04 · Evidence and roadmap

The roadmap is part of the work.

Complex change becomes manageable when the intended outcome is explicit, assumptions are replaced by evidence, dependencies remain visible, and progress requires proof rather than confidence. MOSAIK is the current implementation of that approach.

01
CompletedVerified foundation

Durable knowledge, governance, validation, and recovery

The organization needs an enduring memory and authority layer before orchestration can safely become more capable.

Preserve provider-neutral context, explicit information authority, bounded workflows, human authority, trace, replay, and recovery independently of any one AI provider.

The foundation remains stable while independent capabilities use the same rules, evidence, authority, and decision history.

The foundation has survived bounded interoperability and recovery tests.

18interoperability tests passed
41 / 41recovery items verified
1,769source analyses routed in the July 2026 baseline
19 / 20 + 20 / 20independent agent scores in the provider-replacement test
Evidence basis

Current acceptance test. In a bounded provider-replacement test, two independent AI agents used the same governed case structure and recovered each other's durable state without the original conversations. One agent scored 19/20 and the other scored 20/20; shared invariants passed 4/4, and bidirectional recovery passed 4/4. The result supports MOSAIK's provider-neutral foundation within the tested case.

Explore completed stage detail

One governed foundation, tested across independent agents.

Durable foundationRules, authority, evidence, trace, replay, and recovery
Independent agentsSame governed case structure, no original conversations
Recovered stateShared invariants and bidirectional recovery verified

The durable layer distinguishes evidence, interpretation, inference, proposal, decision, instruction, and runtime state. Replaceable capabilities can change without becoming the authority for organizational meaning.

Orchestration becomes fragile when knowledge, decisions, authority, or recovery depend on the same provider performing the current task. The durable layer had to exist before runtime capability could expand safely.

  • Provider-neutral boot and context behavior passes the accepted interoperability suite.
  • Memory routing, replay, conflict, rollback, and source-class controls pass without failure or skip.
  • The recovery manifest restores every expected item without missing, mismatched, or extra material.
  • Independent AI agents use the same governed case structure and recover durable state without the original conversations.
02
In progressCurrent build focus

One governed path from objective to eligible capability

Users should not have to make architecture, privacy, context, and model-selection decisions one prompt at a time. Ludwig is the planned orchestration function that presents one governed interaction plane while coordinating the capabilities behind it.

Build a thin, provider-neutral orchestration core that turns each objective into a durable ticket, derives only the minimum authorized context, selects the simplest qualified capability, issues narrower temporary execution authority for that route, validates the result, and preserves a recoverable outcome.

A bounded synthetic ticket survives restart, produces a separate resource-selection record, execution grant, and completion-validation result, fails safely under the accepted canaries, and runs within an approved recoverable operating boundary.

Current stage detail

One governed path coordinates the capabilities behind the interaction.

Owner-controlled conversation layerOne place to state objectives and resume work across interfaces
LudwigPolicy, context, routing, validation, and trace
Controlled execution environmentLocal models, hosted agents, harnesses, and deterministic tools

The governed core path

  1. Frame the workRecord the objective, scope, authority, limits, and completion conditions in a durable task ticket.
  2. Compile authorized contextUse the applicable recipe, available resources, and policy rules to assemble only what the task needs.
  3. Select the routeChoose the simplest qualified capability from the permitted options, based on testing and validation against the application’s requirements.
  4. Run bounded workGive the selected capability narrower, temporary authority that it cannot expand for itself.
  5. Validate and preserve the resultReconcile existing state, test the result against the ticket, report its completion status honestly, and retain the evidence and recovery state.

Where a validated result would trigger a consequential action, human review and controlled execution follow in later roadmap stages.

This stage replaces architectural assumptions with testable operating rules. It establishes the common control path required before a live conversation channel, personal context, local reasoning, remote access, or action can be trusted.

  • The durable task ticket, resource-selection record, bounded execution grant, and completion-validation result are separate versioned contracts with defined expected outcomes.
  • Context recipes identify required or eligible resource classes; the registry describes available resources; policy determines permission; and the router selects the simplest qualified route.
  • Synthetic fixtures expose preflight bypass, unauthorized resource access, stale or revoked grants, authority expansion, duplicate action proposals, scope leakage, audit suppression, incomplete work, and false completion.
  • The deliberately naive path fails the canaries before the governed implementation is accepted.
  • A ticket retains its objective, identity, scope, authority, completion contract, budgets, checkpoints, evidence, state, and provenance through restart, retry, cancellation, and timeout.
  • Replaceable workers cannot expand gateway-supplied identity, information scope, authority, tool rights, action rights, budget, or lifetime.
  • Validation reconciles existing state and decides whether the ticket is complete; worker assertions cannot release a result or create a duplicate action.
  • The accepted service footprint remains recoverable and does not materially degrade existing systems.
03
PlannedUseful personal service

Ludwig becomes present and useful

The governed core becomes a useful personal service through one authenticated conversation path, one exact capability, one qualified local reasoning route, and read-only evidence-backed retrieval.

Let one authenticated owner ask through a controlled conversation layer and receive exact deterministic results, validated local-model responses, and authorized memory retrieval through the same governed path.

The owner can use the accepted local service from approved devices; selective, exact, and exhaustive retrieval canaries pass; consequential claims remain traceable; incomplete work cannot appear complete; and unauthorized memory never enters context.

A small but complete read-only service proves identity, transport, deterministic execution, local reasoning, retrieval, evidence, and continuity together before access and input types expand.

