs and effectful plugin connectors;
- extended supplier, procurement, capacity, demand, and continuity modules;
- multiple discovery/provider families beyond the selected reference profiles;
- dynamic built-in-domain retirement/reinstall.
Optional capability code may depend inward on stable ports. The kernel and mandatory first-party domains never depend on it.
The exact selection, capability edges, dependency state, delivery slices,
claim requirements, required runtime profiles, successor-after-skip rule, and
support boundary for every stop are in
Selected Production Profile Manifest and the
generated docs/selected_profile_manifest_v1.txt. This manifest is required
before 0.2.0. Reviewed architectural edges begin as an independently retained
DeclaredMinimum; each entering work package supplies PackageExact as a
non-weakening refinement. Removing a minimum requires a separate approved
dependency-supersession security record. Every dependency must be numerically
earlier than its owner stop. No production claim passes until every row in its
graph closure is package-exact, implemented and evidenced. 0.398.0
revalidates and may narrow claims, but does not create the first selection.
The selected v1 is not:
- a general log lake or full SIEM query/detection engine;
- an EDR, RMM, MDM, or arbitrary remote-shell product;
- an OAuth/OIDC authorization server;
- a secrets manager, KMS, general database, message broker, or object store;
- a substitute for provider-side transactional guarantees;
- an autonomous AI agent;
- an active/active multi-region authoritative-write system.
Connectors ingest or invoke only exact admitted operations. “Automation,” “agent,” “correlation,” or “integration” wording cannot widen these boundaries.
These increments are integration and work-planning overlays, not new release numbers. They may exercise future owner implementations in bounded spike or integration branches, but mainline support and release evidence remain owned by the existing exact stop.
Complete before affected Hosted contracts freeze:
- execute the
0.38.0dependency-free feasibility matrix; - disposition exact PostgreSQL and SQLite access, HTTP client/server and TLS, OIDC/JWK/JWT and WebAuthn, cryptography/randomness, MIME/content handling, Wasm Component runtime, browser/UI build, KMS, and observability needs as safely project-owned or Unsupported;
- define the first single-node topology and TCB: API/UI server, worker, PostgreSQL, blob/quarantine boundary, external identity issuer, and operator boundary; split a process only for a documented trust/failure/scale reason;
- define the first HA successor topology without claiming it;
- record version/platform/license/maintenance/unsafe/native-code evidence and supported Rust target spikes;
- freeze the mandatory core profile and explicit non-goals above.
Failure is a useful outcome: it blocks unsafe implementation and makes the Hosted production claim unavailable under current policy.
Complete the existing Phase A owners without expanding them:
- non-interchangeable nonzero tenant/domain IDs;
- checked time/duration and stable kernel errors;
- cumulative non-resetting allocation/work budgets;
- canonical bounded values/schema/codec;
- typed command/event envelopes and pure decide/evolve;
- in-memory incident proof.
Fuzz, property, finite-model, and compile-fail tests must demonstrate determinism, boundedness, tenant typing, and invalid-ID rejection.
Implement 0.102.0–0.118.0 as the reusable service foundation: typed work
components, tasks, attachment quarantine, assignment, SLA/SLO, approval,
on-call, paging and notification behavior. These are product/application
authorities, not generic operational-kernel internals. Each aggregate or
lifecycle has its own model binding and cannot become a universal ticket table.
After PostgreSQL 0.57.0, 0.119.0–0.121.0 prove the first persistent
incident transport/domain slice with static, deterministic identity and policy
facts. This is deliberately not a production authentication proof:
static tenant/subject/policy facts
-> CreateIncident typed command
-> pure incident decision
-> PostgreSQL expected-version atomic work bundle
-> event + result + audit + outbox
-> disposable incident projection
-> policy-shaped HTTP read and minimal UI
-> projection deletion/rebuild
All Phase D/E Hosted paths before 0.225.0 remain contracts, fakes, or this
test slice. They cannot be deployed with the fake identity profile, described
as production authenticated, or used to claim session/revocation correctness.
