Research question
Across extended tasks and sessions, how reliably does a model preserve explicit constraints, carry corrections forward, and recover after an error, and does that reliability differ between model families and versions?
Why this question
Most evaluations measure a model inside one exchange. Practical work happens over many: a constraint is given early, other work intervenes, a correction is made, and the same kind of decision comes up again later. Failures in that setting are hard to see in single-turn tests and easy to misattribute in anecdote.
What would count as evidence
- A fixed set of constraints and corrections introduced at known points, with retention checked at defined later points.
- Repeated trials per condition, with the denominator stated.
- At least two model families and a record of the exact versions and access dates.
- Controls that separate forgetting from instruction conflict, and conflict from ambiguity in the task itself.
Planned next step
Convert documented observations into a controlled, repeatable protocol, then publish that protocol here before running it.
Hypothesis
Unavailable No formal hypothesis is registered yet. It will be stated in the protocol before any trial is run.
Method
Unavailable The protocol has not been published.
Results
Unavailable No experiment has been run. No result is implied.
Disclosures
- The question was motivated by observations made during extended practical work with AI systems. Those observations are anecdotal and are not presented as findings.
Revision history
- Entry created as a research question.
Cite this entry
Sai Bharath Siliveri (2026). Longitudinal AI reliability. ALGEDI SI Research Labs, ALG-001, research question, revision of 2026-10-15. https://algedi.si/research/longitudinal-ai-reliability/