Experimental design and statistical judgment
Paired controls, preregistered equivalence margins, lineage-level comparisons, and explicit uncertainty. A promising pilot is not treated as a confirmed effect.
Kleos: a pilot that did not replicate →Methods & About
The practice of careful doubt.
Abraham Labs is an independent research practice studying whether mechanisms inspired by human learning improve machine capability, efficiency, transfer, or adaptation.
Post-training and evaluation engineering
Verified training answers, controlled explanation treatments, parameter-efficient fine-tuning, paired adapter seeds, and checkpoint learning curves. Training success is measured separately from transfer to new problems.
Route specialization: teacher versus reference →Auditable systems and failure diagnosis
Frozen inputs, append-only observations, artifact hashes, task-overlap checks, and independently implemented scoring. Interface failures and metric mismatches are documented instead of hidden behind a single accuracy number.
When the student could not use the interface →Current work
From static lessons to adaptive apprenticeship.
The completed loop found a small average adaptive advantage but inconsistent lineages, not reliable superiority. A fresh follow-up is in preparation: richer practice-error feedback, fixed original replay, and starting-parent measurement on the same final pool.
Read the follow-up design →Earlier pilots, regressions, and the decisions they informed →
Falsifiable before fashionable
Substantive studies define the question, null and alternative hypotheses, treatment, controls, metrics, confounds, exclusions, analysis, and stopping rule before execution.
Frozen means frozen
A changed prompt, task set, model digest, or analysis plan creates a new revision. Prior designs are never silently rewritten.
Negative results remain results
A well-powered equivalence result narrows the world. It is not tuned, reframed, or rerun until it becomes exciting.
Claims stop at the evidence
Prompt-conditioned behavior is not called weight learning. Generated text is not treated as proof of consciousness, identity, or faithful introspection.
Independence disclosure
Abraham, without borrowed authority.
Abraham directs the research questions, scientific judgments, and publication decisions. AI assistants support implementation, analysis, and adversarial review. Their agreement is not treated as external replication or peer review.
Abraham Labs is not affiliated with a university or commercial laboratory. The name is a research brand, not a claim of institutional status.
Research & engineering opportunities
Discuss the work.
For opportunities in model evaluation, post-training, or research engineering, get in touch. The reports provide concrete examples of the questions, controls, implementation choices, and limitations behind the work.
Contact Abraham →