diff --git a/app/web/main.py b/app/web/main.py
index 6f745ca..c1aae6c 100644
--- a/app/web/main.py
+++ b/app/web/main.py
@@ -13,13 +13,16 @@ from __future__ import annotations
 import importlib.util
 import json
 import os
+import re
 from functools import lru_cache
 from pathlib import Path
 from typing import Any, Dict, List, Optional, Tuple
 
+from app.web.plotly_parse import parse_plotly_run
+
 import httpx
-from fastapi import FastAPI, Form, Request
-from fastapi.responses import HTMLResponse, JSONResponse
+from fastapi import FastAPI, Form, HTTPException, Request
+from fastapi.responses import HTMLResponse, JSONResponse, Response
 from fastapi.staticfiles import StaticFiles
 from fastapi.templating import Jinja2Templates
 from sklearn.datasets import (
@@ -804,6 +807,32 @@ async def metrics_json() -> JSONResponse:
     return JSONResponse(_scan_metrics())
 
 
+_STEM_RE = re.compile(
+    r"^make_[A-Za-z_]+?_[A-Za-z]+_N\d+_T\d+_J[\d.]+_s\d+$"
+)
+
+
+@lru_cache(maxsize=32)
+def _cached_frames(stem: str) -> str:
+    """Parse <stem>.html and return the frames dict as a JSON string."""
+    html = FIGS_DIR / f"{stem}.html"
+    return json.dumps(parse_plotly_run(html), separators=(",", ":"))
+
+
+@app.get("/api/runs/{stem}/frames.json")
+async def run_frames(stem: str) -> Response:
+    if not _STEM_RE.match(stem):
+        raise HTTPException(400, f"malformed stem: {stem!r}")
+    html = FIGS_DIR / f"{stem}.html"
+    if not html.is_file():
+        raise HTTPException(404, f"no such run: {stem}")
+    try:
+        payload = _cached_frames(stem)
+    except Exception as e:
+        raise HTTPException(500, f"parse failed: {e}")
+    return Response(content=payload, media_type="application/json")
+
+
 @app.get("/health")
 async def health() -> JSONResponse:
     async with httpx.AsyncClient(timeout=3.0) as client:
diff --git a/app/web/plotly_parse.py b/app/web/plotly_parse.py
new file mode 100644
index 0000000..84d4561
--- /dev/null
+++ b/app/web/plotly_parse.py
@@ -0,0 +1,290 @@
+"""Parse flow-emitted Plotly HTML files into a compact per-frame JSON dict.
+
+The embedding-sandbox flow writes one standalone HTML per run at
+``figs/<stem>.html``. Each HTML contains a ``Plotly.newPlot(...)`` call (which
+holds the initial frame) and a ``Plotly.addFrames(...)`` call (which holds the
+full animation — including a redundant copy of frame 0). Point data is encoded
+with plotly's base64+dtype ("bdata") scheme, so we decode it with ``struct``.
+
+The module is stdlib-only.
+"""
+
+from __future__ import annotations
+
+import base64
+import json
+import re
+import struct
+from pathlib import Path
+
+
+_STEM_RE = re.compile(
+    r"^(?P<gen>make_.+?)_(?P<emb>[A-Za-z]+)_N(?P<n>\d+)_T(?P<t>\d+)"
+    r"_J(?P<j>[\d.]+)_s(?P<s>\d+)$"
+)
+
+# plotly's typed-array dtype -> (struct format char, item size bytes)
+_DTYPE_MAP = {
+    "i1": ("b", 1),
+    "u1": ("B", 1),
+    "i2": ("h", 2),
+    "u2": ("H", 2),
+    "i4": ("i", 4),
+    "u4": ("I", 4),
+    "i8": ("q", 8),
+    "u8": ("Q", 8),
+    "f4": ("f", 4),
+    "f8": ("d", 8),
+}
+
+
+def _decode_bdata(obj):
+    # Accepts either a plotly typed-array dict {"dtype","bdata",...} or a plain list.
+    if isinstance(obj, list):
+        return list(obj)
+    if not isinstance(obj, dict) or "bdata" not in obj:
+        raise ValueError(f"expected bdata object or list, got {type(obj).__name__}")
+    dt = obj.get("dtype")
+    if dt not in _DTYPE_MAP:
+        raise ValueError(f"unsupported dtype {dt!r}")
+    fmt, size = _DTYPE_MAP[dt]
+    raw = base64.b64decode(obj["bdata"])
+    count = len(raw) // size
+    return list(struct.unpack("<" + fmt * count, raw[: count * size]))
+
+
+def _extract_call_args(txt: str, call_name: str):
+    # Locate the LAST occurrence of `call_name(` in the file (plotly's JS
+    # bundle mentions Plotly.newPlot in docstrings/code earlier, but the real
+    # user-data call is always emitted at the very bottom).
