contrib: MolecularFullerene - schema-bridge-components

[MolecularFullerene]

Summary

This contribution adds a preregistered component-level follow-up to the already merged molecularfullerene-dsh-schema-ablation study (PR #7).

It runs the real DSH agent loop with the registered official persistent bash executor held fixed, while crossing only the model-visible bash.description × bash.parameters bundles. The Minimal system, str_replace_editor, tool count/order, reasoning effort, execution barriers, and scorer are fixed.

Main result

The main sample is 10 frozen read-only prompts × 4 fixed runtime identities × 4 arms = 160 first-assistant trajectories, after an independent four-cell protocol pilot:

• PP: 26/40 minimal-like (65.0%)
• PO: 21/40 (52.5%)
• OP: 14/40 (35.0%)
• OO: 9/40 (22.5%)

Preregistered task-level Latin randomization tests with Holm correction across D/Q/I:

• description D (one-shot − persistent): −0.300; two-sided p = 0.000110; Holm-adjusted p = 0.000330
• parameters Q: −0.125; p = 0.159868; Holm-adjusted p = 0.319737
• interaction I: 0; p = 1

D is the only primary contrast that passes the preregistered multiplicity correction. The Q bootstrap interval is reported as descriptive and is not treated as confirmatory evidence.

All 164 protocol observations were accepted, with zero retries, zero executor dispatches, and zero failed-unit replacements. Tool-name and argument validity are protocol diagnostics only.

Relationship to existing contributions

This does not modify the prior MolecularFullerene contribution. It creates a separate follow-up folder because the repository's incremental CI permits new contribution directories but does not permit modifying an existing one.

It also complements xiaobright-v4-tool-surface-dose-response: that study varies catalog size/composition across Pro and Flash, while this study keeps exactly two tools and a fixed executor, then isolates description and parameter bundles under a preregistered Pro-only factorial design. It does not reproduce or make claims about that contribution's Project2 results.

Included material

• contribution README with environment, provenance, interpretation, and limitations
• one concise results report
• the strict scorer's complete publication-allowlisted JSON, copied byte-for-byte from the canonical research repository

The public-summary file SHA-256 is 77e3f411a584c5ef32e2939e740693388d4c5f9fe33090406daa27f8c9644f88.

Source provenance:

• preregistration
• frozen implementation
• canonical result publication

Validation and publication boundary

• python3 scripts/check_pr.py --base origin/main --head HEAD: 16/16 checks passed
• git diff --check origin/main...HEAD: passed
• both repository indexes are append-only (zero deleted lines)
• no API key, Authorization value, raw reasoning/text, raw tool arguments, request body, real identity/session/request/call identifier or hash, absolute local path, or per-request timestamp is included
• the deidentified matrix ordinals are reconstructible from the public seed and are explicitly not presented as cryptographic anonymization

Important limitations

The endpoint is a lexical trajectory label, not task ability, answer quality, Project2 performance, an internal persona/route, or evidence of RL overfitting. Only 10 prompt clusters are represented; 160 trajectories are not 160 independent tasks. No model-proposed tool call was executed.

deepseek-v4-pro is a mutable service alias. Account, endpoint cohort, server shard, and time were not randomized. The raw-artifact digest is not a provider signature, and the reported Harness commit does not attest ignored generated lib/ or node_modules bytes.