How it works

GRIDLINE is not an agent and not a model. It is the layer that decides what work is admitted, who reviews it, what it may cost, what may leave, and what may be written — and then coordinates the agents and local models that do the work.

Where it sits

Your agent runtimes and your inference substrate already exist. GRIDLINE sits above both and governs them; it does not replace either.

            ┌──────────────────────────────────────────────────────┐
            │  GRIDLINE — the control plane                        │
            │  what work is admitted · who reviews it · what it    │
            │  may cost · what may leave · what may be written     │
            └───────────────┬──────────────────┬───────────────────┘
                            │                  │
      ┌─────────────────────▼───┐    ┌─────────▼──────────────────┐
      │ agent runtimes          │    │ inference substrate        │
      │ Claude Code · Codex ·   │    │ Ollama on your nodes       │
      │ Hermes (in a sandbox)   │    │ (a router can sit here)    │
      └─────────────────────────┘    └────────────────────────────┘

The review, and the context cut

The point of a second opinion is that it is independent. GRIDLINE enforces that mechanically.

Every finding is re-judged by a model of a different family, in a fresh context that never saw the author's reasoning — measured at 27 pairings with zero same-family among them. A reviewer that inherits the author's chain of thought is not a reviewer; it is an echo. The context cut is what keeps the disagreement real.

Three 4–8B models matched one 30B on defect recall (0.708), at 19× less author time — and the two arms found different defects, so pooling their candidates recalled more than either alone (union 0.875). Diversity pays when you pool candidates and costs when you pool decisions.

And the verdict layer refuses until it is earned: a judge that cannot tell a real defect from a false one is not admitted, and today every local model this project measured is refused. A degenerate judge is caught in the act — one answered REJECT on 35 of 36 cases. The records are here.

The fleet

A control node, worker nodes, and — optionally — a memory node, across the machines you already own.

GRIDLINE has run a three-node fleet on a 10 GB desktop, a 6 GB laptop, and a CPU-only box, spanning a private network with one node in another state. Routing fits a model to a node on the node's measured budget, not its nameplate — the effort that found a VRAM estimate 3.7× wrong. A model that will not fit is refused rather than swapped out mid-run.

command node — runs the control plane and the dashboard, bound to loopback.
worker node — hosts local models and runs review work.
memory node — holds the retrieval index; the indexed chunk is a class that may not leave.

Joining a second node is four steps, two of which write firewall rules, and those scripts are Windows-only today. The control plane itself runs on Linux, macOS, or Windows.

Read on

The boundary — what may leave the fleet, and what it may cost, at the mechanism level.
Evidence — every claim above, linked to the record that produced it.