Research agenda¶
The agenda turns gaps into research decisions. It prioritizes questions that could weaken a claim rather than only confirm it.
Priorities¶
| Priority | Falsifiable question | Minimum design | Progress signal | Risk |
|---|---|---|---|---|
| P1 · DI field | Does DI have its own community, methods and cumulative evidence? | preregistered systematic review and bibliometric map | reproducible criteria and search | mistaking sample absence for real absence |
| P1 · Human–AI | When does human+AI outperform both human and AI alone? | task-level comparison with allocation and defined outcomes | effect, heterogeneity, harm and cost | automation complacency or deskilling |
| P1 · Control | Which controls change behavior when an agent can act? | override, stop and reversal tests | intervention and recovery rates | ceremonial control |
| P2 · PULSE | Does PDAMR reduce material omissions versus the current process? | preregistered pilot with comparator | completeness, latency, burden and ex-ante quality | bureaucracy and gaming |
| P2 · Learning | Does the longitudinal record improve calibration or only documentation? | repeated comparable decisions and feedback | calibration, review and observable change | outcome bias |
| P2 · Decision Experience | Which combination of conversation, visualization and workflow works by task? | factorial or sequential study | accuracy, comprehension, time and accessibility | fluency mistaken for quality |
| P3 · Metaphors | Do biological lenses generate useful, distinguishable predictions? | observable hypotheses without metaphorical language | a prediction that can fail | anthropomorphism |
Minimum contracts¶
Every study in the program should state:
- decision and unit of analysis;
- population, context and exclusions;
- relevant comparator;
- outcomes and guardrails separated from adoption;
- measurement period and missing data;
- human authority and incident protocol;
- heterogeneity analysis and adverse results;
- materials, system version and deviations;
- criterion for strengthening, maintaining or weakening each claim.
Recommended first pilot¶
The B2B case can test instrumentation, but it should not move directly from simulated data to an effectiveness claim. The next decision is whether an owner, an authorized operation and a viable comparator exist. Otherwise, the case remains stopped.
A safer initial test would measure whether PDAMR and the Decision Brief reduce process omissions in historical or simulated decisions without allowing AI to execute actions. That result would inform usability and coverage, not causal sales impact.
Candidate protocol · When to add GraphRAG¶
Research decision. The knowledge-base owner must decide whether to add graph-based retrieval for questions about dependencies, currency or provenance. The two explorations described in the article help frame the test. The comparison remains to be run.
- Separate questions with direct answers from those that require following a typed relationship, such as which document superseded a rule and which cases it applies to. Include questions with insufficient evidence and restricted-access tests.
- Compare a text or hybrid retrieval baseline without a graph against the same setup enriched with relationships. Keep the versioned corpus, permissions, questions and response model the same. A link between documents counts as evidence of a relationship only when its type, validity and provenance can be verified.
- Set thresholds by question type before testing. Measure accuracy with verifiable citations, appropriate abstention, latency and the cost of maintaining relationships. Any unauthorized disclosure blocks adoption.
- Add the graph where the gain on relationship questions justifies its cost and risk. Repeat the test when the corpus or permissions change. If the gain disappears, keep the baseline.
Philia Sophia reading (25 Oct 2026) · Knowledge graphs: your AI needs the right relationships.
Apply the knowledge graph lab to the CDI-BoK to check reading paths and editorial provenance before designing a retrieval test.
Review cadence¶
- quarterly: standards, frameworks and future signals;
- for each candidate release: sources, claims and translations;
- after each study: null and adverse results, and deviations;
- annually: reproducible state-of-the-art search;
- exceptionally: regulatory change, material incident or new contradictory evidence.
Success criterion¶
The program advances when a claim becomes more precise even if its strength decreases. More pages or more portal usage do not demonstrate better decisions.