Ganymede

Rough sketch

Where things sit on the screen, and in what reading order. Drawn, not screenshotted, so it stays true when the styling moves on.

What happens, step by step

In plain language, in the order it happens.

1. 1

   Build the panel

   Loan-level servicing history becomes a monthly borrower panel. Calendar dates are the point: a timing feature from a source without real dates is refused at the feature layer rather than caught in review.

2. 2

   Score two questions, not one

   L1 asks whether the account worsens over the next ninety days. L2 asks whether it recovers without contact. An agent reads those numbers at face value, so calibration is the gate rather than AUC.

3. 3

   Rank by money, not by risk

   The allocator maximises expected recovered value per agent-minute: uplift over self-cure, weighted by exposure, under the capacity the team actually has. Probability never sorts the queue.

4. 4

   Coach inside the measured gap

   328 inter-turn gaps from a real call give a median of 479 ms and a p25 of 292 ms. The budget is 300 ms because that is what the distribution allows. A hint that misses the gap arrives after the moment it was for.

5. 5

   Log the decision, then resolve it

   Every score, hint and override is written down with its experiment arm and its propensity. When the payment arrives or does not, the promise resolves and both lenses retrain on the result.

Decisions, and what they cost

Every choice worth recording has a road not taken. Both are here.

• Rank by expected value

  instead ofRank by probability of default

  Risk-ranking calls a 1,289 euro account early and never reaches a 1.93 million euro one anywhere in the capacity sweep. Twelve agent-minutes cost more than the whole uplift on the small account is worth. At 15% capacity the value ordering recovers 59% more using roughly half the contacts. At 60% the edge falls to 2.4%, because with enough agents to call everyone the ordering stops mattering. The gain lives exactly where the constraint is real.

• Calibration as the release gate

  instead ofAUC as the release gate

  A ranking metric says the order is right. It says nothing about whether 0.7 means seventy percent. The agent reads the number and acts on its face value, so the number has to mean what it says.

• Two hint tiers split by a measured budget

  instead ofOne tier, with a latency target picked in advance

  75% of real turn boundaries are wide enough for a hint at 300 ms. Only 48% are wide enough for a 500 ms model call. Deterministic hints therefore land live, and model-composed strategy waits for the next pause instead of racing a budget it loses.

• Refuse four headline numbers

  instead ofEstimate them from simulation

  Recovery lift, override rate, promise-kept lift from coaching, and conversation features beating tabular features all need live data or a randomised slice. Each could have been estimated into something impressive. Each carries a pending badge instead. The evals module refuses to compute lift on synthetic records at all, because the generator's own priors would leak into the answer.

• Generate every figure on the site from the pipeline

  instead ofWrite the numbers into the page

  A build script regenerates the site data by calling the pipeline, and every value ships with the evidence behind it: measured, backtested, simulated, seeded or pending. A figure with no provenance cannot render. CI fails if the committed data drifts from what the code now produces.

• Fourteen review findings as runnable invariants

  instead ofA postmortem list

  A defect is closed when something automated fails if it comes back, never before. The queue producer is the only path to a queue. The experiment arm is non-nullable. The retrain aborts on a row with no propensity. A list in a document cannot do any of that.