The map is free now. The trust isn't.

On July 1st LangChain shipped OpenWiki: an open-source agent that loops over a repository, writes a wiki for it, and keeps the wiki updated from git diffs. It’s a tool I’ve wanted for years. A simple agent that maps an entire codebase, every module, every contract, every seam, as the first step toward agentic work you can trust.

Credit first. OpenWiki is the right substrate: MIT-licensed, self-hosted, genuinely diff-driven. It records the git head of its last run, reads what changed since, and opens a daily pull request instead of pushing blind. And it’s built for agents. It wires a reference to the wiki into your AGENTS.md or CLAUDE.md, so the coding agents in your repo consult the map while they work.

Then I read the code.

A machine-generated map of a codebase is honest exactly once: at generation. It claims only what it saw. If documentation with no provenance is worth zero trust, an honest machine map is worth slightly above zero. Call it ε: a starting coordinate. The question is what makes the number rise, and what makes it fall.

OpenWiki has no number. To be fair to week-one software, the prompt does real grounding work: it forbids inventing files and APIs, demands claims be grounded in inspected sources, asks for inline source references in the finished pages. What it lacks is any record of standing. No confidence field, no reviewer identity, no timestamp on a claim. Nothing separates a claim verified against the current commit from one inherited twenty commits ago. The per-page evidence plan the agent drafts for itself gets deleted before each run completes. What survives are prose references with no binding to the commit they were checked against. In a calibration lab, that’s an instrument without a certificate: the readings may be fine, but nothing on record says when it was last checked, by whom, against what.

Meanwhile the wiki’s authority is installed at full. Wiring into AGENTS.md means every agent in the repo consumes it as ground truth: not ε, 1. Follow the loop. Agents read the wiki. Agents write code. The nightly job reads the diff and rewrites the wiki. The agents read it again. A map that redraws the territory, with no external reference anywhere in the cycle. The one human touchpoint is merging a daily PR, a yes or no on a markdown diff. Every product in this category leans on that same flat checkpoint.

The failure mode is already on record, from a product with a worse posture. DeepWiki, Cognition’s hosted take on the same idea, indexes public repos without the owner’s involvement and publishes with no review gate. It indexed Bo Lopker’s project and presented an unshipped, non-functional VS Code extension as the “primary installation method.” His verdict after reviewing his own project’s page: “it gets too much confidently wrong to be useful.” OpenWiki’s opt-in, PR-gated design is a real improvement. But a merge button carries no memory, and an unvalidated claim that passes through it still lands in front of your agents as ground truth.

I went looking for anyone in the category who built the missing axis. Swimm comes closest to half of it: auto-sync scores each code change against multiple signals, and when the signal is weak it refuses to guess and queues a human. But as far as I can tell it re-asks the same binary question forever. Nothing accumulates across five hundred correct answers. Mintlify drafts updates and keeps a human gate before every publish, the same flat checkpoint again. Nobody ships the actual requirement: a per-claim trust score that decays when the code under it moves, decays again with plain time, and rises only when someone who knows signs off.

Older industries stopped treating this as a feature a century ago. A calibration comes due on a calendar clock, not a usage odometer. Confidence in an instrument lapses with time even when nobody touched it. Recalibration intervals are computed from measured drift, not habit. A lab that fails a round-robin gets a documented corrective path, and accreditors require director-level sign-off before testing resumes. Trust in every mature measurement discipline is a maintained quantity: built slowly, spent by change, restored by signature.

OpenWiki made the map free. That’s good news, and it’s a fair defense that this is week-one software facing a problem nobody in the category has solved either. But the map can’t tell an agent the one thing it needs most: how much of it to believe, today, after the last twenty commits. That’s not a documentation problem. It’s a trust problem. Disclosure, so you can weigh this paragraph: I maintain Recede, an open trust protocol for agentic work, and nobody has shipped the per-claim answer for generated docs yet, me included.

The map was the fast half.