Short educational notes for reading our experiments critically. Each starts with a distinction, follows a hypothetical example and ends with a test. These notes are not peer-reviewed papers, clinical guidance or completed experiments.
Confidence describes a model’s score or expressed certainty. Evidence is the material that bears on a claim: its source, availability, relevance and relationship to the premises. A persuasive answer may lack a necessary premise; a low-confidence answer may still identify a genuine contradiction.
A system recommends a supplier using a historical compliance certificate. The certificate is authentic, but the current licence is missing. Familiarity with the supplier cannot fill that gap. Preserve the missing premise and request the current record.
Treating semantic similarity, model confidence or source prestige as sufficient support.
A calibrated probability estimates correctness frequency over an evaluated population. It does not establish that this particular answer is justified. Abstention introduces a decision rule: which cases should be deferred, with what review cost and residual error?
Suppose a classifier assigns confidence near 0.8 to a group of cases. Compare their actual correctness frequency on held-out data. Then separately inspect whether excluded cases or changed deployment conditions invalidate the comparison. No measurements are asserted in this example.
Calibrating on the test set, reporting only accepted-case accuracy, or mistaking marginal coverage for certainty about every individual.
An optimization changes an execution path. Numerical equivalence concerns specified tolerances; semantic equivalence concerns the states and decisions that matter to the task. Passing one does not automatically pass the other.
Imagine an evidence score close to an action threshold. A small precision change could satisfy a global numerical tolerance yet cross that threshold. The appropriate comparison includes threshold-adjacent fixtures and the resulting action state.
Accepting throughput improvement before checking semantics, or requiring identical text for every stochastic system without a task-specific criterion.
Latency is elapsed time for an operation; throughput is completed work per unit time. Independent memory requests may overlap. A dependent next-address chain restricts that overlap. Data movement and operational intensity can constrain useful compute.
Two loops visit the same nominal data size. One accesses predictable indices; the other visits a random order. A third obtains each next address from the previous load. Their time per access need not measure the same hardware property.
Calling throughput-derived nanoseconds per access raw DRAM latency, overlooking index storage, or naming an exact cache boundary from a broad transition.
Reconciliation tests whether a workload keeps an auditable identity through transitions. Request, semantic, resource and authority checks are candidate ACE invariants. Each requires a defined scope: logical requests differ from retry attempts, and declared capacity differs from physical allocation.
A request is retried after a timeout. Two attempts may appear in the log, but the logical request needs an explicit outcome. A scheduler may declare eviction before the process stops; capacity release must reflect the chosen accounting policy and observed execution.
Equating lifecycle counters without draining pending work, releasing resources twice, or treating object discovery as mutation authority.
Behavioural measurement connects an observable signal to a defined construct. Attention, preference, comprehension, trust and purchase are different targets. A sensor or model can produce a repeatable signal without establishing the interpretation attached to it.
Longer gaze at packaging could reflect interest, difficulty reading it or confusion. A single attention score cannot distinguish those explanations. Combine appropriate task measures and context rather than declaring intent.
Inferring emotion or purchase certainty from a proxy, ignoring consent and population differences, or substituting synthetic personas for observed participants.
Definitions are starting points. Useful research specifies the task, collects suitable evidence and exposes a claim to failure. Our measured graphs, proposed tests and external readings keep their separate statuses.
Existing results → · Open questions → · The NeuroAlchemist →