+    start = txt.rfind(call_name + "(")
+    if start < 0:
+        raise ValueError(f"{call_name}( not found")
+    open_paren = txt.index("(", start)
+    i = open_paren + 1
+    depth = 1
+    in_str = False
+    escape = False
+    args: list[str] = []
+    cur = i
+    n = len(txt)
+    while i < n:
+        c = txt[i]
+        if in_str:
+            if escape:
+                escape = False
+            elif c == "\\":
+                escape = True
+            elif c == '"':
+                in_str = False
+        else:
+            if c == '"':
+                in_str = True
+            elif c in "([{":
+                depth += 1
+            elif c in ")]}":
+                depth -= 1
+                if depth == 0:
+                    args.append(txt[cur:i].strip())
+                    return args
+            elif c == "," and depth == 1:
+                args.append(txt[cur:i].strip())
+                cur = i + 1
+        i += 1
+    raise ValueError(f"unterminated {call_name} arglist")
+
+
+def _parse_stem(stem: str) -> dict:
+    m = _STEM_RE.match(stem)
+    if not m:
+        raise ValueError(f"stem does not match expected pattern: {stem!r}")
+    return {
+        "stem": stem,
+        "generator": m.group("gen"),
+        "embedder": m.group("emb"),
+        "num_points": int(m.group("n")),
+        "num_timesteps": int(m.group("t")),
+        "jitter_scale": float(m.group("j")),
+        "seed": int(m.group("s")),
+    }
+
+
+def _trace_xy_by_id(trace: dict) -> tuple[list[int], list[float], list[float]]:
+    # Prefer explicit `ids`; fall back to customdata (which in this flow also
+    # carries the point id in column 0); finally, fall back to positional index.
+    if "ids" in trace:
+        ids = _decode_bdata(trace["ids"])
+        # ids may arrive as strings too (plotly coerces). If they're strings
+        # of digits, we leave them as-is but prefer ints when possible.
+        if ids and isinstance(ids[0], str) and ids[0].isdigit():
+            ids = [int(v) for v in ids]
+    elif "customdata" in trace:
+        cd = trace["customdata"]
+        if isinstance(cd, dict):
+            ids = _decode_bdata(cd)
+        else:
+            # shape (N, k): pick first column
+            ids = [row[0] if isinstance(row, (list, tuple)) else row for row in cd]
+    else:
+        ids = list(range(len(_decode_bdata(trace["x"]))))
+    xs = _decode_bdata(trace["x"])
+    ys = _decode_bdata(trace["y"])
+    if not (len(ids) == len(xs) == len(ys)):
+        raise ValueError(
+            f"trace length mismatch: ids={len(ids)} x={len(xs)} y={len(ys)}"
+        )
+    return ids, xs, ys
+
+
+def _frame_xy_ordered(traces: list[dict], point_ids: list[int]):
+    # Merge possibly multiple traces (one per label-group) back into a single
+    # (x, y) pair ordered by point_ids. Ids missing from this frame — points
+    # added or removed between frames by the jitter_add_remove path — become
+    # `None` so downstream code can skip them.
+    by_id: dict[int, tuple[float, float]] = {}
+    for t in traces:
+        ids, xs, ys = _trace_xy_by_id(t)
+        for pid, x, y in zip(ids, xs, ys):
+            by_id[pid] = (x, y)
+    xs_out = [by_id[pid][0] if pid in by_id else None for pid in point_ids]
+    ys_out = [by_id[pid][1] if pid in by_id else None for pid in point_ids]
+    return xs_out, ys_out
+
+
+def _labels_for_ids(traces: list[dict], point_ids: list[int]) -> list[str]:
+    # If traces carry distinct names/legendgroups, use that as the per-point
+    # class label. When all traces share an empty name (the single-trace case
+    # emitted by this flow), labels are uniformly "".
+    names = [t.get("name") or t.get("legendgroup") or "" for t in traces]
+    if len(traces) <= 1 or all(n == names[0] for n in names):
+        return ["" for _ in point_ids]
+    by_id: dict[int, str] = {}
+    for t, name in zip(traces, names):
+        ids, _, _ = _trace_xy_by_id(t)
+        for pid in ids:
+            by_id[pid] = name
+    return [by_id.get(pid, "") for pid in point_ids]
+
+
+def parse_plotly_run(html_path) -> dict:
+    """Parse a flow-emitted plotly HTML into a frames dict suitable for
+    the three.js comparison page. Raises ValueError on unrecognised shape."""
