BUILDER & AUTHOR
Shreyas Gowda S
Builder and author of ChronoRAG-G
Designed and built ChronoRAG-G around a simple requirement: temporal AI should show what evidence it used, when that evidence was valid, and where the system failed.
FROZEN EVALUATION REPLAY · C01
Seven facts. One quarter.Every answer component traceable.
Watch a complex question break into evidence obligations, bind to time-valid records, and rebuild into a grounded answer.
Loading the frozen C01 replay…
Reading sanitized evidence data from this site only.
ChronoRAG-G reduces temporal hallucination by keeping each requested fact tied to the evidence and period that support it.
THREE THINGS THE TRACE MUST KEEP TRUE
Time truth. Evidence truth. Inspectable decisions.
01Time truth
Was this evidence valid for the period being asked about?
ChronoRAG-G keeps temporal meaning attached to evidence instead of treating every similar passage as interchangeable.
02Evidence truth
What actually supports each part of the answer?
Answer components remain connected to stable evidence records rather than collapsing into one opaque context window.
03Inspectable trace
Can we reconstruct what happened?
The replay exposes what was required, what was retrieved, what was used, and where a failure occurred.
INTERACTIVE CASE VIEWER · C01
Follow one required fact through the evidence path.
Select an obligation, record, answer component, or period. Open the trace when you want the expert view.
Loading C01 from local replay data…
No remote request or live model is involved.
Loading the grounded-refusal replay…
Loading the failure-localization replays…
MORE THAN RETRIEVAL
Retrieval finds information.ChronoRAG-G keeps the answer accountable to it.
A difficult question can require several entities, metrics, and periods at once. ChronoRAG-G keeps those requirements visible, finds evidence for each one, checks when that evidence applies, carries evidence ownership into answer construction, and preserves the trace for audit.
That is how the system reduces temporal hallucination: the answer is built around time-valid evidence obligations instead of treating all relevant-looking context as interchangeable.
Question
What is actually being asked?
Scope
Which entities, metrics, operations, and periods matter?
Evidence obligations
What individual facts must exist before the answer is complete?
Time
Does each record support the requested period and temporal role?
Evidence ownership
Which evidence belongs to which required answer component?
Answer
Construct the result from resolved evidence.
Audit
Keep enough trace to inspect success, refusal, or failure.
FROZEN EVALUATION
The interaction is the proof.The numbers show the scale.
Evaluated on 1,005 temporal questions over 480 updated earnings-call transcripts.
80.70%
Audited answer accuracy
811 accepted answers across the complete 1,005-question evaluation.
90.44%
of the published gold-evidence oracle score
ChronoRAG-G reached this share of the LLM-GT oracle score while retrieving its own evidence.
101 / 101
New-query unanswerable cases
Every new-query case that lacked sufficient answer evidence was correctly handled as unanswerable in the frozen evaluation.
Frozen evaluation replay · ECT-QA · updated financial corpus
RESEARCH MENTORS
Research guidance and academic mentorship.
Dr Shivakumara Palaiahnakote
Lecturer in Computer Vision
University of Salford, UK
School of Science, Engineering & Environment
Dr Thippeswamy G
Professor & Principal
Nagarjuna College of Engineering and Technology, Bengaluru