CHRONOS works the question.
A session brings models, tools, memory, and evidence to bear on one hard question.
ProjectForty2 / Training Signal
Training signal from real work.
CHRONOS does not ask models to sound creative. It compares what failed to move a hard question with what was associated with measurable progress, then turns that gap into durable preference data.
Same question / same pressure / better trace
signal online
paired trace
Same problem. Different quality.
The useful signal is the difference between spinning in place and finding a new handle.
Rejected
Clean prose that restates known territory.
Stored
A trace that exposes structure the next session can use.
How it works
Preference data with evidence attached.
The training signal stays grounded because it comes from the same research surface, Atlas memory, and verification discipline the public pages expose.
Failure and progress share the same frame.
The pair is useful because both outputs faced the same anchor, context, and pressure.
The gap becomes reusable signal.
The stored pair teaches what advancing looked like, where the weaker trace stalled, and what evidence made the difference.
Why it matters
Evidence-linked preference data.
Instead of asking annotators to label creativity after the fact, CHRONOS records whether a trace helped the work move. The page stays simple because the supporting evidence lives in the session history and public evidence layer.
The answer is judged against the problem.
Atlas context and ledgers keep the comparison tied to the actual research frontier.
The correction is specific.
A model weak on synthesis gets synthesis pairs. A model weak on verification gets verification pairs.
Better sessions produce harder pairs.
The loop tightens as the questions get harder and the useful traces become more precise.
Train on work that made measurable progress.
CHRONOS turns research progress into preference data without hiding the evidence trail that made the preference meaningful.