Archived ideas

The Archive

Ideas that lived, then didn't. Each one is recorded: what it was, and the honest reason it was abandoned. Not a roadmap — a record.

Built and archived

EXP-005

Apr 2026 – Oct 2026 · was beta

What was built: A CAD-as-code CLI and Rust library. Two-dimensional geometry is declared in .cf TOML files (layers, arrays and mirrors, dimensions). cadspec serve previews it live in the browser; the tool validates and formats the source and compiles it to DXF, with PNG and SVG previews for AI agents. Shipped as 0.1.0 betas.
What we learned: Declared sources diff and review cleanly, and the same source always compiles to the same DXF. The hard part was not the format. It was everything around the drawing: the viewer, the validation, and the interop with the tools people already use.
Why it closed: Existing tools already cover this better. ezdxf does 2D DXF from code, CadQuery does 3D, OpenSCAD has its own scripts and libraries, and FreeCAD has an MCP server. cadSpec could not compete with them, so no more tokens go into it.
Use instead: ezdxf for 2D DXF. CadQuery for 3D parametric models. OpenSCAD for scripted solids. FreeCAD with its MCP server for agent-driven CAD.
Archived repository: github.com/UniverLab/cadspec

What it was: Decentralized compute network for science — CPU sharing across nodes, with defined roles and a minimal instruction language (LIC).
Why it died: AI gradient vectors cannot be cleanly split and distributed; the latency is a fundamental problem. The model that works for BOINC-style compute breaks for deep learning inference.

What it was: Personal memory system with spaced repetition (SM-2) and a pair of agents wired on top for review and generation.
Why it died: Cards generate friction when you're orchestrating agents — nobody stops to do flashcard review mid-session. The SM-2 attention loop and the agent work loop do not coexist well.

What it was: Skills that self-generate and evolve by observing the PTY stream of other agents in real time. Concept taken from Hermes Agent.
Why it died: An agent re-analyzing what other agents are doing is high cost and divides focus. Post-hoc, passive extraction might be feasible — but not while the agent is in the work loop.

What it was: RAG-style chunking applied to the genome: split sequences by biological meaning (gene boundaries, regulatory regions) rather than arbitrary fixed windows.
Why it died: Just a concept. No implementation, no dataset, no collaborator. Seeded here because it might germinate somewhere else.

What it was: Forest simulation for Canopy's idle terminal space — prey, predators, plants, and cellular-automaton rules to fill blank screen real estate.
Why it died: Dropped in favor of the Brian's Brain automaton, which is simpler and already implemented. wildterm made it to a spec. Somewhere in a branch.

What it was: A cross-platform mobile app for controlling Canopy from a phone — starting, stopping, and monitoring agents on the go.
Why it died: The attack surface is too wide. Exposing a daemon to the network — even on localhost — requires auth, encryption, and a considered threat model. None of that exists. Building the feature before the security layer is the wrong order.

What it was: A newsletter for lab updates and a metrics dashboard tracking downloads per CLI.
Why it died: Infrastructure cost and no good fit. Cloudflare Workers has no SMTP, so email delivery would need a third-party relay — a dependency for low-signal traffic. The download count can wait until the tools are worth counting.

The record is kept because what dies here may live elsewhere.