Threat model backtest

Threat Model — Backtest (validate against history)

Phase 3.6. A draft that has never been backtested is untested software. The project's own history is a free test suite; use it before presenting the draft for sign-off. This is a producer-side quality gate — the corpus does not go into the published document (per the leave-out list: CVE history is not the threat model).

Procedure

1. Assemble a stratified corpus — start with the last 10–30 inbound security findings: published advisories, reports closed as "not a bug" / "by design", issues labeled security, and scanner/fuzzer/AI-analysis output. If a vendored security-context.md is present in the working directory (a runner's pre-fetch of exactly this material), seed the corpus from it before searching elsewhere. Prefer contested items, but do not let recency or controversy leave component families or contract dimensions untested.

   • Cover every in-scope component family.
   • Cover every applicable contract-dimension row: numeric limits, failure atomicity, topology, callbacks, serialization, lifecycle, concurrency, and resource complexity.
   • For a large corpus, cluster by (component, sink, attack class, required attacker capability) and route at least one representative from every cluster. Increase the corpus beyond 30 when necessary to avoid an untested cluster; report both item count and cluster count.
   • Record each item's actual historical outcome where one exists — fixed, wontfix, by-design, out-of-scope, or unknown — with the advisory or issue URL it came from. That label is the ground truth step 3 scores against, and without it the backtest cannot fail. Set the outcome aside while routing (step 2 is blind); compare only afterwards.
   • When no historical record is reachable, say so rather than inventing one. Synthesize cases to exercise the matrix, mark every one synthesized, and write the §1.1 note verbatim: "no historical corpus was available; the backtest routed N synthesized cases only." A self-invented corpus reported as history is worse than no backtest, because it reads as evidence.

2. Route each item blind — using only the draft (not hindsight knowledge of how it was actually resolved), apply the §1.1 triager quick-start and assign exactly one §1.17 disposition, citing the licensing section. Routing rules and the closed disposition set are in output-structure.md.

3. Score the routing. The two directions of error are not symmetric. Wrongly closing a real vulnerability is far worse than wrongly escalating a non-finding: an over-escalating model wastes maintainer time, an over-closing one hands a reporter a "not a bug" on a live issue. Score accordingly — this asymmetry decides every fix below.

   Signature	Meaning	Fix
   Closes an item the project actually fixed	the model closes a true positive — the one disqualifying outcome	Blocking. Fix before sign-off. Narrow the licensing §1.12 disclaimer, §1.7 trusted marking, or §1.3 scope line until the item routes VALID or escalates. Never widen a disclaimer to reach a close.
   Routes to MODEL-GAP	the model is missing a decision	propose, in this order: an unresolved matrix row plus a proposed-answer §1.18 question; a conditional §1.11 guarantee; or an explicit §1.12 disclaimer. Prefer the option that leaves the report escalating over the option that closes it
   Routes plausibly to two or more dispositions	two sections overlap or contradict	sharpen them until the routing is unique
   Routes to a disposition that contradicts how the maintainer actually resolved it	the model is wrong, or the historical call was	high-value §1.18 question — do not paper over it

   Disclaiming is the cheapest way to make a MODEL-GAP disappear, which makes it the easiest way to pass this gate while making the model worse. A disclaimer added because a corpus item routed badly is reverse-engineered from the answer: it must still be true of the project as it is, cite a real source, and stay inside the scope that source covers.

4. Feed §1.15 — a recurring pattern is a candidate known non-finding only when its final outcome closes as BY-DESIGN: property-disclaimed or an already-established KNOWN-NON-FINDING. Never promote VALID-HARDENING or MODEL-GAP into a non-finding, and never promote an OUT-OF-MODEL:* route: a pattern that closes because the code is out of scope, the build is unsupported, or the root cause sits in a dependency keeps that disposition. KNOWN-NON-FINDING is first in the precedence order, so relabelling one of those routes promotes it above the very checks that decided it. Require documented/maintainer provenance and record the exact component or sink, symptom/attack class, preconditions, and stable claim or obligation IDs that discharge it; textual resemblance alone is not a match. Every candidate must satisfy the four §1.15 rules in output-structure.md — in particular, no entry may match on the reporter's evidence (no reproducer, no demonstrated reachability) rather than on the behaviour of the code.

5. Export 2–4 worked routing examples (§1.11) — the second and last thing the corpus may contribute to the document. Pick items that show the routing algorithm working, de-identify them, and give each one line: sink, required attacker capability, symptom, disposition, licensing claim. At least one must route VALID. A triager who only ever sees closes learns that the model's job is to say no; one worked VALID shows where the project takes responsibility, and it is the example that makes the rest credible. Budget about 15 lines. No CVE IDs, no reporter names, no dates — those are corpus content and stay on the producer side per the leave-out list.

   Steps 4 and 5 are the only backtest corpus items that may enter the document. The §1.1 note carries aggregate figures only, never an item.

6. Close the coverage loop — update the contract-dimension matrix after every revision, then reroute the affected cluster. An accepted model has no unexplained applicable cell and no unowned MODEL-GAP; an unratified draft may retain gaps only when each is represented by an unresolved row and §1.18 question.

Record the result

The §1.1 note is the only part a reader sees, so it must state what the backtest actually proved rather than that it happened. Report, in one short paragraph:

• Corpus size and shape — item count, cluster count, and contract-dimension coverage.
• Provenance split — how many items carry a real historical outcome versus how many were synthesized. When none are real, use the verbatim sentence from step 1. "Routed 22 findings" reads as history; say which of them was.
• Disposition histogram — how many items landed on each §1.17 disposition.
• The fail-safe figure — how many historically-fixed items route to a closing disposition. The target is zero, and a shortfall is the one number that blocks sign-off.
• How much the model closes — the share of the corpus that closes outright. A model that closes nearly everything is either exceptionally well documented or quietly over-disclaiming; say which.
• Contradictions, with pointers to the §1.18 questions they raised.

Example: "Backtested 22 findings (14 clusters, all 8 applicable dimensions); 17 carry a real historical outcome, 5 synthesized. Routed 9 VALID, 6 BY-DESIGN: property-disclaimed, 5 OUT-OF-MODEL, 2 escalated, 0 MODEL-GAP. All 11 historically-fixed items routed VALID. 50% of the corpus closes. Two contradictions raised Q7 and Q9."

Do not publish the corpus itself.

Output

Write the routing table to .threat-model/backtest.md — a producer-side artifact, deliberately in a dot-directory rather than beside threat-model.md. Every row carries a real advisory or issue URL, which the leave-out list keeps out of the published model, so a table sitting next to the deliverable gets published by the first git add -A. Add .threat-model/ to the project's .gitignore if it is not already covered. One row per item:

id | source | component | cluster | dimension | disposition | licensing § | historical outcome | pass/fail

where source is a real advisory or issue URL, or the literal synthesized. Handing the table back only in conversation materializes nothing a reviewer can check afterwards, which is how a backtest comes to self-certify.

Also hand back to the orchestrator: the coverage report, the model decisions/revisions triggered (for threat-model-authoring to apply), the §1.18 questions for unresolved gaps, the §1.15 candidates, the §1.11 worked routing examples, and the header backtest note.