Field Lab
Practical article-based experiments with a protocol, an inspectable output, success criteria, and honest limits.
20 published
Every page has one canonical home and remains connected to its primary learning domain.
A documented ten-item pilot for turning one bounded task into an AI-assisted workflow with explicit inputs, checks, failure routes, and accountability.
Across one hundred frozen citation cases, which faults are observable from structured records and which still require semantic judgment? Executed on frozen inputs with inspectable
When thirty bounded source statements are truncated at three fixed word budgets, how often does the prespecified qualification remain visible? Executed on frozen inputs with inspec
How do four locked search strategies trade recall against precision on a frozen collection with known eligibility labels? Executed on frozen inputs with inspectable results, negati
Which graph properties make a proposed learning path impossible, overconstrained, or dependent on a single bottleneck? Executed on frozen inputs with inspectable results, negative
Across forty-eight designed claims, how often does a source with a different institutional role reveal conflict or a narrower scope? Executed on frozen inputs with inspectable resu
How much known relevant material becomes retrievable when a term index expands from titles to abstracts and then to full text? Executed on frozen inputs with inspectable results, n
Which scalar, list, nested provenance, type, and Unicode fields survive JSON, flat CSV, and Markdown export transformations? Executed on frozen inputs with inspectable results, neg
Which evidence-boundary fields survive when sixty structured statements are transformed into claim-only, dated, scoped, or full evidence cards? Executed on frozen inputs with inspe
How does checkpoint performance change when error observability, evidence access, reviewer capacity, and stop authority vary independently? Executed on frozen inputs with inspectab
How do four human-review thresholds trade review cost against expected error loss across tasks with different consequences and reversibility? Executed on frozen inputs with inspect
How do four scheduling rules respond to twelve frozen interruptions without creating an unmanageable review backlog? Executed on frozen inputs with inspectable results, negative fi
How differently do blocked, round-robin, paired, and seeded-shuffle generators distribute ninety-six practice items from eight categories? Executed on frozen inputs with inspectabl
How do sixty test protocols distribute across near, moderate, and far transfer when four dimensions of task distance are declared explicitly? Executed on frozen inputs with inspect
A reproducible 160-string test of NFC, NFKC, search-key equality, width variants, combining marks, compatibility forms, and directional controls.
When forty claims contain objections, rebuttals, scope limits, and unresolved conflicts, what is lost by flattening them into linear summaries? Executed on frozen inputs with inspe
Which combinations of observability, direction, threshold, time window, and scenario exclusivity make an indicator discriminating rather than decorative? Executed on frozen inputs
How do fixed, abrupt, performance-contingent, and error-contingent fading rules behave on frozen performance traces? Executed on frozen inputs with inspectable results, negative fi
Across sixty bounded claims and four short-form constraints, when does an explicit qualification survive instead of being traded for a cleaner headline? Executed on frozen inputs w
Across one hundred quality decisions in five knowledge-work domains, which decisions reduce to declared observables and which require contextual interpretation? Executed on frozen