Threadout

Evidence-backed commitment integrity — find what slipped between the conversation and the merge.

What failed

Something gets agreed to on a call. Then it either reaches the work or it does not, and nothing is watching that seam.

Four shapes, and only the first one looks like forgetting. Dropped — agreed aloud, never tracked. Buried — captured, then displaced by something louder. Intentionally parked — a real decision to defer, but the reason was never written down, so the same argument gets had again next quarter. Delivered differently — shipped, but not as decided.

What I observed

Retrospectives catch these late, and only when someone happens to remember what was originally promised. That makes human recall the instrument. Recall is the thing that failed in the first place.

What I built

Threadout reconciles one initiative across a conversation export, a project tracker, and GitHub, and produces a short list of gaps with the evidence that produced each one.

The design positions are all refusals.

• The product shell stores no evidence. It is a thin multi-tenant surface in front of Conduit, which stays the private evidence and detection authority. One place a customer’s evidence lives, so one place to delete it from.
• Candidates are bounded before they are scored — by project, by time window, by who was involved. Missing data narrows the bounding rather than forking the logic.
• Ambiguous is its own outcome. Resolution is three-way: resolved, unresolved, ambiguous. Ambiguous is never quietly relabelled into one of the other two, because that is exactly where a system this shape earns false confidence.
• A finding with no cited evidence never ships. Below the confidence threshold, a finding appears as a held count and nothing more.
• A human review the publish path cannot bypass. Nothing goes out without a durable record binding one authorized person’s approval to the exact version of the content they approved. The guard denies with no side effects.

The governing rule, which decides most of the smaller arguments: if it is not short and mostly right, it does not get sent.

What the evidence showed

The parts that are real are real. The resolver is working code — bounding, the three-way outcome, the confidence gate, and a strict timestamp parser that rejects an instant it cannot prove rather than passing a rounded one. The control plane is applied and tested: tenant isolation, the job lifecycle, and the review receipt above. The public site is live and the demo runs.

The parts that are not, are not, and this is the part I will not smooth over. There are no live connectors. The conversation export, the tracker, and GitHub exist as a typed contract and honest copy, not as running integrations. Meeting transcription is an optional expansion and it is not connected either. The demo is synthetic — a fictional corpus I wrote, resolved in the browser. No real conversation has been through this system.

So the resolver has never been wrong about a real commitment, because it has never seen one. Everything above that sentence is engineering I can show. None of it is evidence that the idea works.

What comes next

Connect one real initiative, run the weekly list against it, and publish whether it held — including the count of findings a human threw out. That number is the only one that settles anything.