Move the model.Keep the capability.
Companies replace models and can silently lose behavior their application depends on. InheritBench finds what broke, helps recover it, and proves whether the replacement is ready to ship.
What InheritBench does
Find what broke
Measure which required behaviors disappeared after the model change.
Recover the behavior
Target the gaps and select a recovered successor under control.
Decide whether to ship
Run final evaluation and safety checks against explicit readiness rules.
The browser evaluates precomputed predictions. Model training and inference happen outside the browser.
The replacement risk
A replacement can sound right and still break your application.
Teams switch models for cost, latency, licensing, infrastructure, or vendor reasons. General benchmarks do not prove that application-specific rules survived the move.
Plausible text can still use the wrong policy code, bypass an approval, violate a contract, or choose an unsafe action. InheritBench tests the behavior the application actually requires.
The reference replacement produced valid-looking output while preserving only 0 of 32 required behaviors.
How it works
Diagnose → Recover → Assure
Turn a model replacement into three clear questions: what was lost, can it be recovered, and is the successor ready to ship?
Diagnose
Compare the replacement with the behavior your application already depends on.
Recover
Target the missing behavior and compare recovery candidates without using final-test results.
Assure
Apply the same evaluation and safety rules to decide whether the recovered successor is ready to ship.
Try it yourself
Choose → Run → Review
Choose a built-in candidate, run the evaluation in your browser, and review the readiness decision. Then introduce a controlled failure and see the decision change.
Reference result
Qwen capability recovered on an OLMo successor.
Four recovery candidates were compared without looking at final-test results. Candidate 0 was selected, then evaluated once on clean and adversarial records.
- Source
- Qwen2.5-0.5B
- Successor
- OLMo-2-0425-1B
- Selected candidate
- Candidate 0
64 / 64
Operational correctness
63 / 64
Exact-contract fidelity
64 / 64
Strict validity
2 on 1 adversarial record
Safety findings
CONDITIONAL_PASS
Readiness
Clean behavior recovered fully. Adversarial evidence retained 2 safety findings on 1 record, producing a conditional deployment decision.