A makerspace for privacy-first, local-first research & projects.

The hardened daily driver. Local-only AI workbench for research, writing, and software work — evidence-disciplined, replayable, no playful framing.

When you actually need the answer to be right. Every request flows through a typed pipeline engine, executed by a roster of specialist agents over scoped context-augmented memory. Serious Mode is on by default — research, build, and writing pipelines run with evidence discipline. Image / video lanes have been stripped out on purpose. No API keys. No cloud. No frontier model calls. Ever.

• Typed pipelines — research / build / code-fix flows where every stage emits against an OutputContract. Contract failures surface as findings, never silently propagate as fabrication.
• Specialist roster — planner · researcher · auditor · skeptic · synthesizer · verifier · memory_critic. Skeptic sidecar writes its rationale to a separate critique file.
• Scoped CAG memory — domain → topic → thread → project → run, with promotion gates, contradiction detection, and an append-only decision ledger.
• Replayable audits — every run resumable from any node; regression harness flags semantic drift across replays via embedding cosine.

research_pipeline
  intake → domain_framing → source_discovery
        → evidence_analysis → nuance_pass
        → synthesis → cag_promotion_gate

build_pipeline
  requirements → architecture → implementation_plan
              → patch_artifact_generation → verification

code_fix_pipeline
  planner → code_localizer → patch_writer
         → reviewer → test_fixer → finalizer

// every stage emits against a typed OutputContract.
// contract failures surface as findings — never
// silently propagate as fabrication.

A Hermes-first operational AI substrate with filesystem-grounded memory, governance, and recoverable continuity.

Not a chatbot. Not a SaaS wrapper. The governed operating layer that AI workloads run inside. OverCR provides identity, doctrine, boot continuity, subagent routing, packet validation, workflow choreography, and approval-aware operational structure. Operational Vigilance, Execution, Recovery, Command & Routing.

• Filesystem truth — canonical state lives on disk. Chat history is ephemeral.
• Untrusted by default — model output passes L1–L6 validation before any state advancement.
• Bounded workflows — subagents never route directly to each other; OverCR is sovereign.
• Approval gates — no autonomous outbound contact; no autonomous filesystem mutation. Ever.

core_guarantees
  ├─ filesystem truth is authoritative
  │    chat history is ephemeral · disk is canonical
  │
  ├─ model output is untrusted
  │    must be sanitized + validated before
  │    any state advancement
  │
  ├─ overcr is sovereign
  │    subagents never route to each other;
  │    all packets flow through the substrate
  │
  ├─ no autonomous outbound contact
  │    external action requires operator approval
  │
  └─ no autonomous filesystem mutation
       workers produce plans. plans need a human.

Oathweaver's wilder sibling. Same lineage, looser rules — image generation, image-to-video, creative writing, unrestricted topics, more lanes, more breakage.

The fun workbench. Where Oathweaver runs hardened and disciplined, Foxforge runs loose — with creative pools, image generation via ComfyUI, image-to-video (Wan2.2 / SVD XT fallback), an unrestricted-topic lane, and a wider Make surface. More lanes means more moving parts, which means more rough edges. Treat it like a playground: fast iteration, novel outputs, the occasional broken corner. Not what you reach for when you need the answer to be right.

• Image & video lanes — ComfyUI integration for image generation (Pony XL / SD presets) and image-to-video via Wan2.2, with SVD XT as the lower-VRAM fallback.
• Creative Make pool — novels, memoirs, books, screenplays, with continuity-aware scene writers and a voice critic. Unrestricted-topic content routes through an abliterated model.
• Seven Make pools total — essay, longform, content, specialist, creative, web app, desktop app. Each its own multi-agent pipeline.
• Same engine, looser leash — LangGraph turn graph, four-persona research, citation linking. Just with the safeties dialed down and the toy box opened.

lanes  (same engine, looser leash)

  ┌──────────────┬──────────────┬──────────────┐
  │ Research     │ Make         │ Talk         │
  │ 4 personas, │ 7 pools —    │ Reynard      │
  │ handoffs,   │ essay /      │ dolphin3:8b  │
  │ citation    │ longform /   │ free-form    │
  │ linker      │ content /    │ conversation │
  │             │ specialist / │              │
  │             │ creative /   │              │
  │             │ web/desktop  │              │
  └──────────────┴──────────────┴──────────────┘

media  (the part Oathweaver dropped)
  │
  ├─ ComfyUI          image gen · Pony XL / SD
  ├─ Wan2.2           image-to-video (16+ GB VRAM)
  └─ SVD XT           image-to-video (4–6 GB VRAM)

// playground, not production.
// more lanes → more moving parts → more breakage.
// experimental phase — APIs and configs may change.

A series of prompts for producing deep, taste-oriented analyses of codebases..

Not code reviews. Not security audits. Each pass interrogates a different layer of how the author thinks — what they take to be obvious, where their judgment shows up, what their machinery actually enforces, what survives when lifted out of context. The series is designed to be run in order. Each pass is meant to be performed without knowing what later passes ask for. The diagnostic value of any single pass depends on the earlier passes being honest about their own blindness. Used to audit all Guideboard project releases.

• The First Read — The first impression — coverage-driven, opinionated, blind to its own blind spots
• The Discounted Artifact — The largest or most formal artifact the first pass under-read
• The Trace — Whether the repo's stated obligations are actually enforced by the running code
• The Twin — What the repo's choices look like next to one adjacent repo with a different mental model
• The Lift — What can be torn off the repo and survive on its own