MODEL
REGRESSION
FORENSICS.

A working research direction for tracing behavioral regressions through training lineage and verifying likely causes through controlled intervention.

01 / PROBLEM

Observe the regression. Find the cause.

Model teams can observe that a checkpoint regressed, yet still struggle to identify which training change caused the behavioral failure. This project asks whether behavioral differencing, training-lineage search, and targeted interventions can turn regression debugging into a reproducible engineering workflow.

02 / RESEARCH QUESTION
Can an automated debugger localize the training change responsible for a model regression, then verify that diagnosis through intervention rather than correlation alone?
03 / PROPOSED LOOP
BEHAVIOR DIFFREGRESSION CLUSTERLINEAGE SEARCHINTERVENTIONVERIFIED CAUSE
ADetect

Measure a behavioral delta between baseline and candidate checkpoints.

BLocalize

Rank changed data, configuration, or training phases as candidate causes.

CIntervene

Remove, replace, or replay the suspected change under controlled conditions.

DVerify

Test recovery on held-out cases and unrelated capabilities.

04 / NOVELTY STATUS

Adjacent work exists. The gap must be earned.

This project does not currently claim novelty. The working wedge is end-to-end regression forensics that joins behavioral differencing, training lineage, candidate-cause ranking, and intervention-backed verification. The related-work review must establish whether that combination is sufficiently distinct and useful.

05 / EXPERIMENT 000

Start with one planted regression.

Create a small controlled post-training experiment with a known root cause. Detect the behavioral change, search the recorded training lineage, rank candidate causes, execute the minimal intervention, and test whether the behavior recovers on held-out examples without damaging unrelated capabilities.

CONTROLLED EXPERIMENT / SCHEMANOT RUN
BASELINEopen model + clean SFT mixture
PLANTED CHANGEsingle known data/config intervention
OBSERVEtarget eval regression + behavioral traces
DIAGNOSErank candidate causes from lineage
VERIFYretrain/replay without suspected cause
SUCCESSheld-out recovery + stable unrelated evals
06 / RESEARCH LOG
LOG 000

Defined the initial question and began reviewing adjacent work.

EXP 000

Plant one reproducible post-training regression and test whether a controlled intervention recovers the held-out behavior.

07 / REPRODUCE

Reproduction starts with Experiment 000.

Commands, configs, eval definitions, and summarized results will be added after the first experiment has been run and reproduced from a clean checkout.

$ experiment 000 has not been executed yet
NOTE / 2026-08

This page separates the current research question from completed evidence. Experimental results will be added only after they are reproduced.