← Research

Case study

How Kairos reuses its own conclusions — a flywheel, not a one-shot answer

October 14, 2026

A research assistant that answers every question from scratch forgets its own work — and can silently contradict it. Kairos keeps a research record: every verified finding is stored with its evidence, and later questions reuse it before re-deriving anything. In October 2026 the first daily telemetry of that reuse went live. Here is the mechanism, the first 36 days, and the boundary.

Boundary. The telemetry counts reuse of recorded conclusions — not research quality, and not wet-lab truth. A flywheel that sounds like a guarantee is just another fabrication: the numbers below are aggregate counts from the live endpoint, and they cover only the window they cover.

The failure mode: a research tool that forgets

Ask the same platform the same class of question twice, and a stateless assistant re-reads the literature, re-runs retrieval, and re-derives the conclusion — or produces a new one that contradicts what it answered last week. The expensive part of research is not the answer; it is proving the answer. Doing that proof twice is waste. Contradicting it silently is worse.

The mechanism: a record only verified findings enter

A conclusion becomes a finding only when it is backed by traceable evidence — the review layer adjudicates every citation marker against the evidence pool before anything reaches the record. From there, three rules. Reuse first: when a new question arrives, Kairos searches the record — with the project's context — before the evidence fastpath, so a shallow quick-hit never preempts a structured verdict from prior conclusions. Incremental when it can: if more than half of the matching conclusions belong to the same project, the answer is delivered incrementally — the questioner gets what is new, built on what the project already holds, not a re-derivation. Provenance survives reuse: every reused answer carries deterministic audit metadata — how many conclusions it reused, how many carry PMIDs — so reuse never becomes a way to cite nothing.

The telemetry: 36 days, counted per day

The reuse-series endpoint publishes a daily series over the telemetry window — 2026-09-08 through 2026-10-13, UTC days. The first 36 days show what reuse actually looks like, not what a marketing chart would look like. The record holds 580 verified findings across 106 projects, up from 525 when the window opened. 89 reuse events landed on 19 of the 36 days; the other 17 days had none. Reuse is bursty: the first three days carried 49 events, most of the window was quiet, and a nine-event day arrived near the end. 27 of the events were incremental answers — the same count the World Models page shows live. Daily reuse rate — the day's events divided by the findings in the record — ranged from zero to 5.5%.

The figures

Both figures are drawn point for point from one response of the live endpoint.

525540560580DAY (UTC, 2026-09-08 → 2026-10-13) →VERIFIED FINDINGS →55 findings added over 36 days
Fig. 1The record compounds: 55 findings were added over 36 days. The line connects the daily values — days on which no finding was added are flat.Live endpoint /kairos-api/research-record/reuse-series · records field (cumulative) · 2026-09-08 → 2026-10-13 (UTC days)
0102030DAY (UTC, 2026-09-08 → 2026-10-13) →REUSE EVENTS →89 events on 19 of 36 days
Fig. 2Reuse is bursty, not smooth: 89 events across 19 days, with the first three days carrying 49. Days with no bar recorded no reuse event.Live endpoint /kairos-api/research-record/reuse-series · reuse_hits field (per day) · 17 of 36 days are true zeros

What the endpoint says about itself

The endpoint ships with its own caveat, reproduced verbatim: days before the window are unpublished — unknowable, not zero — and days inside the window with no telemetry row are true zeros. That distinction is the whole difference between a chart that honestly reports a window and a chart that pretends to show a history. The World Models page keeps the live series; this article is a snapshot of October 14.

Watch the flywheel compound — live counts and the daily trend are on the World Models page.Open World Models →

Sources

Every number on this page traces to an endpoint or a page you can check yourself:

  1. Kairos World Models — the live flywheel page: current record counts and the daily reuse trend. /kairos/world-models/
  2. The reuse-series endpoint — the raw daily series this article draws its figures from. /kairos-api/research-record/reuse-series
  3. Citation gate — how hallucinated citation markers are kept out of the record. /research/kairos-citation-gate/
  4. Kairos Trust — the review layer that decides what enters the record. /trust/