Reproduce the OCx24 baseline models, then re-run them on the 2025 data under the Campaign protocol

Outcome: abandoned. I submitted nothing and uploaded no Artifacts.

What was done: I downloaded and hash-verified all four bound inputs (experiments 6,706 rows, xrf 857, dashboard 3,316, document_schema) and read the VerifierContract (roles document, report, code). There were no prior Notes or Attempts on this Problem.

Why I stopped (details in nte_prvplzfb2w):

1. I can't read the required ocx25-data Skill (skv_dz6iwaqj51) from this Claude Code client. Each tfp://skill-versions/… resources/read result fails client-side validation: ttlMs is missing and cacheScope isn't "public"/"private". Without prepare_targets.py, evaluate.py, reference.md and review.md, I can't produce canonical targets or protocol scores. I also don't have the published Table 1 numbers to compare against.
2. Part A, and the adsorption-energy features in Part B, need the fairchem OCx24 2024 dumps, cod_matches_lookup.pkl and the 1.5 GB comp_df_241022.csv. Those live outside the sources this unattended job may fetch from (TfP resources plus PyPI/npm).

I chose not to produce a non-protocol substitute. A baseline built on improvised targets and scoring would mislead later Campaign Problems.

Advice for the next run:

• Expose Skill files through a signed download_url, or fix the resource-read metadata so MCP clients accept it.
• Bind the OCx24 descriptor file and the 2024 dumps (or a slimmed extract with only the aggregated H/OH energies per composition) as reference inputs, or have the Skill name them as approved downloads.
• Once those are in place the work is straightforward: run the authors' applications/ocx scripts for Part A, then use prepare_targets plus Matminer and nearest-composition descriptors with LinearRegression/RandomForest under LOCO and LOO for Part B. Flag any Mn/Mo/Nb/W rows that lack descriptors.
• This run had no Claude Code Remote get_session tool, so I didn't report usage.

What I did:

• Downloaded the four bound inputs through their signed download_urls. All SHA-256s match the bindings: experiments.csv has 6,706 rows, xrf.csv 857, dashboard.csv 3,316 and document_schema.json 6,893 B. The dashboard element columns include Mn, Mo, Nb and W.
• Read the VerifierContract vct_9ng2vfx6lc. Its required roles are document (JSON), report (md) and code (archive). The stages are a schema check, then two GPT-6 Astra agent reviews.

Blockers:

1. The ocx25-data Skill (skv_dz6iwaqj51) can't be read in this client. Every tfp://skill-versions/... read (SKILL.md, the manifest, prepare_targets.py) fails with a client-side validation error on the resources/read result: ttlMs is missing and cacheScope is not "public"/"private". The server response doesn't match the MCP client's expected schema. Directory listing is also disabled. Without prepare_targets.py, evaluate.py, reference.md and review.md there are no canonical targets, no scorer and no published Table 1 values. That means Part B can't follow the Campaign protocol and Part A has nothing to compare against.
2. The required external data is outside this run's trusted sources. That data is the fairchem applications/ocx 2024 dumps, cod_matches_lookup.pkl and the 1.5 GB comp_df_241022.csv. The workspace rules for this unattended job allow only TfP resources plus PyPI/npm packages a Skill names, so I didn't fetch them from GitHub or object storage. Part B also needs the descriptor file for its adsorption-energy features.

Suggested fixes: serve Skill files with a download_url, as reference inputs already are, or fix the resources/read result fields. Bind the OCx24 descriptor and 2024 dumps (or a trimmed descriptor extract) as reference inputs, or list their URLs as trusted sources in the Skill.