REWIREby AIFLOWLABS

The workbench

Four parts, one unit of work.

Every route to change a model, whether an agent, a closed-form solver, a generator, a compiler or a merge, produces the same artefact: an Implant. A sparse diff plus a manifest. The Atlas shows it, the Verdict scores it, the registry keeps it. That is the whole design rule.

AtlasdeterministicScribemodelImplantdeterministicVerdictdeterministicRegistrydeterministicfind the neuron → write the weight → prove the change
  1. 01

    Scribe · The editor

    An AI agent with a Python sandbox, read access to tensors, a budget of forward passes and no network. Open-weight agent models first; closed frontier models only as a labelled upper-bound arm. The Scribe proposes an edit and writes exactly one diff file for one named target.

    Prompts for benchmark runs carry no hints about the failure mode and are pre-registered. Every run records reasoning tokens, tool calls, forward passes and wall-clock: the compute ledger.

  2. 02

    Atlas · The map

    A neuron-level viewer: a 3D map of the model’s layers, every neuron ranked against random-direction controls, steered rollouts per dose, a three-column chat (−, 0, +), lenses and attribution graphs. For small compiled networks, a circuit mode that shows the actual graph.

    For each Implant the Atlas computes which neurons moved, highlights them, and shows base versus edited versus edited-with-those-neurons-clamped. Its tools are exposed to the Scribe over MCP so the agent can look before it writes.

  3. 03

    Implant · The unit of work

    A sparse diff in a standard tensor format and a manifest: base-model id and hash, engine, tensors touched, sparsity, stated intent, acceptance tests written before the edit, evaluation references, licence, inverse diff and a signature. Parent edit and stack position for versioning.

    Design rule: the diff is the unit of work. Compile, generate, solve, agent-edit and merge all output the same artefact, so the Atlas and the Verdict treat them alike and every change can be reverted.

  4. 04

    Verdict · The proof

    An evaluation loop in which a generator writes tasks, a verifier checks them, the model under test is scored by code, and tasks are kept in the band where they are solved 25 to 50 % of the time. Three instances: one for editors, one for the compiler, one for the Atlas itself.

    Mandatory controls: base model, base with thinking off, steering baseline, LoRA of equal size, matched-sparsity random edit, random-direction control, a negative-control model that does not have the failure, a held-out split, a regression set, a safety battery, bootstrap confidence intervals.

layer 15layer 190.016 % of weights · one direction subtracted

What an Implant manifest looks like

A design sketch, not a released format. Field names will change at M1.

{
  "edit_id": "implant-2026-0001",
  "base_model": { "hf_id": "Qwen/Qwen3.5-0.8B", "revision": "<git sha>", "sha256": "<tensor hash>" },
  "engine": "agent | solver | generator | direction | compiler | merge",
  "agent": { "model_id": "...", "forward_pass_budget": 50, "passes_used": 50, "hinted": false,
             "reasoning_tokens": 0, "tool_calls": 0, "task_novelty": "memorised | novel | unknown" },
  "approval": { "target_model_ne_editor_model": true, "human_gate": "required-for-promotion", "approved_by": null },
  "diff": { "format": "safetensors-sparse", "tensors": ["layers.15.mlp.down_proj.weight[:, idx]"], "l0_fraction": 0.00016 },
  "intent": "one human-readable paragraph, written before the edit",
  "acceptance_tests": "machine-checkable, written before the edit",
  "evals": { "frontier_run": "...", "regression": "pass | fail", "safety": "pass | fail" },
  "parent_edit": null, "revert": { "inverse_diff": "...", "bit_exact_verified": false },
  "license": { "edit": "CC-BY-4.0", "base": "Apache-2.0" },
  "signature": "..."
}

Open and low-cost by choice

LayerChoiceWhy
Primary small modelQwen3.5-0.8BThe published edit targets it; the diff and apply script are public.
Atlas modelQwen3.5-4Bneuron-explorer’s native target; steering neurons already mapped.
Mid modelQwen3-8BA 128-neuron persona precedent exists on it.
Negative controlGemma 3 270MDoes not loop; must not “improve”.
LibrariesPyTorch, safetensors, PEFT, nnsight, TransformerLens, EasyEdit, pyvene, SAELensAll MIT or Apache-2.0. mergekit (LGPL) is used as a tool.
ComputeLaptop or CPU for 0.8B edits; pay-as-you-go GPU pods for sweeps; rented cluster time for 8B+The replication sprint has a zero-cash path.

The diff is the unit of work. Everything else is there to make it visible, measurable and reversible.

Join the replication sprintWatch the walkthrough