Introducing node&norm
Delegating work to AI leaves people and institutions responsible for its consequences. node&norm examines whether they can question decisions, withdraw authority and establish what happened. This independent research program connects governance, evaluation and assurance, with AGI preparedness as a longer-term purpose.
Explore the connections
- 1Question
- 2Connection
- 3Source
1. Choose a question
Loading the reviewed map…
Search & filter records
At the widest view, ribbons count direct research–governance links by theme. Zoom in for lines between visible records. Select a record for its reviewed links and, on wide screens, an outward continuation through those records.
List view
Browse or search the same map records without panning or zooming.
No matching records. Try a different search or clear the filters.
Columns group records by type. Stronger lines are reviewed direct links; fainter lines continue outward through an intermediate record. “One step on” does not establish a direct link, causation or validation. The card provides the same records and intermediates as lists.
Drag to explore · + to zoom
Map key & essay symbols
In the map
- Research question
- Governance topic
- Risk record
- Incident case
- Solid: research relevance or related governance topics
- Dashed: editorial risk links
- Dotted: selected incident context
Five editorial themes organize the records. Positions support navigation; they do not measure similarity, risk or strength of evidence.
How the six essay illustrations work
In the essay illustrations
The essay illustrations reuse these shapes to tell a story. Their meanings below differ from the map’s record types and reviewed connections.
- A person or institution that can act
- A record, rule or decision
- An error or risk
- An affected person or open matter
- An established link in the illustration
- A conditional or unconfirmed link
- A citron intervention or unresolved step
01 Responsibility & repair
The record is corrected, and the harm stays open. The citron arc is the part still owed.
02 Practical human control
A reviewer sits on the loop. The bar shows whether the decision still travels on to the person it affects.
03 Human influence
Two reviewers, the same evidence and rules, one shared conclusion. The crossing lines are what they had in common.
04 Evaluation rules
The score runs high. One disqualifying error draws a bar across the whole result.
05 Decision records
Approval is on record. The step that would reach the person has no record yet, and that gap is marked.
06 Conditional approval
Past the gate, one path continues and one waits. Two conditions are closed and one remains open.
These diagrams illustrate the essays’ questions. They are not mapped records or evidence that a control worked.
Sources, connections and limits
How the themes work & what we plan to test →
What the 640 assessments cover. Each of eight research areas was reviewed against each of 80 curated governance topics: 8 × 80 = 640 assessments of relevance. That review identified 137 direct research–governance links. Background context is retained in the cards; connections not established by the review remain absent. This total counts reviewed pairings, not every line or record in the map. Risk and incident connections are reviewed separately; they have not all been compared with every research area or governance topic.
| Collection | In the map | Review pool |
|---|---|---|
| MIT risk records | 530 | Up to 1,589 |
| AIID incident cases | 216 | 1,786 |
| Risk and incident total | 746 | Up to 3,375 |
834 dots now; up to 3,463 if every candidate were accepted. Both totals include the eight research areas and 80 governance topics. The review pool includes records already mapped. It is an upper bound for review; the final published selection depends on source quality and relevance, and some candidates may be excluded or deferred.
How the review pool is counted
The retained MIT snapshot contains 1,835 source rows. Exclusion and identity checks leave 1,589 records in the review pool, including those already mapped. The AIID public table displayed 1,786 incidents when checked on November 3, 2026. Its count may change, and multiple reports about one incident are not additional incident dots. The earlier combined source pool of 3,621 records was counted before the MIT exclusions; the working risk-and-incident review pool is 3,375. These are source records, not necessarily distinct hazards or independently confirmed harms. The MIT retrieval date remains October 8, 2026.
Additions are prepared in batches of 25 risk records and 10 incident cases, with source and editorial checks before publication. The original five risk mappings retain their October 8, 2026 review; the first ten batches of 25 risk additions were reviewed on November 2 and 3, batches eleven through fifteen on November 6, batch sixteen on November 7, batches seventeen through twenty on November 12, and batch twenty-one on November 13, using retained source metadata. These risk additions do not claim newly completed full-paper reviews. Selected incident cases retain their supporting sources, evidence status and review notes. They include historical automation, research observations, company disclosures and allegations clarified by later investigations; inclusion does not establish that every case demonstrates harm. The selection is not representative or an automated incident feed.
The layout is authored for navigation, not a vector embedding. Links express editorial relevance, not causation, legal applicability, compliance, research validation or effective control. At the widest view, ribbons summarize direct research–governance links by theme. Zooming in reveals individual recorded links. Selecting a record highlights its reviewed direct links. On wide screens, the trace also follows existing links outward through intermediate records; “one step on” is not a reviewed direct link to the selected record. The card lists the same records and intermediates on every screen size. The full-network control hides overview connections without hiding the selected trace. Themes are provisional navigation categories, not a validated taxonomy or a causal model.
Risk metadata: MIT AI Risk Repository, adapted under CC BY 4.0. Selected incident metadata: AI Incident Database, Responsible AI Collaborative. The adapted incident selection and editorial crosswalk are available under CC BY-SA 4.0; summaries and connections are editorial additions. Original incident IDs, titles, dates and displayed editor credits remain attached.
Theme assignments and reasons (JSON) ↗ · Risk records and batch review scope (JSON) ↗ · 640 research–governance assessments (CSV) · Research–governance crosswalk (JSON) ↗ · Incident selection, evidence and credits (JSON) ↗