Skip to content

feat: deliver Sprint 6 transfer evaluation and realism views - #27

Merged
mohamed2ze merged 2 commits into
mainfrom
sprint6/transfer-realism
Nov 18, 2026
Merged

mohamed2ze merged 2 commits into
mainfrom
sprint6/transfer-realism

Conversation

@mohamed2ze

Copy link
Copy Markdown
Collaborator

What

  • Add offline surrogate-crafted FGSM/PGD/post-projected CW transfer evaluation and an independent matched target white-box comparator.
  • Bind distinct source/target models, shared preprocessing/partitions, exact cohorts, budgets, seeds and metric denominators; keep candidate transfer explicitly gated.
  • Add read-only transfer/threat-context API, bilingual v0.6.0 threat/realism page, source/target journey decisions, proposed version-pinned mappings and generated evidence table.

Why

Deliver Sprint 6's research, FastAPI/dashboard and bilingual journey work packages without representing exploratory or synthetic results as frozen-benchmark acceptance.

The owner explicitly assigned training and all model/data downloads to a colleague. No local training on EMBER, dataset/model downloads, real transfer measurements, official-test evaluation, live malware or problem-space lab was performed. Existing measured white-box playback is retained; black-box measurements stay pending.

How tested

  • make lint
  • make test via final make check: 618 passed, 6 expected optional skips
  • make coverage: 618 passed, 6 skipped; 90% combined coverage
  • make security: no known audited dependency vulnerabilities; no medium/high Bandit findings
  • Synthetic CPU ART transfer tests for all three families, weights-only bundle loading, source-only crafting, inference-only target evaluation, forged/incompatible partitions, zero controls, malformed artifacts, API/UI states
  • Architecture/docs/hygiene focused checks; JavaScript syntax; task-file pre-commit hooks
  • Real MNIST demo: not needed; committed measurements unchanged

Browser: desktop recorded playback and English/Egyptian-Arabic/RTL text checked; 958px document had no page overflow. One next-step keyboard progression/focus was observed. Subsequent automated keyboard/screenshot interactions were inconclusive/unavailable; full keyboard and mobile viewport review remains a gap. Node component tests are not browser/mobile proof.

All-files pre-commit detected a preexisting missing EOF newline in the unrelated historical plan. Restored that hook-only change; scoped task hooks pass. No hook configuration or check was weakened.

Checklist

  • Conventional Commits
  • No datasets/weights/.env/mlruns/arrays committed
  • Implementation security boundaries inspected; existing prediction/freeze gates unchanged
  • Docs/decisions/progress updated and discoverable
  • Deployment triggers inspected: existing CI only, image publishing disabled; no deployment
  • Research acceptance: M1, real transfer measurements, technical policy/mapping review and problem-space evidence remain pending

Evidence: docs/sprint6.md, docs/sprint6_report.md, docs/sprint6_mappings.md, generated docs/results/sprint6_comparison.md.

@mohamed2ze
mohamed2ze merged commit a529a7b into main Nov 18, 2026
3 checks passed
@mohamed2ze
mohamed2ze deleted the sprint6/transfer-realism branch November 18, 2026 12:52
@mohamed2ze
mohamed2ze restored the sprint6/transfer-realism branch November 18, 2026 12:58
@mohamed2ze
mohamed2ze deleted the sprint6/transfer-realism branch November 18, 2026 13:01
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.

1 participant