KAIROS: keeping an uncertain timeline honest

A timeline contains several kinds of time

In a fictional river dispute, a bulletin published on 12 January says a pumping station “stopped operating on Saturday.” An analyst imports it that afternoon. On 14 January the publisher corrects the account: the station stopped late on Friday, while the loss of service was noticed on Saturday. A timeline that stores only the bulletin date cannot express the distinction. A timeline that overwrites Saturday with Friday hides what the earlier briefing relied on.

KAIROS explores how to retain these differences. It is not a live chronology service. Our proposed adapter would carry event-time candidates, source publication metadata, and the system’s record history into DIALECTICA for review. The record history describes what a system stored; it should not be presented as what an institution knew, believed, or could reasonably have known.

Resolve a date only when its anchor is justified

SUTime is an established rule-based system for recognizing and normalizing temporal expressions. Its paper explains the need for a reference date when resolving expressions such as “last Friday.”[1] That is a practical warning for our design: a quoted statement may require its original speaking date, not the date of the page that later republishes it.

A candidate needs the original expression, proposed normalized value or range, chosen anchor, timezone where relevant, and the reason for the choice. “Early January” should remain a bounded approximation. An unknown timezone should not quietly become UTC. Calendar dates, durations, recurring schedules, and deadlines also need different treatment.

Relations need room for an unresolved answer

Allen’s interval algebra describes thirteen basic relations between definite intervals, including before, overlaps, during, and meets.[2] Incomplete reports may support more than one possible relation. The proposed representation should retain a set of possibilities or an unresolved result rather than force one label. It should also state whether a relation concerns an event’s start, its end, or its whole duration.

This is an annotation problem as well as an engineering problem. Ning, Wu, and Roth’s MATRES work identifies confusion about event endpoints and proposes a multi-axis scheme using start points.[3] We take that as a reason to define the question carefully, not as evidence that one representation solves all temporal reasoning. A threatened strike, an agreed deadline, and a completed closure cannot be placed on the same factual timeline without recording their different status.

Corrections should remain inspectable

The proposed workflow keeps the first record, adds the correction, and links the two. A reviewer can then ask which account supported a particular version of a brief. A correction to event time does not automatically invalidate every associated claim; it may instead require a narrower change to an inference or recommendation.

For PRAXIS, the intended result is a reviewed capsule containing the chronology, unresolved dates, and relevant source history. KAIROS would supply candidate temporal information through a DIALECTICA-compatible adapter. It would not write directly into the workspace’s canonical records or certify that a reconstructed sequence is true.

Test temporal errors before drawing conclusions

We would test date-expression boundaries, anchor selection, timezone handling, and relation accuracy separately. The fixtures should include copied reports, daylight-saving changes, vague ranges, conflicting corrections, and a conditional event that never occurs. An abstention is appropriate when the text does not support a unique answer; evaluation should distinguish that from an avoidable omission.

A further test would reconstruct the same briefing at two record-history cutoffs and verify which sources were available in each version. Finally, ask whether the resulting prose overstates explanation: a protest following a policy announcement is an ordering observation, not proof that the announcement caused it. Any causal hypothesis needs additional evidence and competing explanations. No benchmark result is claimed here.

Keep uncertain time distinct from a convenient date

KAIROS is a historical temporal research reference, not a current composed production graph. In the reporting fixture, October publication and workspace capture are distinct fields even when their dates coincide. January and July are proposed timing at month precision, not adopted effective days. A proposed DIALECTICA adapter must retain unknowns, interval endpoints, source versions and unavailable evidence before any integration claim. [4]