Runtime, not retrospect.
Agents and pipelines read from synchronised caches and produce confidence-weighted output. The categorical surface, governed in production.
What it is.
For dimension management to work in a world where agents are the dominant producers of dimensional values, the canonical reference has to be readable at machine speed, available continuously, and consultable as part of the production flow rather than after the fact. CleanDims provides the runtime layer that makes this possible.
Every agent or pipeline that produces or consumes dimensional values maintains a local cache, synchronised continuously to the canonical reference. Classifications happen against the cache, at memory speed, without a round trip to a remote system. The cache knows two things about itself: which version of the reference it holds, and the elapsed time since it last successfully synchronised.
This is not data quality monitoring, which detects problems after they have entered the data. CleanDims operates inside the production flow, at the moment the value is being produced.
Staleness-aware confidence.
An agent classifying from a current cache produces full-confidence output without a warning. An agent classifying from a cache that is one or more versions behind, or that has been unable to sync for longer than its configured tolerance, produces output with a confidence penalty. The penalty is reflected in the same visible-warning mechanism that surfaces unmapped values to consumers.
This collapses three failure modes (fail-open, fail-closed, fail-soft) into one architectural pattern. There is one regime: the agent always produces output, against the freshest cache available, with a confidence signal that reflects cache freshness. The agent never halts. The agent's job is to produce; the consumer's job is to decide what to do with output that is not fully fresh.
Consumer-side policy.
Different consumers of the same agent's output have different tolerances for staleness. A regulatory consumer of a vendor classification halts at zero tolerance. An analytical consumer of the same classification tolerates hours of staleness. CleanDims surfaces the confidence signal; consumers configure policies that determine when to act, when to warn, and when to halt.
This is the locus shift that the target state describes: the halt decision moves from the agent (which lacks context about how the output will be used) to the consumer (which has the context that matters). Different consumers configure different policies; the same agent serves all of them.
A finance agent classifying expense reports.
A finance agent classifies expense reports into spending categories as employees submit them. The agent reads its category dimension from a local cache synchronised to the canonical reference. When an employee submits an expense for a vendor the agent has not seen before, the agent's first action is to attempt to resolve the vendor name against the cached aliases.
If the match is high-confidence, the resolution is silent. If the match is low-confidence, the agent classifies anyway, attaches a warning to its output, and a request is opened on the steward's queue. The expense report flows through to approval, with the warning surfaced to the approver. The approver decides whether to act on the warning or proceed; in parallel, the steward sees the new variant in their queue and resolves it, either as an alias to an existing canonical or as a new canonical entry. The next expense from that vendor resolves silently.