Large Universe Models for software teams

What it ingests

Commits, pull requests and review comments, CI results and flake history, incident timelines and postmortems, issue trackers, deployment and rollback events, funnel and retention metrics, support ticket volume and content, and calendars.

Ship dates as posteriors

A date on a roadmap is an estimate with uncertainty that has been rounded to a point and then defended socially. The estimate stops updating the moment it is committed to, because updating it is a conversation nobody wants to have.

A Large Universe Model holds the date as a distribution and moves it on evidence:

• Merge velocity on the epic falls for two consecutive weeks
• A dependency's own estimate slips
• The flake rate in the affected suite doubles, which historically precedes a slip
• Scope is cut in a thread, tightening the distribution

The output is not a more optimistic or pessimistic date. It is a date that moves when the commits say so rather than when the standup does, with an interval that widens honestly when the evidence is thin.

Divergence detection

The most useful single behaviour is noticing when independent sources stop agreeing.

The roadmap says the migration is on track. The incident channel shows three rollbacks this week touching the migration path. The funnel shows a conversion dip in exactly the cohort routed through the new service.

Each of those lives in a different system, owned by a different team, and each looks unremarkable alone. A Large Universe Model holds beliefs fed by all three and surfaces the contradiction, which is usually the first observable sign of a problem that surfaces socially three weeks later.

Onboarding as a query against belief state

Ask a Large Universe Model for the state of a service and the answer is current as of this morning, with sources: what it does, what it depends on, what broke recently and why, which parts are under active change, who last touched them, and which of the documented assumptions no longer hold.

A Large Language Model can only summarise the documentation, which is wrong in the specific ways documentation is always wrong. Retrieval fetches the same stale documents faster. The belief state is the only representation that was maintained rather than written once.

Alerting on broken confidence

Conventional monitoring alerts on thresholds. A Large Universe Model alerts on a confident belief breaking — which is a much rarer and much more informative event.

If it has held for months that a service's p99 latency sits in a narrow band, and the band breaks, that is worth an alert even if the absolute number is within the configured threshold. Conversely, a threshold crossing during a deployment the model already expected to cause one is not news.

Cohort drift

Churn is visible in a dashboard after it has happened. The behavioural precursors — session depth falling, a feature quietly abandoned, support contacts rising in tone — arrive continuously and individually mean nothing. A maintained belief per cohort catches the drift while it is still cheap to reverse.

Limits

None of this substitutes for a team that talks to each other. A Large Universe Model surfaces divergence; deciding what to do about it is a human judgement involving priorities the model does not hold. The failure mode to design against is a system that produces more signals than anyone reads — which is why revision magnitude, not threshold crossing, has to be the trigger.