Build the Vertical Slice

Build Mosaic as a sequence of complete slices. Each slice produces a runnable system and preserves the earlier tests.

Slice 1 · One deterministic run

Implement:

• Task, Budget, RunState, and RunError;
• Model and Verifier traits;
• scripted model;
• one structured output;
• in-memory trace.

Acceptance: a test covers success, invalid output, one repair, and exhausted repair budget.

Slice 2 · Forge CLI and file traces

Add:

• clap commands;
• YAML task parsing;
• JSONL trace persistence;
• strict replay;
• cancellation;
• redaction.

Acceptance: replay reaches the same terminal state with no external calls.

Slice 3 · Text and image evidence

Add content-addressed blobs, text locators, image regions, OCR, and an embedding adapter.

enum EvidenceQuery {
    Text(String),
    SimilarTo(EvidenceId),
    Region { image: EvidenceId, area: NormalizedRegion },
}

Acceptance: a generated claim cannot cite a nonexistent image region.

Slice 4 · Audio and video timelines

Add timed transcript spans, sampled frames, shot boundaries, and audio events. Start with external decoders behind traits before optimizing.

Acceptance:

• no decoded interval escapes source duration;
• the pipeline is bounded under a large input;
• cancellation terminates decoding;
• artifacts retain transform provenance.

Slice 5 · Real models

Add one remote adapter and one local capability:

• normalize request/response metadata;
• protect secrets;
• record model artifact or provider model ID;
• support timeout and retry policy;
• benchmark the local model.

Acceptance: either provider can replace the scripted model without changing program logic.

Slice 6 · Tools and authorization

Add a read-only source tool and a preview-producing transformation tool. Every tool declares effect and capabilities.

let authorized = authorizer.check(actor, run, proposed_call)?;
let receipt = tool_runtime.execute(authorized).await?;
let observed = verifier.observe_effect(&receipt).await?;

Acceptance: a model-generated call cannot bypass authorization, even during replay or retry.

Slice 7 · Durable API

Add Axum, SQLx, workers, object storage, leases, event IDs, and resumable SSE.

Acceptance:

• worker crash does not duplicate an effect;
• server restart preserves run state;
• client reconnect receives missed events;
• tenant isolation holds at retrieval and tool boundaries.

Slice 8 · Evaluation and routing

Add:

• versioned fixtures;
• deterministic graders;
• calibrated model grader;
• experiment reports;
• small/local route and strong fallback;
• quality, latency, and cost thresholds.

Acceptance: release tooling rejects a candidate that improves mean score while increasing severe failures.

Slice 9 · Multimodal generation

Add candidate generation adapters for at least two output modalities—for example image storyboards and speech narration.

Do not begin with final video generation. Use a funnel:

many cheap briefs
 → a few storyboard candidates
 → verified image + narration pairs
 → expensive video finalists
 → human-approved artifact

Acceptance: every output has provenance, safety status, cost, grader results, and approval state.

Slice 10 · Production hardening

• structured traces and metrics;
• per-tenant quotas;
• audit log;
• provider and local model health;
• data retention and deletion;
• chaos tests for timeouts and partial failures;
• canary and rollback;
• runbook.

The capstone demonstration

Run Signal Room and Cutroom on Mosaic. Then show:

1. the same evidence primitives support both;
2. local preprocessing reduces expensive model context;
3. a small-model path handles routine cases;
4. verifier failures trigger targeted repair or routing;
5. held-out evals establish the trade-off;
6. a risky generated artifact cannot publish without approval.

That demonstration proves you built an AI application platform—not a provider wrapper.