- The Insights report stopped being a list of labels and became a measurement. Until yesterday the report could tell you that a conversation had a problem in it, because a teammate had written the question and a model had answered it against the transcript. What it could not tell you is whether that problem mattered. Every project now also carries observable outcomes, which are facts rather than opinions, read straight off what actually happened: the client left without ever running a search, the search produced nobody worth shortlisting, the client never picked anyone, the talent declined, or never replied, or replied and never sent an offer. For each label the system then compares how often an outcome happens when that label is present against how often it happens when it is absent, and reports the gap with a confidence interval around it and a letter grade for how much weight the finding deserves, built from three things kept separate: how big the effect is, how sure we can be it is not noise, and how much evidence it rests on. No model is ever asked what is to blame. The model only reads a conversation and applies a label; the arithmetic does the rest, which is the whole point, because a model asked "was this the reason" will always find you one. It also shows which labels travel together, and it reaches across agents, so a label on a client conversation can be weighed against what happened to the talent on the other side of it days later.
- A probe that asks how early we could have known. Knowing a conversation went wrong is worth much less than knowing when it first became visible. The probe replays a conversation as a growing prefix, the first message, then the first two, and so on, scoring each clip on its own, and gives back two numbers per label: how far into the conversation the sign first appears, and how much it flickers on and off as more of the conversation arrives. That separates a signal we could genuinely have acted on at the second message from one that only looks obvious once you already know the ending. It is deliberately expensive and deliberately separate, kept in its own store away from the cheap grading that runs on every report, and it runs as a proper background job so a big replay is not something anyone waits on.
- The list of questions gained version control. The set of checks is the instrument, and quietly changing the instrument halfway through invalidates everything measured with it. The whole set, the questions and the grades alike, can now be saved as a named snapshot, listed, compared and reverted to, and reverting never destroys anything: it restores the snapshot's questions and retires the rest rather than deleting them. Alongside it, a set can be exported whole and compared against another, which is what the new final testing tab is for: holding the current live set next to a proposed one and seeing exactly what changed before anything is measured with it.
- The assistant can now change the checks, not just suggest them, and it stopped forgetting the conversation. When it proposes new questions they arrive as a single message with a checkbox beside each one, all ticked, and one button to apply the ones you kept, instead of a scatter of separate cards to accept one at a time. It can now also edit or retire a check that already exists rather than only adding more. The conversation itself is now one continuous session rather than one per report, so it survives switching between reports or rerunning one, with a deliberate Reset as the only way to start fresh, and it is told when the report or the view in front of you has changed since it last heard from you. It also learned to describe itself honestly: asked what it can do, it answers with what it actually has rather than a plausible-sounding list, and says plainly what it cannot do yet.
- A retry that would have doubled every number it recorded. The new probe was the first job of its kind, and the rule that every background job must be explicitly assigned a lane, put in place the day before after six job types had been enrolled in a speed promise by accident, caught it on its very first run. It belongs on the slow lane with the other bounded model sweeps. The second half matters more for what the probe is for: a job on that lane can be retried, and the probe wrote its results by adding to them, so a retry after a partial run would have counted part of the work twice and quietly inflated the very numbers the whole thing exists to produce. It now clears a run's results before recording, so a retry rewrites rather than adds.
- The one thing deliberately not built. The obvious next step is to feed all of this back into what the agents actually say to clients, catching a problem while the conversation is still happening rather than in a report afterwards. That is exactly the kind of change that cannot be checked by ordinary tests, because it alters how a stochastic conversation unfolds, and the only thing that can check it is the live behavioural suite that costs about fifty-five dollars a run. So it was designed, written down, and left unbuilt on purpose rather than shipped on the assumption it would be fine.
A quiet Saturday, 4 commits, and all of it went into one thing: turning the Insights area from something that describes our conversations into something that measures them. Every project now carries plain observable outcomes alongside its labels, facts read off what actually happened rather than anything a model was asked to judge, and each label is scored against them statistically, with a confidence interval and a letter grade that keeps the size of the effect, the certainty and the amount of evidence behind it as three separate things. A new probe replays each conversation as a growing prefix to find out how early a problem was visible and how steady that signal is. The set of questions itself gained version control, because changing the instrument halfway through invalidates the measurement, and nothing is ever hard deleted. The assistant can now edit and retire checks rather than only propose them, applies a batch in one press, and keeps one continuous conversation across reports. And the day before's rule that every background job must declare its lane caught the new probe on its first run, along with a retry that would have counted part of its own results twice.