Threat-model prompt pack release notes

2.0

22 August 2026 · commit 36cadae

Method coverage

• P-stride now writes cumulative batches with a row budget. It must list expected, considered, and remaining elements, and cannot advance while coverage remains.
• An empty LLM subset now has an explicit decision. PHANTOM-B is marked not applicable after operator confirmation; it is never reported as complete on an empty set.
• Repeated components can inherit analysis only when the diagram shows a scaling group or managed pair. Each pair still requires a configuration-divergence threat.

Evidence and actions

• Existing controls now state which inventory ids the diagram shows them protecting and what diagram relation supports that coverage.
• Control absences record that an expected control is not shown at a named referent. They do not claim the deployed system lacks the control.
• Threat evidence references resolve to control-absence ids. Mitigate and eliminate actions bind to a component, store, flow, or trust-boundary id.
• Role records now state what each role notices first and which claims it declines to infer.

Chain and exports

• P-rank is an optional builder review order based on attacker position and explicit preconditions. It does not claim likelihood, impact, or residual risk.
• Optional Track B now branches after P-qa and returns to P-report. P-report is the only step that writes report.markdown.
• Markdown and JSON exports serialize the final matrix without rewriting the report.
• The CSV header keeps SRF columns for both tracks and adds review_position and evidence_refs.
• The pack root now carries the scoring dimensions and the rule that automated and human scores are not averaged.

Compatibility

• Version 2.0 is a schema and workflow break from 1.x. A 1.x JSON file may fail the current schema because it lacks method-coverage, applicability, evidence, or replica fields.
• The v2 schema requires chain_meta.prompt_pack_version to equal 2.0. Keep an older report with its original pack file when audit reproduction matters.
• CSV import maps made for 1.3 or 1.4 need two new columns: review_position after threat_id and evidence_refs after preconditions.
• The report route changed. Run optional Track B after P-qa, then run P-report and the four exports once.
• Version 2.0 also fixes report and JSON templates that wrote prompt_pack_version as 1.2 instead of reading the pack root version.

1.4

21 August 2026 · commit 289e6f3

• Added P-export-diagram as the fourth regular export.
• The export writes Mermaid flowchart source to save as a .mmd file.
• Threat ids are placed on their existing diagram referents; the export cannot invent inventory nodes.

1.3

21 August 2026 · commit 44b7c84

• Added P-export-csv with one RFC 4180 row per threat.
• The CSV kept SRF columns when Track B was skipped so one import map could accept either track.
• Added a one-chat shortcut that fetches prompts.json and runs the chain.

1.2

21 August 2026 · commit eba6ff6

• Added attacker assumptions, attacker positions, existing controls, claim boundaries, and action validation to the chain.
• Added P-report so a reviewer could read the threat model without parsing the JSON.
• Added markdown and completed-JSON export prompts.
• Left reviewer fields empty for a human sign-off.

1.0

21 August 2026 · commit 1c27f59

• Added the diagram-normalization and solution-description chain for image, Mermaid, and SVG inputs.
• Applied STRIDE to in-scope architecture elements and PHANTOM-B to the LLM subset.
• Added Track B for SRF layer, persona, and party assignments.
• Added evaluation baselines and a JSON Schema for stored results.
• A same-version maintenance commit added chain next-step labels and explicit JSON filenames.