The Atelier Lab

The Atelier Lab · research in the open

A private workbench for thinking with machines.

Atelier turns a local computer into a dependable research partner — one place to read evidence, ask grounded questions, change code carefully, and learn what actually works.

Explore the system View the repository
Local-firstModel-agnosticEvidence before fluency
atelier / local sessionLOCAL_ONLY
# one loop, from source to confidence
atelier › ingest ~/papers/attention.pdf
✓ extracted · characterized · indexed 42 pages
atelier › ask --show-context
"What is the strongest limitation?"
✓ 8 passages retrieved · citations attached
atelier › verify --artifact claim.json
✓ evidence recorded · certificate ready
3 operations0 uploads1 audit trail
01 The idea

The useful unit is not a chatbot. It is a research loop.

Most AI interfaces begin with a blank prompt. Atelier begins with your approved workspace, your sources, and a clear boundary around what the system is allowed to do.

It is the operational center of The Atelier Lab: a local CLI, a replaceable web surface, and a shared service for documents, repositories, models, memory, and deterministic tools. The user stays in the loop; the system makes evidence, provenance, and verification easier to inspect.

01Retrieve first.02Reason with context.03Verify before you trust it.
02 How it works

From a file to a defensible answer.

Every request passes through explicit stages, so a fluent response never has to stand in for evidence.

characterize → route → retrieve → reason → verify
01◎

Characterize

Parse the artifact first. Detect its shape, source, sections, figures, schemas, and what kind of work it needs.

deterministic foundationRead more ↗ 02↗

Route

Choose the cheapest capable model and workflow based on domain, modality, privacy, context, and difficulty.

capability-firstRead more ↗ 03⌕

Retrieve

Combine semantic and lexical search with section-aware ranking, diversity, and citations.

hybrid evidenceRead more ↗ 04∴

Reason

Use local models, guarded tools, and durable project context to answer, build, compare, or explore.

human-directedRead more ↗ 05✓

Verify

Run tests, inspect diffs, check feasibility, preserve provenance, and leave an artifact another person can inspect.

confidence with receiptsRead more ↗
The shared layer

One workbench. Many interfaces.

CLI, web, and Finder actions all talk to the same local service. Your workspace and privacy policy travel with the task.

Read the workbench plan
03 What is already here

A small system with serious edges.

The baseline is intentionally practical: enough surface area to do real work, enough structure to measure it.

v1.0 verified locally
A / 01

Knowledge mode

Ingest papers, DOCX, PPTX, spreadsheets, notes, images, EPUBs, archives, and source code into a local library.

User guide
A / 02

Research mode

Characterize papers, retrieve grounded passages, inspect figures, compare sources, and keep the citations attached.

Research method
A / 03

Build mode

Inspect repositories, make constrained edits, run targeted tests, review the diff, and produce a verification certificate.

Experiment standard
A / 04

Scientific tools

Inspect and simulate small quantum circuits, validate optimization candidates, and keep deterministic checks close to the reasoning.

Current architecture
A / 05

Memory

Store explicit project facts, task state, artifacts, decisions, provenance, and expiry without persisting conversations by default.

Memory model
A / 06

Privacy boundary

Approved workspaces, capability-scoped tools, LOCAL_ONLY by default, secret redaction, and explicit network escalation.

Limitations
04 Evidence, not theatre

Measure the loop. Record the misses.

The lab treats negative results and hard limits as useful output. These are small frozen suites — signals to investigate, not claims of general intelligence.

Browse the benchmarks
Knowledge answers17 / 18 correct
Code tasks13 / 13 solved
Combined tasks10 / 10 solved
Router lift43.8% → 100%
Frozen local suites · measured on Apple Silicon · see the repository for scope and caveats.
39.3tok/sqwen3:4b decode
223chunksverified three-paper index
36GBunified memory constraint
$0clouddefault local path
05 What happens next

Close the loops, then push the frontier.

The roadmap is a practice system: every phase ends in an artifact, a measured result, or an honest failure analysis.

Open the living roadmap
Now · Phase A

Deployment mastery

Close the current experiments and make the local serving trade-offs legible.

  • Run the 20k-step foundation experiment
  • Measure quantization and batch-size curves
  • Stand up a reproducible serving benchmark
Exit bar · explain the throughput / latency curve from your own numbers.
Next · Phase B

Architecture depth

Build the model from first principles, then test each design choice.

  • Annotated transformer from scratch
  • RoPE, RMSNorm, SwiGLU, and GQA ablations
  • Mini scaling-law study on the laptop
Exit bar · understand the deltas, not just the vocabulary.
Later · Phases C–E

Original contribution

Move from learning the stack to asking a question the stack can answer.

  • SFT → DPO → GRPO on a small model
  • Kernels, efficiency, MoE, and parallelism
  • Publish a reproducible open result
Exit bar · a stranger can rerun it.

Open work · careful systems · local evidence

Build a better relationship with the machine.

Read the source, run the experiments, and follow the work as it gets more capable — without becoming less inspectable.

Explore on GitHub Read the project README