ChronoGuard

ChronoGuard runs an agent as if it were working at a chosen past date, then measures how well that blinding actually held. It is domain-agnostic, tied to no agent framework, and aimed at local models served by Ollama.

Reasoning as of a past date fails in two ways that look identical from outside. The first is that a tool returns a document published after the as-of date, and filtering handles that. Every tool’s output is normalized into one record type carrying a publication timestamp, and a decorator puts the guard in front of any Python callable that returns evidence. Undated and unparseable records are dropped by default, the boundary is exclusive so day-precision corpora cannot smuggle in a day of hindsight, and the agent never learns what was withheld.

The second way is the weights, which already encode facts from after the date. No filter reaches that, so the only honest move is to measure it. A probe asks questions that were only answerable after the as-of date, with no tools at all, and scores a control group next to them so a model that simply cannot answer anything does not read as blinded. A claim classifier then checks the final answer against the evidence the agent actually received.

chronoguard report "When will Halden ship Meridian, and what will it cost per seat?"

The verdict takes the worst of the three measurements and carries its reasons with it. Nothing can push it down, and a run on a model whose training postdates the simulated date never reports as low risk however well the filter held.