Training now builds the conversion label from the success date configured on your success criteria, instead of inferring an anchor from the record. If your definition of a conversion depends on a specific date field, the label your next model trains on matches it.
The marketing site and the product run on different hosts, and the app’s session cookie lives on the app’s host, so ax1om.ai could not tell a signed-in customer from a first-time visitor. Every docs page pitched everyone the same way.
The app now sets one first-party cookie, ax1om_lifecycle, on .ax1om.ai at
login and whenever your plan changes. Its value is a single word from a closed
set: trial, free, pro or enterprise. There is no identifier in it, no
email address, and nothing that could be joined back to a person. It is set for
30 days with SameSite=Lax and Secure, it is cleared at every sign-out, and
the app never reads it back. It is not written while an admin is acting as a
customer, so an impersonated session does not publish the target org’s stage on
the admin’s browser.
ax1om.ai reads it to pick which pair of links a docs page shows. A visitor is offered a way to start; a Pro reader is offered the app and the reference instead of a signup form; a Free reader on the scoring page is told plainly that live scoring is on Pro. The absence of the cookie means visitor, and an unrecognised value is treated as one.
The tool reference on the docs site is generated from the MCP server’s own tool manifest, and the build fails in either direction if the two disagree. An agent developer reading the page is reading what the server actually exposes, not a description of it that was accurate once.
Proposed nuances can now be confirmed or declined by an org owner or admin. Only a confirmed entry changes a number, and an agent never confirms.
An org owner or admin can now let an agent read row-level records. Absent means off, and revoking takes effect on the next request.
The hourly rescore skipped its score-level pass because the job image did not carry the setting the API pins. It runs.
Every training run now holds out later records rather than a random sample, so validation reads as a forward test.
All 31 tool reference pages now carry the tool's parameters and return shape, generated from the server's own manifest rather than written by hand.
Training runs recorded no split kind, so the per-rule readout fell back to a random holdout. New runs record it; models trained before this do not.
Every act on a model returns its validation state, and an org chooses whether a not-passed model is refused or proceeds with the state attached.
The model wizard advances the path per stage, Back and Forward move between stages, and score create and data sources carry their view in the query string.
A new MCP tool returns the columns a workspace holds, so an agent can plan a call instead of discovering the shape by failing one.