ReferentGround

AI agents now generate code at a scale humans cannot review line by line. Approving ten-thousand-line pull requests by gut feel is how production incidents, security holes, and compliance failures get shipped. The reviewer has the responsibility but no longer has the surface area—and after the merge, there is no record of what was actually understood when the approval happened.

ReferentGround is an attempt at both problems. It reads a codebase deterministically and produces a versioned graph of what is actually there—capabilities, paths, boundaries, ownership—so the unit of review can move from lines to what changed in the system, and so the judgment itself becomes evidence: replayable, diffable, tied to the revision it was made against.

What it explores

• Truth before explanation. A static, deterministic extraction layer first. Anything an LLM contributes is grounded in code that was already identified by the parsers—never invented.
• One substrate, two surfaces. A graph for humans to navigate, and bounded context packs for agents to consume. Both are projections of the same versioned snapshot, so they cannot drift apart.
• Evidence by construction. Provenance and quality gates run through everything: a context pack, a capability path, an AI-assisted change—each can be replayed, diffed, and validated across repository revisions. Token budgets are hard ceilings, not aspirations.
• Capability-level diffs. The point of the graph is the diff. Did checkout suddenly start calling billing directly? Did the auth boundary move in this PR? Which paths touch sensitive data this week that didn’t last week?
• Governance, not vibes. Reviewable capability diffs, private evals, and institutional memory—a learning layer, so an organisation’s judgment about AI-written change accumulates instead of evaporating with each review.

Why it matters

Most AI coding failures are context failures. Agents are given partial, lossy, stale context and then asked to make changes with incomplete understanding. Most reviewer failures are the inverse: too much context, presented at the wrong altitude. Both problems share a substrate.

The bet is that the codebase itself—parsed, versioned, made navigable—is a more honest source of truth about a running system than any document, diagram, or runtime dashboard. And that when the question arrives, as it will, of what did we know when we shipped this—the answer should be a replayable artifact, not a recollection.

Status

In active development—a TypeScript/Bun CLI today, a cloud platform forming behind it. Local-first by design. Public details will follow as the work matures; the underlying ideas show up in the writing.