Skip to content

Research: Frame neuro-symbolic SAST as a proof/evidence-tiering donor for OwnIR (kill-first) #460

Open

Description

@PhysShell

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

  1. 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.
  2. 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.
  3. 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.
  4. 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

Bounded investigation

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

  1. PhysShell commented on Nov 15, 2026

    @PhysShell
    OwnerAuthor

    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 via uvx frame-mcp can 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. LLM llm_verified means sink matches Frame's sink model, not end-to-end exploitable path. A clean scan does not prove safety, and frame 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.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions