Evidence over claims. Assurance over automation.

Research & field notes

Public work with visible status, method, and limits.

This library grows as concept notes, method notes, apparatus lessons, and study records. Unfinished work is labelled unfinished. Synthetic examples live under Examples, not as client outcomes.

Field notes

Concepts and methods you can reuse

Signature distinctions and operational maps. Working definitions are labelled as such. Each note includes a checklist, matrix, or decision rule.

Concept note

Evidence debt: the proof you meant to collect later

The gap between claims an organization currently relies on and the evidence it can currently produce under the scope that now applies. Inventory fields, interest, and a decision rule before repeating a claim.

Read the concept note
Method note

Why every regulatory change needs a record chain

Step, authoritative record, failure signal, owner, evidence required. Why each column exists, how missing one creates evidence debt, and when to reopen the map after a published amendment.

Read the method note
Concept note

A citation is not support

Separate citation presence, retrieval relevance, entailment, and decision usefulness so relevant passages stop laundering unsupported claims.

Read the concept note
Apparatus lesson

When agreement is zero: how a claim auditor failed its gold

A deterministic claim auditor cleared engineering gates and failed blind human gold. Lexical match is not support. Independent first-pass audit still requires human disposition.

Read the apparatus lesson

Study records and technical notes

Longer programmes with explicit limits

Active empirical work and architecture notes. Status language stays honest.

Active empirical study

Do structured workflow scaffolds reduce unsupported claims?

A controlled study testing whether provenance, uncertainty, disconfirmation, and final claim-audit controls reduce unsupported claims in AI-assisted research without winning by saying less.

Open the study record
Technical note

Intent graphs and goal orchestration in AI systems

Intent graphs across semantic routing, task-dependency planning, skill composition, and commercial search journeys, with a MainFrame implementation lens.

Read the technical note
Working paper

Reliability benchmark landscape and credible differentiation

Mapping work on false-completion detection, appropriate abstention when evidence is absent, and evaluator self-calibration against blind human gold.

Examples live elsewhere

Method walkthroughs are under Examples, not the research feed.

Synthetic OOS reconstruction, regulatory-change maps, and the buyer-readiness claims package show deliverable shape without borrowing client credibility. Permissioned peer collaboration notes appear there when approved.

Status vocabulary

  • Concept / method note: bounded public analysis
  • Apparatus lesson: measurement or tooling failure with limits
  • Active study: incomplete analysis; no premature causal claim
  • Working paper: circulating draft, not peer-reviewed
  • Public synthetic: invented scenario, labelled

Need this kind of research attached to your own decision?

The Evidence Intelligence Research Sprint converts a consequential question into a controlled source register, analysis, decision memo, appendix, and briefing.