Failure Receipt
Retain the episode, task condition, observed outcome, provenance, and known unknowns.
Technology
Turn failure into a testable next-data decision.
CPF is an evidence-first research framework for tracing a model failure through controlled intervention, bounded hypothesis formation, and a collectable DataPlan. It is designed to make a claim inspectable—not to infer causality from a single trace.
System architecture
An inspectable system, not a black box.
CPF sits beside a team’s existing models, evaluation tools, simulators, data systems, and training loop. Its role is to structure the evidence between a failure and the next decision.
Controlled Branches
Compare causal, placebo, and no-change conditions against a comparable anchor.
DataGap Hypothesis
Express the candidate missing experience as a bounded, falsifiable statement.
DataPlan
Specify what to collect next, the sampling range, and the validation measure.
CPF core loop
One failure. One controlled loop.
A useful diagnosis does not stop at explanation. It produces a next action whose value can be tested after retraining on held-out evaluation.
See the method
- 01
Observe
Record the failure and preserve the trace.
- 02
Intervene
Change one relevant factor under a controlled branch.
- 03
Compare
Separate the intervention from placebo and no-change controls.
- 04
Hypothesize
State what evidence may be missing, including uncertainty.
- 05
Plan
Form a bounded collection recommendation and success measure.
- 06
Validate
Retrain and compare before claiming the loop improved anything.
Designed to connect
Works with the stack a team already has.
Models & checkpoints
Start from an existing policy, model interface, or observable evaluation output.
Evaluation & simulation
Define repeatable interventions and control conditions around one task failure.
Data & retraining
Return a traceable DataPlan that can be collected and tested in the team’s own loop.
Technical maturity
Current, in validation, and next.
The public technology story mirrors the research status: it distinguishes working foundations from experiments still underway and long-term research goals.
Evidence objects
External CPF overview, traceability requirements, and a bounded intervention workflow.
Phase 1 validation
Episode bundle verification, controlled simulator studies, and retraining-validation design.
Physical causal learner
A longer-term system that learns from validated causal experience loops in physical tasks.
Read the evidence-first technical overview.
A public view of CPF’s workflow, boundaries, and the validation still required.