After the exact tenant, subject, session/OIDC, RBAC/ABAC, policy-lifecycle,
redaction, and authorization-conformance owners through 0.225.0 exist, prove
the smallest real cross-phase path:
OIDC-authenticated user
-> tenant-scoped policy and authorized-field plan
-> CreateIncident typed command
-> pure incident decision
-> PostgreSQL expected-version atomic work bundle
-> event + result + audit + outbox
-> disposable incident projection
-> authorized HTTP read and minimal UI
-> projection deletion/rebuild
The 0.225.0 conformance campaign consumes, rather than duplicates, the
existing journal, model, PostgreSQL, policy, API/UI, incident, tenant/subject/
session/OIDC, audit, and projection owners. It includes exact retry,
changed-material conflict, stale writer, crash at each bundle component,
notification fake, tenant-ID collision, policy and session revocation,
backup/restore, and upgrade. A real notification provider is added only after
its transmission profile is admitted.
No five-backend abstraction, broad domain suite, or production authentication
claim may substitute for this proof. Missing product behavior fails 0.225.0
and returns to its existing owner or a new successor minor; the conformance
commit cannot implement it. Passing evidence cannot retroactively promote
0.119.0.
Add scheduler/timer/lease/fence/queue behavior, basic workflow, quotas and measured fairness, external effect intent/dispatch/reconciliation, externally anchored integrity checkpoints, observability, failure injection, replay, restore, and rolling upgrade. Each authority uses a separate model and local owner transaction.
Deliver the mandatory ITSM set through real policy, storage, workflow, API, and notification paths. Shared work-item values and read composition never become a universal mutable aggregate, repository, patch command, or nullable mega-table.
Deliver the mandatory asset/service/SecOps profile with one real discovery and selected ingestion profiles. Graph and search remain disposable projections; raw facts/evidence remain immutable or governed erasable references; reconciliation is a pure pinned-input decision.
Search and knowledge 0.286.0–0.300.0 are a shared projection/application
closure used by both operable checkpoints, not security-only functionality.
Pure Wasm components precede effectful connector profiles. Builders, extended business modules, semantic search, federation, AI, additional databases, and vendor packs remain separately selectable and cannot become kernel dependencies.
Operate and harden the same selected minimum profile through packaging, startup/upgrade compatibility, HA fencing/failover, backup/restore/DR/failback, load/soak/fairness/rebuild/chaos, external security assessment/remediation, and the final support/refusal matrix. Phase O verifies decisions already made; it does not discover the first viable topology.
The combined VITHEIM-PRODUCTION-V1 promise still requires service operations
and security/asset operations. To obtain feedback without weakening that
promise, the selected manifest labels two independently operated evidence
slices:
core-service-operations: kernel, PostgreSQL, identity/policy, core ITSM, workflow basics, attachments, notifications, knowledge/search, and its API/UI paths.security-asset-operations: SecOps, vulnerability, custody, assets, services, selected discovery, bitemporal graph, impact, and correlation.
Each checkpoint has its own workload envelope, backup/restore exercise,
operational runbook, support boundary, and exact selected-stop evidence. The
second composes with the first through typed APIs and facts. Neither checkpoint
alone is Vitheim 1.0.0; production requires both plus the common
qualification closure.
The following owners remain separate even in one deployable:
- migration job owns migration progress only;
- import staging owns untrusted candidate bytes and validation only;
- activation owns local cutover CAS and exact domain-owner receipts;
- integrity owns event/checkpoint continuity;
- evidence custody owns retention, hold, archive, and deletion evidence;
- capacity/quota owns reservations and settlement;
- external witness owns append-only observations and equivocation state;
- corruption recovery owns quarantine and recovery authorization;
- tenant/domain retirement coordinates typed child outcomes but never embeds child internal state machines.
Interactions are immutable typed requests and authenticated receipts folded idempotently into one local owner. A combined transaction contains only the minimal guard rows needed to linearize that owner's action. Shared lock-order documentation does not create a shared authority or justify a mega-state machine.
For the default v1 profile, built-in domain schemas remain installed and may be
disabled. Dynamic built-in-domain retirement/reinstall and its advanced
0.156.0–0.210.0 profile remain Unsupported. Tenant deletion still follows its
mandatory lifecycle. Plugin uninstall removes an installation but does not
claim domain/data retirement.
For the default v1 storage profile, migration means versioned in-place PostgreSQL schema/data upgrade plus staged canonical export/import. Online arbitrary-backend migration remains Unsupported.
- Tenant scope is structural in every ID, port request, durable key, uniqueness/ foreign relationship, inbox/outbox, cache/index/blob/queue key, canonical preimage, pooled-session context, backup, restore, log, metric, and operator path. Generated/validated storage descriptors are the primary schema defense; handwritten predicates are not.
