Technical paper · LIV
Validity Estimates for Facts in Enterprise Agent Memory.
A per-fact validity estimate learned from an organisation’s own verification history, tested against simple schedules on public change data and in simulation, with every hypothesis registered before the test.
Version 1.0.0 · 9 November 2026 · Apache-2.0 (text and code), (c) 2026 FlowX.AI
Preprint, not peer reviewed. Most results come from simulators under stated assumptions, not real-world rates; on public Wikidata change histories the learned model showed no gain over the simple baselines. CASS is the companion paper. Read CASS →
- agent memory
- data freshness
- survival analysis
- interval censoring
- re-verification
- Wikidata
- simulation benchmark
- pre-registration
Two checks, spent where they count
An agent acts on facts it stored weeks or months ago. LIV estimates, per fact, the chance that the value is still true, learned from the organisation’s own change and re-verification history, and spends a fixed checking budget where a stale value is likely and would cost most. The fields below are illustrative; the results are the paper’s.
Five stored facts, each with the time since it was last verified. A validity estimate appears per fact. A budget of two checks first goes to the two facts least likely to be still true; weighting by what a stale value would cost swaps a cheap fact for a high-consequence one. The checks return: one value has changed and is updated, one is confirmed. The panel then states the pre-registered results: in the tacitdrift simulator, 18.8 to 71.5 percent more stale facts found per unit of check cost than the better of per-class Poisson and a calendar schedule in all 8 pre-registered cells with heterogeneous entities and activity signals; on Wikidata statement lifespans, no improvement detected (−0.3 percent, interval −2.1 to 1.7).
Core thesis
In simulation, a validity model learned from an organisation's own verification history found more stale facts per check than simple change-rate schedules when per-fact signals existed;on public Wikidata change histories it did not.Weighting re-verification by what a stale fact would cost lowered cost per case, and the learned scorer added a further saving below the pre-specified worthwhile size.
Abstract
Agents that act on stored enterprise facts act on values that may have changed since they were last verified. We study a per-fact validity estimate with calibration as a target, learned from an organisation's own change and re-verification events, that gives the probability that a stored value has not changed since its last check and chooses which facts to re-verify under a fixed budget. The ingredients are established; we evaluate their combination for agent memory, with the per-class Poisson estimator of Cho and Garcia-Molina as the main baseline, on LIV-Bench: tacitdrift, a seeded simulator whose parameters encode stated assumptions rather than real-world rates, and a task built from Wikidata statement lifespans.
In the simulator, at an equal budget, a discrete-time hazard model finds 18.8 to 71.5 percent more stale facts per unit of check cost than the better of per-class Poisson and a calendar schedule in all 8 pre-registered cells with heterogeneous entities and activity signals, and does not beat per-class Poisson in the negative control. On Wikidata we detected no improvement (-0.3 percent, interval -2.1 to 1.7), also with co-change features added by amendment, and 4 other model-based baselines found more stale facts per check.
In a simulated loan-intake process, consequence-weighted re-verification lowered test-window cost per case by 6.8 percent against the better freshness-only policy, short of the pre-registered 10 percent, and raised failed actions. In a second pre-registered experiment with complete case follow-up and frozen models, an avoidability-aware allocation lowered total cost per enrolled case by 1.55 units (interval 1.33 to 1.77) against a freshness-only allocation that uses the same validity scores but not the rule and cost tables, at least the smallest worthwhile saving of 1 unit, without lowering completion; the pilot decisions before that test are disclosed. In a third pre-registered experiment in the same simulated process that changed only the scorer under that allocation, the learned validity model lowered total cost per enrolled case against per-class Poisson by 0.60 units (interval 0.53 to 0.68), less than the smallest worthwhile saving, without lowering completion.
Part of the FlowX.AI paper series.
Each paper names its method and the evidence behind it, including where the method did not pay off.