Agent provenance
Design toward making the source of claims, outputs, and recommendations visible before they influence decisions.
Pre-alpha local-first AI infrastructure
Mythica helps turn scattered AI work into reviewable context without treating agent output as automatic truth, approval, or completion.
It is a local-first ecosystem for agent memory, evidence, evaluation, governance, and human-controlled workflows.
Current status
Early, real, and intentionally bounded
Active work launch
No active work launch implied
Source-ingest work
Parked and not ready
External harnesses
Research references only
AI work needs more than agents.
Useful AI-assisted work still needs memory, evidence, evaluation, governance, and human authority. Mythica explores how those pieces can stay visible and reviewable over time.
Runtime evidence
Design workflows where observations, helper outputs, and run evidence remain distinct from approval or authority.
Governed workflows
Build around human/operator decision points instead of treating agent output as automatic action.
The operating loop
Mythica organizes AI-assisted work around a simple loop: Work -> Evidence -> Provenance -> Evaluation -> Governance -> Human Decision.
Work
Humans, tools, scripts, and agents create outputs.
Evidence
Artifacts, traces, claims, and files can be captured for review.
Provenance
The origin and movement of information stays visible.
Evaluation
Behavior can be compared for quality, drift, and regression.
Governance
Evidence is reviewed within explicit authority boundaries.
Human decision
People retain authority over approval, truth, and next action.
Why it exists
AI work spreads across chats, tools, repositories, documents, logs, model outputs, and human decisions. Mythica is about keeping that work understandable instead of letting context disappear.
Disconnected sessions
Long-running work needs context and provenance that survive individual conversations and tools.
Unclear evidence
Claims, artifacts, validation posture, and approval state need to remain separate.
Project drift
Prompts, agents, branches, and documents need lineage and decision boundaries.
Ecosystem layers
The ecosystem is organized as separate layers so memory, evidence, evaluation, governance, and operator control do not collapse into one unsafe automation surface.
Mythica Context Intelligence
The context intelligence layer is the memory and provenance direction for structured, reviewable context. Current source-ingest work is parked and not ready.
Mythica Operator Stack
The operator and governance layer is the direction for keeping evidence, review posture, approval state, and durable decisions separate.
Studio
Studio is the operator workspace direction for viewing context, preparing handoffs, and reviewing artifacts without turning display into authority.
Evaluator
Evaluator is a proposed future layer for comparison, regression detection, and reviewable evaluation reports.
Evaluation research lab
The evaluation research lab direction informs reproducible experiments, evidence posture, and regression-style review.
Legacy Mythica
Legacy Mythica preserves the infrastructure, creative AI, local compute, and experimentation history that shaped the current system.
Current progress
Recent progress is about making the ecosystem harder to overstate: recording what is mainline, what is parked, and what is still research.
Source-bounded progress
Mythica now has bounded internal helpers and review-side building blocks, while keeping readiness and authority claims out of scope.
- MCI has count-only and metadata-only review-side building blocks.
- Source-ingest work remains parked and not ready.
- Technical boundary design exists as a design direction, not implemented enforcement.
- PI and NemoClaw/Hermes harness research remain parked as references, not selected tools.
Research references
Some work is useful as design input without being selected, trusted, installed, or ready for operational use.
MCI boundaries
Count-only and metadata-only helpers support review-side analysis, but they do not make ingest ready.
Technical boundaries
Current design work frames the difference between policy instructions and tested technical boundaries.
Harness patterns
PI and NemoClaw/Hermes research inform future thinking without implying trust, install, or adoption.
Trust boundaries
Evidence movement is not authority movement. Mythica separates what happened from what was approved.
- What happened.
- What was observed.
- What evidence exists.
- What was evaluated, reviewed, or approved.
- Memory ingestion creates durable truth.
- Visibility means authority.
- Task execution means completion.
- Evaluation scores replace review.
- Agent claims are the same as evidence.
What is not ready yet
Clear limits are part of the project. Mythica should be understandable without implying it is more mature than it is.
- Mythica is not a finished product or production-ready governance platform.
- Source-ingest work is not ready, and real source expansion is not authorized by these sources.
- Technical boundary design is not implemented enforcement.
- External harnesses are not selected, trusted, installed, approved, or install-ready.
- No active work launch, validation, approval, production readiness, or formal acceptance is implied.
- No timeline or launch promise is made here.
Recent progress
Recent work adds evaluator and harness research to the current Mythica picture while keeping pre-alpha boundaries around readiness, trust, and authority.
Early synthetic review signals
Evaluator work is still early synthetic and local. It is focused on review-assist signals such as false completion claims, missing evidence, unknown preservation, unclear handoffs, and safer review loops.
Sandbox and trust requirements
Harness research is clarifying sandbox and trust requirements before any package install, runtime execution, external tool adoption, package trust, or harness trust is considered.
Technical boundary preflight
A technical boundary preflight exists as a non-destructive review-only test candidate. It is not implemented enforcement and does not create runtime authority.
No readiness claim
This is not production evaluator readiness, benchmark readiness, SDK or package readiness, Docker/Kubernetes readiness, validation, approval, active work launch, formal acceptance, or authority movement.
Current focus
Near-term focus remains on source-bounded review, safer wording, and clearer boundaries before deeper automation or runtime work.
Keep public claims source-bounded
Turn internal evidence into public copy without exposing private details or overstating readiness.
Refine source-ingest safety boundaries
Continue treating count-only and metadata-only work as bounded internal helpers until source-ingest work is explicitly selected and reviewed.
Study technical boundaries carefully
Use design and non-destructive preflight work before any technical permission or runtime changes.
Legacy and evolution
Mythica grew from practical infrastructure, creative AI, local compute, and early research into evidence and authority across AI-assisted work.
Resilient infrastructure
Backup, edge, server, and operational ownership shaped the local-first bias.
Creative AI and local systems
Generated artifacts, visualization, local GPU work, and practical AI systems shaped the research path.
Governed agentic work
The current direction focuses on reviewable context, evidence boundaries, and human authority.