Skip to content

Replace legacy VM execution with fail-closed Isogloss - #6

Open
owengregson wants to merge 55 commits into
mainfrom
codex/traveling-isogloss-replacement
Open

Replace legacy VM execution with fail-closed Isogloss#6
owengregson wants to merge 55 commits into
mainfrom
codex/traveling-isogloss-replacement

Conversation

@owengregson

Copy link
Copy Markdown
Owner

What changed

  • Replaced the shipped VM transform with guarded Isogloss source execution. Accepted pure return regions are lowered into scalarized 3×3 BPRF realizations with exact runtime domain guards; configured failures stop the build and never retain the original relation as fallback.
  • Physically removed the VM runtime, interpreter/loader, bytecode encoder and encryption backend, opcode mutation/shuffle infrastructure, legacy presets/options, manifests, benchmarks, and execution-specific tests.
  • Added canonical semantic IR/CFG, effect-delimited region planning, sound direct-call evidence, exact pure-region lowering, and constructive black-box learnability upper-bound gates.
  • Added statefully masked moving-cover custody, fixed 4/8/16/32 transcript buckets, keyed rejection-sampled placement, owner/client schema separation, and fail-closed product planning for local, custodied, private-function, and attested profiles.
  • Cut the public API, package identity, CLI, browser worker, playground, website copy, and README over to Isogloss. Nonlocal profiles require a real owner-proven boundary and forbid a complete local fallback.

Threat-model constraints

holographic-local is intentionally reported as client-complete with hardnessLowerBound: null. It raises analysis cost but does not claim secrecy against unrestricted execution/instrumentation. The known exact degree-9/three-input oracle attack remains 220 queries. Custodied/private/TEE modes are admitted only through the owner planner with the required real boundary/capabilities.

The custody attack harness confirms corrected hidden placement is not empirically distinguishable above its declared baselines (46.35% full classifier vs 50% binary chance; 25.78% stage-count classifier vs 25% chance), while honestly retaining declared bucket leakage.

Root causes fixed during cutover

  • Late discovery of a source input shifted already-emitted step references. The compiler now freezes the complete ingress prefix before assigning step references, with a permanent multi-stage reproducer.
  • The public planner imported a server/reference custody module just to obtain transcript bucket constants. Bucket classes now live in a dependency-free product module, and the package-entry isolation test proves reference evaluators/custodians remain unreachable.

Validation

  • bun test: 179 passed, 0 failed, 34,277 assertions
  • bun run typecheck: passed
  • package bun run build: passed, including declarations
  • browser worker rebuild: passed
  • web bun run typecheck && bun run build: passed, static export generated
  • built package API/CLI smoke tests: passed
  • npm pack --dry-run: 6 expected files, 59.8 kB tarball
  • full local benchmark: checksum agreement; 25.061 ms build, 211.69× size expansion, 56.80× steady-state runtime overhead for the guarded kernel (engineering cost only, not a hardness claim)

owengregson and others added 30 commits July 24, 2026 11:43
Design doc + runnable launch packet for an Opus 5 / Fable 5 parallel-workflow
ideation campaign that forces novelty by construction: forbidden-solution
lists, assumption inversion, cross-domain analogy seeding, and forced fusion.
Scope is the whole next-gen tool (seven tracks); blue-sky generation with a
separate feasibility/novelty tagging pass. Agent-facing prompts use neutral
product/IP-protection framing. Plan only; the workflow is opt-in.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Copilot AI review requested due to automatic review settings July 25, 2026 11:39

Copilot AI left a comment

Copy link
Copy Markdown

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Copilot wasn't able to review this pull request because it exceeds the maximum number of lines (20,000). Try reducing the number of changed lines and requesting a review from Copilot again.

@github-actions

github-actions Bot commented Jul 25, 2026

Copy link
Copy Markdown

Dependency Review

✅ No vulnerabilities or license issues or OpenSSF Scorecard issues found.

Scanned Files

  • .github/workflows/ci.yml

Copilot AI review requested due to automatic review settings July 25, 2026 11:43

Copilot AI left a comment

Copy link
Copy Markdown

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Copilot wasn't able to review this pull request because it exceeds the maximum number of lines (20,000). Try reducing the number of changed lines and requesting a review from Copilot again.

Copilot AI review requested due to automatic review settings July 25, 2026 11:45

Copilot AI left a comment

Copy link
Copy Markdown

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Copilot wasn't able to review this pull request because it exceeds the maximum number of lines (20,000). Try reducing the number of changed lines and requesting a review from Copilot again.

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

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

2 participants