Repository navigation
Description
Classification
Bounded prior-art research / P2 backlog / read-only first / NO production integration authorized.
Source: LambdaSec Frame, architecture & evaluation (2026-07-02).
Question: Can Frame's separation-logic/SMT engine, LLM hypothesis layer and evidence-tiering design reveal an independently demonstrable correctness or evidence gap in Own.NET's Roslyn → OwnIR → Rust ownership/effects/aliasing pipeline?
Frame is SAST with taint, frameworks, authorization and vulnerability rules, not an ownership/Dispose checker or interchangeable OwnIR backend. Review architecture, not marketing or drop-in adoption.
The specific ideas to evaluate
- Proof boundary and provenance: classify hypotheses, recognized-sink grounding, modeled path evidence, solver verdicts, runnable reproduction and fix verification separately. A sink recognized by the engine is not automatically a proof that the whole hypothesized execution path is feasible.
- Separation logic / bi-abduction / frame inference: identify exactly which heap/resource obligations its core encodes and how unsupported code, unknown solver results, summaries and aliases are handled. Compare to Own.NET's existing ownership facts and transfer semantics, not name similarity.
- Independent witness: consider whether a tiny test-only validation of a known ownership diagnostic could produce a reproducible counterexample or meaningful proof assumption. No blanket trust in a 'verified' field.
- LLM as proposal only: do not allow a model's discovery/triage to upgrade Own.NET diagnostic certainty or suppress a warning without independent, explicit evidence.
Related Own.NET work; avoid duplication
- Research: Doop/Soufflé-style relational points-to and alias analysis (bounded spike) #458 Doop/Soufflé points-to (relational alias analysis), Research: COVA-inspired targeted path-condition oracle for ownership soundness (bounded spike) #431 independent COVA path-condition oracle.
- Research: FORVES2-style conditional semantic validation and proof-assumption audit for OwnIR #445 semantic-assumption audit; Research: evaluate Bifrost for C# code intelligence, agent navigation, and evidence contracts #426 evidence/completeness; Research: AI-assisted ownership-analysis setup with auditable assumptions (TrustInSoft / Microsoft) #452 auditable proof assumptions; Research: semantic mutation testing harness for Own.NET analyzer coverage (bounded pilot) #405 mutation harness.
- Research: requirements-to-code conformance checking (BDD + reviewed formal contracts + OwnIR evidence) #448 requirements-to-code conformance (distinct consumer; possible Assurance-as-Code connection).
Bounded investigation
- Pin Frame commit, license, relevant solver/SL-COMP materials, documented C# coverage, exact evidence-tier schema and command semantics. Inspect code paths rather than taking README 'sound' at face value.
- Map two real Own.NET failure or uncertainty cases (e.g., alias-dependent conditional release, interprocedural dispose/transfer) to Frame's modeled abstractions; document what is not representable.
- Define a matrix: claimed property | encoded semantics | assumptions | UNKNOWN/timeout | executable witness | independent oracle. Keep proved within a model, tested, LLM-proposed, and unresolved distinct.
- If a narrow test-only experiment looks feasible, preregister 3–5 positive/negative fixtures with independent oracle, cost/complexity budget, and no changes to production diagnostics; execute only after separate GO.
- Compare measured incremental value against Research: COVA-inspired targeted path-condition oracle for ownership soundness (bounded spike) #431/Research: FORVES2-style conditional semantic validation and proof-assumption audit for OwnIR #445/Research: Doop/Soufflé-style relational points-to and alias analysis (bounded spike) #458. If existing mechanisms already cover the cases, KILL.
Caveat on upstream numbers
Frame's paper reports recall 0.67 / precision 0.51 on a pooled, model-judged corpus, with Frame contributing to the pool. These are not verified Own.NET benchmark numbers and do not establish general semantic soundness.
Deliverable / decision
A short source-backed KEEP / EXPERIMENT / KILL memo, pinned links, gap map, candidate minimal fixture(s), independent-oracle plan, and explicit non-goals. No automatic dependency, implementation, CI gate or 'proven safe' claims.
Activity
PhysShell commentedon Nov 15, 2026
Follow-up: LambdaSec agent skill and MCP front end for Frame (2026-11-15)
Upstream: LambdaSec Skills, frame-security-scan/SKILL.md, Frame MCP.
Concrete command interface for a potential test-only trial:
frame scan … -f json/ SARIF,frame verify findings.json,--changed REF,--baseline FILE, optional--hunt-authz, optional--ai. The skill has explicit approval gates for hosted LLM code exfiltration, in-place fixes and exploit workflows; the plugin MCP path viauvx frame-mcpcan auto-install on launch. Do not enable or install through this issue.Keep this a C# SAST benchmark, not an ownership oracle. Frame says symbolic frontends cover C#, whereas Go gets only LLM detection with
--ai. LLMllm_verifiedmeans sink matches Frame's sink model, not end-to-end exploitable path. A clean scan does not prove safety, andframe verify'gone' is not evidence of fixed program semantics; run independent tests if later approved.For #460's evidence-tiering investigation, first inspect JSON provenance/UNKNOWN statuses from pinned versions on synthetic examples. Evaluate as an alternative to or supplement for the existing proof-assumption and path-oracle work. No production scanner, dependency or fix workflow now.