+    path = Path(html_path)
+    stem = path.stem
+    meta = _parse_stem(stem)
+
+    txt = path.read_text(encoding="utf-8")
+
+    new_args = _extract_call_args(txt, "Plotly.newPlot")
+    if len(new_args) < 3:
+        raise ValueError(f"Plotly.newPlot expected >=3 args, got {len(new_args)}")
+    initial_traces = json.loads(new_args[1])
+    layout = json.loads(new_args[2])
+
+    add_args = _extract_call_args(txt, "Plotly.addFrames")
+    if len(add_args) < 2:
+        raise ValueError(f"Plotly.addFrames expected 2 args, got {len(add_args)}")
+    raw_frames = json.loads(add_args[1])
+    if not isinstance(raw_frames, list) or not raw_frames:
+        raise ValueError("Plotly.addFrames second arg is not a non-empty list")
+
+    # Establish the stable id set as the UNION of ids across every frame —
+    # the flow's jitter_add_remove path intentionally adds/removes points
+    # between frames, so no single frame is authoritative. Order is: first
+    # appearance across frames, then by numeric id within each frame's new ids.
+    seen_ids: list[int] = []
+    seen_set: set = set()
+    for fr in raw_frames:
+        for t in fr.get("data") or []:
+            ids, _, _ = _trace_xy_by_id(t)
+            for pid in ids:
+                if pid not in seen_set:
+                    seen_set.add(pid)
+                    seen_ids.append(pid)
+    if not seen_ids:
+        raise ValueError("no point ids recovered from any frame")
+
+    first_frame_traces = raw_frames[0].get("data") or initial_traces
+    labels = _labels_for_ids(first_frame_traces, seen_ids)
+
+    frames_out = []
+    times = []
+    for i, fr in enumerate(raw_frames):
+        traces = fr.get("data")
+        if not traces:
+            raise ValueError(f"frame {i} has no 'data'")
+        xs, ys = _frame_xy_ordered(traces, seen_ids)
+        frames_out.append({"x": xs, "y": ys})
+        times.append(fr.get("name", str(i)))
+
+    # Global bounds across all frames (skipping None placeholders for
+    # points that only exist in some frames).
+    all_x = [v for fr in frames_out for v in fr["x"] if v is not None]
+    all_y = [v for fr in frames_out for v in fr["y"] if v is not None]
+    if not all_x or not all_y:
+        raise ValueError("no finite x/y values across frames")
+    xmin, xmax = min(all_x), max(all_x)
+    ymin, ymax = min(all_y), max(all_y)
+
+    title = ""
+    try:
+        title = layout["title"]["text"] if isinstance(layout.get("title"), dict) else str(layout.get("title") or "")
+    except Exception:
+        title = ""
+    meta["title"] = title
+
+    return {
+        "meta": meta,
+        "point_ids": seen_ids,
+        "labels": labels,
+        "times": times,
+        "frames": frames_out,
+        "bounds": {"x": [xmin, xmax], "y": [ymin, ymax]},
+    }
+
+
+if __name__ == "__main__":
+    import sys
+    import traceback
+
+    figs_dir = Path("/home/mm/work/dr-sandbox/figs")
+    html_files = sorted(figs_dir.glob("*.html"))
+    if not html_files:
+        print("no .html files in", figs_dir)
+        sys.exit(1)
+
+    failures = []
+    for p in html_files:
+        try:
+            out = parse_plotly_run(p)
+        except Exception as e:
+            failures.append((p.name, str(e)))
+            print(f"FAIL  {p.name}: {e}")
+            traceback.print_exc()
+            continue
+        m = out["meta"]
+        pids = out["point_ids"]
+        frames = out["frames"]
+        nT = len(frames)
+        f0 = frames[0]
+        f0x = [v for v in f0["x"] if v is not None]
+        f0y = [v for v in f0["y"] if v is not None]
+        xr = (min(f0x), max(f0x)) if f0x else (float("nan"), float("nan"))
+        yr = (min(f0y), max(f0y)) if f0y else (float("nan"), float("nan"))
+        consistent = all(len(fr["x"]) == len(pids) and len(fr["y"]) == len(pids) for fr in frames)
+        present_per_frame = [sum(v is not None for v in fr["x"]) for fr in frames]
+        present_rng = (min(present_per_frame), max(present_per_frame))
+        print(
+            f"OK    {m['stem']}  |ids|={len(pids)} T={nT}  "
+            f"present/frame=[{present_rng[0]},{present_rng[1]}]  "
+            f"f0 x=[{xr[0]:+.3f},{xr[1]:+.3f}] y=[{yr[0]:+.3f},{yr[1]:+.3f}]  "
+            f"consistent={consistent}"
+        )
+
+    print()
+    if failures:
+        print(f"{len(failures)} failure(s):")
+        for name, msg in failures:
+            print(f"  {name}: {msg}")
+    else:
+        print(f"all {len(html_files)} files parsed OK")