diff --git a/proposals/manifold-ci-benchmarking-discovery.md b/proposals/manifold-ci-benchmarking-discovery.md
new file mode 100644
index 0000000..06a117a
--- /dev/null
+++ b/proposals/manifold-ci-benchmarking-discovery.md
@@ -0,0 +1,115 @@
+# Manifold CI + Benchmarking Discovery and Proposal
+
+This draft is written so it can be posted directly to the issue thread as the “summary + proposal” requested by maintainers.
+
+## 1) Current-state summary (what to inventory in `elalish/manifold`)
+
+> Note: this repository (`opencax/GSoC`) only contains idea listings, so the concrete CI/benchmark status must be collected in the `elalish/manifold` repo.
+
+### CI inventory checklist
+
+- Enumerate all workflows in `.github/workflows/` and list:
+  - Trigger type (`push`, `pull_request`, schedule, manual).
+  - OS matrix and compiler matrix.
+  - Build types (`Debug`/`Release`) and feature flags.
+  - Test coverage (unit/integration/fuzz/examples).
+- Capture gaps against the issue request:
+  - Are sanitizers present? (ASan/UBSan/TSan).
+  - Are deterministic-cross-platform checks present?
+  - Are performance regressions tracked in CI?
+
+### Benchmark inventory checklist
+
+- Identify any existing benchmark binaries/scripts and where they run.
+- Identify storage strategy for results (if any):
+  - Artifacts only, committed JSON, external dashboard, or none.
+- Check if there is historical trend visibility today.
+
+## 2) Proposed implementation plan
+
+### Phase A — Strengthen correctness CI
+
+1. **Add sanitizer jobs (initially Linux):**
+   - New workflow job(s) to build + run tests with `-fsanitize=address,undefined`.
+   - Keep runtime bounded by selecting representative test targets.
+   - Upload sanitizer logs as artifacts.
+
+2. **Add determinism check across platforms:**
+   - Define canonical benchmark/test inputs.
+   - Emit normalized machine-readable outputs (JSON/text with stable formatting).
+   - Compare outputs across Linux/macOS/Windows jobs:
+     - Either strict byte compare after normalization.
+     - Or tolerance-based comparison for floating-point fields.
+
+3. **Guardrails for flaky checks:**
+   - Separate deterministic checks from perf checks.
+   - Add retry only where external dependencies are involved.
+
+### Phase B — Continuous benchmarking in GitHub Actions
+
+1. **Benchmark workflow design:**
+   - Trigger on `push` to main + manual dispatch + nightly schedule.
+   - Run on fixed runner type to reduce noise.
+   - Persist raw benchmark output as versioned JSON artifacts.
+
+2. **Baseline + comparison strategy:**
+   - Compare PR branch result against:
+     - latest `main`, and/or
+     - rolling baseline (median of last N runs).
+   - Compute percent change and flag regressions using configurable thresholds.
+
+3. **Data publication path:**
+   - Option A: store JSON history in a `gh-pages` branch.
+   - Option B: push to a dedicated benchmark-results repo.
+   - Option C: keep artifacts only (lowest effort, weakest UX).
+
+### Phase C — Frontend benchmark dashboard
+
+1. **Frontend scope (simple first version):**
+   - Table + sparkline/trend chart per benchmark.
+   - Filters by OS/compiler/config/date range.
+   - Regression badges (green/yellow/red).
+
+2. **Tech choices:**
+   - If manifold already has web docs tooling: integrate there.
+   - Otherwise static HTML+TypeScript page fed by JSON.
+
+3. **MVP delivery criteria:**
+   - A maintainer can open one URL and inspect trends/regressions.
+   - Data refreshes automatically from CI outputs.
+
+### Phase D — Port OpenSCAD-style benchmark cases
+
+1. Select representative workloads:
+   - Boolean-heavy solids.
+   - Deep CSG trees.
+   - Mesh conversion/stress cases.
+2. Port input generators and normalize runtime parameters.
+3. Tag benchmarks by category and expected complexity.
+
+## 3) Concrete first-week tasks
+
+- [ ] Post CI/benchmark inventory table from the current manifold repo.
+- [ ] Open PR 1: add Linux ASan+UBSan workflow job.
+- [ ] Open PR 2: add benchmark workflow that uploads JSON artifacts.
+- [ ] Open PR 3: scaffold dashboard page reading those JSON files.
+- [ ] Open PR 4: port first 3–5 OpenSCAD-derived benchmark scenarios.
+
+## 4) Suggested comment to post in the issue thread
+
+I reviewed the project direction and propose starting with a discovery pass, then delivering in small PRs:
+
+1. Inventory existing manifold workflows/tests/benchmarks and map gaps (sanitizers, deterministic checks, historical perf tracking).
+2. Add Linux ASan+UBSan CI job first (fast feedback, low risk).
+3. Add a scheduled benchmark workflow producing JSON artifacts and PR-vs-main regression summaries.
+4. Publish those JSON results into a lightweight dashboard (static TS/HTML) with trend charts and regression badges.
+5. Port a first set of OpenSCAD-style workloads (boolean-heavy, deep CSG, mesh stress) into the benchmark suite.
+
+If this direction sounds good, I can start by posting the CI/benchmark inventory table and then open the sanitizer PR as the first concrete step.
+
+## 5) Risks and mitigations
+
+- **CI runtime growth:** keep sanitizer/benchmark jobs in separate workflows and use selective triggers.
+- **Benchmark noise:** pin runner type, run multiple iterations, compare medians.
+- **Cross-platform float drift:** normalize outputs and apply field-specific tolerances.
+- **Dashboard maintenance cost:** start with static rendering and JSON schema versioning.