# Channel-specific reported-risk networks

The two networks reproduce the paper's frozen non-circular top-three mapping. OECD AIM is the primary channel with 10,449 incidents; AIID is the secondary channel with 1,751 incidents. AIM has 516 activated cards and 7,052 unique card-pair edges. AIID has 459 activated cards and 2,808 edges.

The estimand is the structure of candidate risks revealed in collected reports. Counts and network measures do not estimate population incidence, causality, or severity. The channels contain linked records and are not independent samples of all real-world harms.

## Inputs and reproduction

Use the frozen `channel_analysis_20261024/inputs/channel_edges_k5.csv.gz`, `channel_analytic_records.csv.gz`, and `01_channel_results/T1_revealed_network.csv`. Place the network metric CSV beside the two input files, then run from the site checkout with pandas, NumPy, and NetworkX installed.

```bash
python scripts/build_channel_incidents.py --inputs /path/to/frozen/channel-inputs
```

The seed is 20261021. Input hashes are recorded in `validation.json` and `networks.json`. The script asserts channel counts, unique candidate cards per incident, inventory identity, and agreement with the paper's graph outputs. The input records used here contain analytic metadata rather than incident narrative text.

## Calculation

For each incident, retain ranks 1–3 and form the three unordered pairs among its distinct candidate cards. Every unique pair forms an undirected edge, and repeated pairs increment its count. Degree is the number of distinct neighbors in the full channel graph. Cross-domain incident share counts records whose three cards cover at least two L1 domains. Cross-domain edge share counts unique edges joining different L1 domains. These are different denominators.

A revelation ratio divides a domain's share of card assignments by its share of the 622-card inventory. AIM contributes 31,347 assignments; AIID contributes 5,253. This ratio measures relative representation against inventory composition, not prevalence. Domain assortativity uses the unweighted network topology. Network O/E values are taken from the paper's 1,000 within-channel label permutations.

## Display

The overview starts with the 25 nodes with the highest observed incident counts per domain plus the control-transfer card, retains the 240 strongest induced ties, and removes displayed isolates. The full view includes all active nodes and edges. Selecting a card preserves the graph layout and viewport, and opens a persistent right-hand incident panel with the source title, exact source summary, date, and original incident URL. The list switches between AIM and AIID and loads 15 entries at a time. Zooming in reveals full risk-name labels with collision-aware placement. Filtering never recomputes the reported full-network statistics.

L4 colors come directly from `data/semantic_space.json`. The same card has the same fill as in AI risk space, including semantic shades across L3 families. The domain legend uses the existing L1 anchor colors. Nodes are circles with white borders, and edges follow the existing quadratic-curve convention. A red outline identifies `P_SYS_CONTROL_003` without changing its fill.

Node area increases linearly with full-channel degree above a minimum area. Edge width increases with the square root of pair co-activation rate per 1,000 channel incidents. Both panels share these visual scales. The graphs are stacked vertically. The joint force layout uses a 1200-by-680 canvas, stronger repulsion, and a collision-separation pass with 22 extra coordinate units between overview node boundaries. Common cards retain shared coordinates for channel comparisons. Layout distances are not calibrated semantic, causal, or geographic distances.

## Files

- `networks.json` contains both complete aggregate networks, risk definitions, colors, full-network metrics, and display coordinates.
- `nodes.csv` contains degree and activation counts for all activated cards in each channel.
- `edges.csv` contains each card pair, its co-activation count, its per-1,000 rate, and cross-domain status.
- `validation.json` contains input hashes, counts, checks, seed, and rendering definitions.

The incident-detail assets contain titles and summaries from the existing public source-record release, with original database URLs and source attribution. Their title/description/AI-system hashes match every frozen mapping input. Underlying source article full texts and human-adjudication material are not exported.

## Linked incident details

```bash
python scripts/build_channel_incident_details.py --inputs /path/to/frozen/channel-inputs
```

This second script joins the existing public `source_records.csv.gz` to the paper’s frozen records by channel-specific ID. It asserts 12,200 exact title/description/AI-system hash matches and checks that every card’s list length equals its activation count. The generated `incidents-AIM.json` and `incidents-AIID.json` are fetched only when that source is selected. `incident_details_validation.json` records the input/output hashes and counts. Source descriptions are shown with source attribution, and all incident links retain their original OECD AIM or AIID URLs.

The right panel is reserved from the initial render. It initially displays an Overview with inventory counts, L1/L3 semantic color examples, and explanations of node size, curved links, the bridge ring, and zoom labels. Selecting a node replaces the Overview with its incidents without changing graph width. The Overview button restores the introduction.

## Selected-card centrality

The detail panel reports normalized degree centrality and exact normalized betweenness for the active channel. Both use the full undirected, unweighted channel network rather than the displayed overview. Degree centrality is distinct neighbors divided by n minus one. Betweenness excludes endpoints and uses the undirected normalization factor 2 divided by (n minus one) times (n minus two). Values switch with the source tab. Regenerate the added fields with `python scripts/add_channel_centrality.py`.

Node sizing uses observed incident counts with a common scale across channels. SVG radius is sqrt(6.25 + 317.75 * count / maximum count), where the maximum is taken over both full channel networks. Thus area above the minimum visible area is linear in counts. Counts represent retrieved incident records, not population prevalence. Centrality metrics and within-channel ranks remain available in the detail panel.

Display adjustment. AIID radii use a 1.5625 multiplier (125% of the preceding 1.25 display multiplier) for readability. AIM retains its count-based radius scale and starts with a modest left/up translation. Fit and reset restore these channel-specific defaults. Compare exact counts across channels in the detail panel rather than comparing areas directly.

## Annual network navigation

Run `python scripts/build_channel_years.py` to build `years.json` from frozen incident dates and the by-card record index. No retrieval is rerun. Each record retains three distinct cards, and both activation counts and co-activation pair counts sum to three times the annual record count. Year totals reconcile to 10,449 AIM and 1,751 AIID records. The histogram uses event years as stored in the snapshots; missing years have zero records. The default is all years. Selecting a year rebuilds display counts, the overview selection, and the source list from that year. Exact normalized centrality and competition ranks use the full annual channel graph. Global coordinates and node count-to-area scales remain fixed across years, apart from the previously documented AIM collision separation. The snapshot end years are partial years.

Risk-card tables, overview candidate selection, and automatic name-label priority use observed incident count in descending order, with card ID breaking ties. Centrality ranks retain their own metric definitions. Within a single card, source incident records remain newest first because each row is one incident. Annual bars remain chronological.

L4 color authority. The incidents renderer now reads each card color by exact ID from `data/semantic_space.json`, the same payload used by AI risk space. It does not generate an independent palette. This applies to all-years and annual views. Domain anchors remain distinct from boundary-blended semantic card shades.