- PostgreSQL RLS/roles are defense in depth. SQLite truthfully refuses hostile process/filesystem-administrator isolation.
- Atomic command/consumer bundles commit result, events, audit, quota, and outbox together with expected-version/idempotency checks. An incapable backend refuses.
- Privileged-storage tamper evidence requires an independently retained checkpoint anchor. Without it, the claim is local corruption detection only.
- Lifecycle commands are narrow legal transitions. Assignment cannot mint roles, delegation, visibility, or eligibility; current epochs fence queued work.
- Policy produces an authorized read/field/traversal plan before sensitive selection. Notifications reauthorize at delivery or consume one narrow immutable disclosure grant.
- Graph traversal authorizes node, edge, intermediate node, field, and purpose before inclusion and bounds depth/nodes/edges/paths/time/work.
- Asset/SecOps reconciliation never mutates raw source facts. Absence, suppression, deduplication, confidence, scanner output, and remediation evidence retain their distinct meanings.
- Evidence custody preserves provenance, parent digests, uncertain/conflicting clocks, access/copy/transform/export/hold/disposition history, and truthful Complete/Incomplete/Unavailable/IntegrityFailed states.
- Attachment “clean” is a versioned scanner observation. Current authorization, verdict freshness, extraction budgets, safe origin/content disposition, and sandboxed rendering remain mandatory.
- Wasm fuel does not replace host-call, memory, instance, output, log, concurrency, time, destination, secret-handle, cancellation, and OS/process isolation controls.
- AI stays optional, advisory, proposal-only, and dependency-isolated from repositories, adapters, dispatch, workflow effects, secrets, and admin ports.
- Declarative registries name identity/owner/schema/dependencies/evidence only; executable transition semantics remain code-owned by the current feature.
Every Supported deployment profile publishes a versioned
MinimumProfileWorkloadEnvelopeV1 derived from reproducible benchmarks. It
contains conservative numeric maxima and refusal behavior for:
- command/event/payload/batch and attachment/archive expansion sizes;
- aggregate replay/snapshot work and projection/search/graph rebuild time;
- p50/p95/p99 command and read latency plus sustained/burst command, outbox, and inbox rates;
- tenants, subjects, roles, relationships, workflows, timers, alerts, assets, graph edges, search documents, and concurrent operations;
- per-tenant/worker CPU, memory, disk, connection, queue, and recovery reserves;
- projection/search/graph lag and provider-outage backlog drain;
- backup duration, restore duration, RPO, RTO, and integrity-check duration;
- Wasm fuel/host-call/memory/instance/output/time limits where enabled;
- AI context/output/token/concurrency limits where enabled.
The first vertical slice publishes an initial single-node envelope. Later profiles may strengthen or expand it only with new evidence. Requests outside a profile are refused before allocation or partial writes.
The consolidated review items are acceptance changes in existing owners:
0.205.0/0.209.0use one atomic Open/Sealed transmission-claim admission cut; sealing cannot race claim creation.0.207.0reports a successor operational only after its local activation has committed and an authenticated applied receipt is folded.0.208.0requires authenticated local and Vitheim permanent fencing of a failed successor before a replacement can claim single-issuer continuity.0.210.0gives preparation exhaustion a non-resetting explicit terminal/ recovery outcome; ordinary retry cannot manufacture another lineage.
They do not create more versions. If dynamic domain/status retirement is not selected for v1, the entire advanced profile remains Unsupported together.
This planning pass is complete when:
- the selected profile and increments govern
1.0.0evidence; - no dependency, topology, or product claim bypasses Increment 0;
- Phase C/retirement coordinators satisfy the authority split;
- every critical implemented protocol has a focused package and executable model;
SelectedProfileManifestV1remains exact, closed, acyclic, and consistent with numerically ordered, non-weakening PackageExact prerequisites and claim-specific closures;- every entering stop has a validated
ImplementationWorkPackageV1, and every critical stop has anExecutableModelBindingV1; every implemented test resolves, executes and binds exact immutable evidence; - numeric workload envelopes begin with the first vertical slice;
- optional and out-of-scope capabilities are discoverably refused; and
- the four late-protocol corrections are enforced by their existing owners.
Future findings belong in roadmap planning only when they expose a missing authority owner, impossible dependency/order, structural tenant/domain bypass, false production claim, or unrecoverable safety/liveness contradiction. Ordinary defects, race cases, fields, and negative tests belong to the owning design/code review and test suite.