  • Signed actor and room identity remain attached to every task, retrieval request, evidence item, and audit event.
  • One exact deterministic capability works without receiving private context it does not need.
  • One qualified local reasoning route returns structured, evidence-referenced results and recovers from interruption.
  • Each connected capability reports the permissions and isolation controls actually enforced; required safeguards fail closed when unavailable.
  • Selective lookup, exact evidence, exhaustive set, comparative, and exploratory work report the correct completion state.
  • Recall remains inspectable, proposed durable-memory changes remain staged, and continuity survives device, service, model, or provider change.
04
PlannedSecure reach and files

Secure reach and governed file understanding

Once the local read-only path is stable, MOSAIK can extend its reach and work with authorized files without weakening identity, privacy, coverage, or recovery.

Make the accepted Ludwig service securely available where the owner needs it and add governed document processing with source-level traceability and honest unsupported-item handling.

The same accepted service works locally and remotely through the approved public boundary, and an authorized document produces a grounded result with traceable sources and an accurate coverage status.

Remote access and files expand both usefulness and exposure. They should be introduced only after the core service can already enforce identity, scope, evidence, failure, and recovery boundaries.

  • Approved devices retain reliable access, notification, reconnection, and session revocation across network changes.
  • Only the approved conversation surface is publicly reachable; internal services and administrative interfaces remain private.
  • Attachments are retrieved through an authorized interface and checked for owner, scope, type, size, hash, and retention before processing.
  • Text extraction, later optional modalities, and every unsupported or failed item retain source, parser, location, and coverage status.
  • Temporary processing copies expire without altering or deleting source material.
05
PlannedMeasured capability choice

Evidence-driven capability and intelligent routing

Higher-capability workers, local specialists, deterministic tools, and combined routes should enter the system only when measured evidence shows that they improve an accepted result.

Qualify replaceable higher-capability routes and promote task-sensitive routing or bounded combinations only where they improve quality, privacy, reliability, cost, speed, context demand, or reviewer effort.

Routing is reproducible, information remains within eligible boundaries, every accepted connector can be removed without losing canonical meaning, and each promoted route or combination shows measured benefit over the simplest qualified baseline.

The expensive mistake is not using strong capability. It is using more capability, context, cost, or review than the accepted result requires. Stable tasks and evidence are needed before route comparisons become meaningful.

  • Each higher-capability connector completes a bounded task without exposing credentials or taking unapproved action.
  • Route eligibility applies privacy, identity, authority, tools, context, modality, consequence, availability, and cost before performance preferences.
  • Route selection runs in shadow mode until results justify promotion for a specific low-risk task class.
  • Delegated work can preserve or narrow, but cannot expand, its originating identity, information scope, action authority, tool rights, budget, or lifetime.
  • Combined work preserves source spans, disagreement, uncertainty, failure, and route identity rather than passing only worker summaries.
  • Specialist routes meet the same quality, privacy, traceability, validation, and recovery thresholds as the baseline.
06
PlannedControlled fulfillment

Controlled fulfillment and dependable operation

A validated answer becomes more consequential when it can change durable information or trigger action. Fulfillment therefore arrives with bound approval, verification, rollback, monitoring, and recovery.

Complete one governed proposal-only writeback, one approved low-risk action, and the operational hardening needed to run the personal service reliably.

The personal service is useful, observable, recoverable, revocable, supportable, and replaceable enough to admit another user safely.

Action and durable change require stronger controls than analysis. Multi-user activation should wait until the personal service can detect failure, reject stale authority, restore state, rotate access, and recover without silent change.

  • Every proposed change presents its exact target, owner, destination, difference, evidence, source freshness, and rollback before approval.
  • Approval remains bound to the authenticated user, request, result, scope, and expiration; changed, old, duplicate, or withdrawn approval fails closed.
  • Every approved write or action produces a redacted execution receipt showing its authority, effective safeguards, attempted and completed effects, verification result, and rollback status.
  • One low-risk action completes with post-action verification and a tested rollback path.
  • Monitoring, redacted logs, backup, restore, credential rotation, cancellation, degraded operation, incident response, and complete shutdown have usable runbooks and canaries.
  • A sustained personal-service canary passes the sovereign, recoverable, and captive tests against actual dependencies.
07
PlannedPrivate and shared use

Private and shared collaboration

MOSAIK expands from one private service to two independent users and a shared room without merging personal authority, memory, consent, or approval.

Support two separately authenticated and consenting users, independent personal contexts, an explicitly governed shared context, and appropriate owner or joint approval for collaborative actions.

Both users can work privately and together with persistent memory; prohibited cross-user retrieval fails closed; shared information and approvals survive correction, withdrawal, and membership change; and canonical meaning remains independent of models, clients, indexes, runtimes, and providers.

Collaboration is not a larger version of single-user access. It introduces independent ownership, consent, sharing, withdrawal, membership, and joint-authority boundaries that must remain visible throughout retrieval, memory, action, and recovery.

  • The second user directly consents and receives independent identity, personal authority, correction, export, withdrawal, and removal paths before personal context is loaded.
  • Each user can access a private service while every attempt to retrieve the other user's private context fails closed.
  • The shared room uses shared authority by default and accepts a private item only after its owner explicitly shares it.
  • Conflicting personal preferences remain distinct rather than being silently merged.
  • Consequential shared actions require the appropriate owner or both users unless an approved policy explicitly delegates authority.
  • Collaborative behavior remains correct through simultaneous requests, correction, withdrawal, membership change, partial failure, and any required communication media.

The strategy behind the sequence

Build tomorrow's capability while today's system still gives you room to choose.

Kodak, Xerox, Sears, and an emerging German counterexample show why adaptation is hardest while the existing system still works. The roadmap applies the same lesson: build the next capability before changing becomes unavoidable.

Read “The Danger of Defending What Works”