This page explains the current Cairn research results in plain language. It includes what worked, what only partly worked, the important numbers and the limits of each result.
A successful test tells us something useful, but it does not prove that every possible version of the problem is solved.
If a test passes, we explain what passed. If part of a test is incomplete or fails, that stays in the record too. The goal is to make the results useful without overselling them.
The first Q1 test asked whether a replacement AI model could pick up from the same preserved history without inventing a different version of the current state.
Two materially different model families recovered the same predefined operational state, the same unfinished commitments, the same revision relationships and the same resume point.
Neither model added unsupported claims to the authoritative state, and the two models did not disagree on the predefined assertions being scored.
This supports the idea that important working state can live outside the model and be handed to a replacement model in a form it can recover reliably.
One source-tracing weakness showed up in the trial. We kept that weakness in the result instead of rewriting the test after seeing it.
This test was about recovering defined operational state. It does not mean two different AI models will have the same personality, reasoning style, wording or judgment in open-ended situations.
The follow-up asked whether the replacement model really needed a huge amount of historical text, or whether a smaller structured package could carry the important source relationships.
Claude and Grok met every required target and recovered the same required source relationships.
Gemini remains recorded as a partial result. We did not quietly remove it because the other two performed better.
When the important history and its sources were represented directly, the completed model arms could recover the required evidence without being given the entire past as one giant block of context.
Q2 asks whether a long-lived system can recover efficiently as its history grows, instead of rereading everything from the beginning every time it resumes.
The largest required Trial 001 test contained one million protected synthetic events.
Verified checkpoint recovery rebuilt the same final state while reading only the most recent 10,000-event tail.
The dependency-aware path only had to apply 232 events that could actually change the final state or unfinished commitments.
When checkpoints, dependency data and protected history were deliberately damaged, none of the corruption tests silently returned a false verified state.
For this controlled workload, checkpoint recovery rebuilt the same final state as full replay while doing far less work.
Trial 002 made the test stricter. We kept the recovery window fixed at exactly 10,000 events while the total lifetime history grew from 100,000 to 1,000,000 events.
Verified checkpoint recovery with a 10,000-event recovery window.
Verified checkpoint recovery with the same 10,000-event recovery window.
Time to verify and replay the entire 100,000-event history.
Time to verify and replay the entire one-million-event history.
The 1M recovery time divided by the 100K recovery time. The test allowed anything up to 2.0.
Normal checkpoint recovery did not reread event bodies from before the checkpoint at either required scale.
The dependency-aware path applied 229 relevant events at 100K total history and 232 relevant events at 1M total history while still rebuilding the same final state.
We deliberately damaged checkpoint data, dependency data and protected history. The system rejected the bad data and either recovered safely or stopped instead of pretending the result was valid.
Full replay became roughly 10× slower when total history grew 10×. Fixed-window recovery stayed almost exactly the same.
The results are encouraging, but there are still clear limits to what we can say from them.
They were controlled research fixtures built so the tests could be repeated exactly. Real systems may have different event patterns.
The milliseconds on this page show how the approaches scaled in the test environment. They are not a promised production response time.
The models recovered the predefined operational state. They can still reason and behave differently.
Fast checkpoint recovery is for normal resume. A full origin-to-current verification remains a different and more expensive operation.
The required Trial 002 ladder ended at one million events. Ten million remains an extended target.
We have not yet completed the direct comparison between structured history and brute-force long context.
The next question is straightforward: does structured history actually outperform simply giving the model more and more old context?
We will keep Q3 marked open until that comparison is run and scored.
Keep the history.
Test the recovery.
Say what the results actually show.