[I07] Access check for the canonical rerun (PhysioNet + HF gate)

[duckyquang]

Task metadata

Field	Value
Task ID	I07
Category	Infrastructure
Priority	Medium
Complexity	Low
Est. days	1
Replaces	#14 (closed — written to gate D01, which is long done)

Features / method

Confirm access for whoever runs the canonical rerun, including the gated Hugging Face embedding dataset.

Description

The closed version existed to block cohort construction until every extractor was credentialed. That gate is spent — D01-D06 are done and the paired cohort exists. What matters now is narrower: the people who will actually run #62-#68 need working access, and one access path the original never mentioned has since become the primary one.

Reported results no longer come from raw MIMIC extraction. They come from pre-extracted embeddings on Hugging Face (MITCriticalData/mimic-iv-echo-jepa-embeddings), which is gated and requires PhysioNet credentials, plus the HuBERT-ECG Parquet. Anyone reproducing the pipeline needs the HF gate approved, not just the PhysioNet DUAs.

Steps:

1. Confirm PhysioNet credentialing and signed DUAs (MIMIC-IV-Echo 0.1, MIMIC-IV-ECG 1.0, MIMIC-IV 3.1) for the people working [D07] Paired cohort recovery — 4,103 shared subjects to 1,208 rows #62-[W07] Restate reported numbers from the canonical run #68 — not the whole team, just the active ones.
2. Confirm HF gate approval on MITCriticalData/mimic-iv-echo-jepa-embeddings for the same people.
3. Confirm read access to the HuBERT-ECG Parquet, and record where the canonical copy lives. configs/encoder/ecg_fm.yaml still points hubert_ecg_parquet at a /path/to/shared/ placeholder.
4. Record the real resolved paths so the config placeholders in configs/paths/default.yaml and configs/encoder/*.yaml stop being placeholders.

Acceptance criteria:

• Active contributors on the Task A milestone confirmed for PhysioNet + HF gate.
• Canonical location of the HuBERT-ECG Parquet documented.
• Config placeholder paths replaced with real ones, or documented as intentionally local-override.

[duckyquang]

Config side of this is done in PR #74; the people side needs the people, so here's the concrete state and a checklist.

Resolved in PR #74:

• configs/encoder/ecg_fm.yaml: hubert_ecg_parquet now points at the repo-local canonical path (data/interim/hubert_ecg_embeddings.parquet, what make reproduce-echo-hubert-local expects) instead of /path/to/shared/.
• The remaining /path/to/ entries (EchoJEPA + ECG-FM weights, data_root) are now explicitly documented as intentional local overrides, with the canonical locations named in the comments (echo weights on ORCD under weights/, per docs/embeddings.md §3).

Measured, not assumed: I checked HF access from this machine directly. MITCriticalData/mimic-iv-echo-jepa-embeddings is gated: manual, and content downloads 401 without an approved token — there is no HF token on this machine at all. So the clip-level echo manifest rebuild (which gates the #69 pooled-vs-clip comparison and the real cross-attention rerun in #72) is blocked on exactly this checklist, not on code. The vjepa2.1-vitl-mimic-pt-100 shard set is only 2.17 GB total, so once someone with gate approval pulls it, everything downstream runs on a laptop.

Checklist for the people actually working #62–#68 — @kevzho @paiksca @gullyboyslok, can you each confirm (or say which is missing):

1. PhysioNet credentialing + signed DUAs for MIMIC-IV-Echo 0.1, MIMIC-IV-ECG 1.0, MIMIC-IV 3.1.
2. HF gate approval on MITCriticalData/mimic-iv-echo-jepa-embeddings.
3. Where your copy of the HuBERT-ECG parquet lives — and specifically @kevzho, since the canonical joined manifest came from your machine: where is the source-of-truth copy of hubert_ecg_embeddings.parquet (ORCD? a team share?)? That's the one thing the docs still can't name precisely.

Once the three confirmations land and the parquet's canonical home is named, this issue can close.