Corning/code-adapter-harness-evidence

Code adapter harness: selected research evidence

This bundle contains 420 SciBench task trajectories: seven named development configurations × two weight conditions × 30 questions. Six configurations passed their frozen workflow audits. legacy_workflow_diagnostic is an explicitly incomplete overall workflow, retained only for diagnostic inspection.

Start with the research handoff (中文), results table, schema, and GitHub implementation.

What the evidence supports

Bounded delivery preserves the code arm's 21/30 score while reducing its completion tokens from 170,510 to 101,620, relative to stable delivery. Requests fall from 232 to 177 and execution attempts from 157 to 106. Delivery changes from 30/30 to 29/30. This is useful engineering progress; it is not evidence of a consistent accuracy advantage from the trained code adapter.

Configuration	Sham correct /30	Code correct /30	Sham correct /29	Code correct /29	Code delivered /30
stable_delivery	24	21	24	20	30
strict_evidence	7	3	7	3	3
bounded_delivery	22	21	22	21	29
question_first	23	19	22	19	27
fresh_review	22	19	22	19	27
shared_drafts	22	21	21	21	26

The 29-task interpretation excludes only scibench_diff_032, whose missing numeric damping coefficient was declared before these runs. All 30 results remain available. Strict evidence formatting introduces substantial contract failures; its low score should not be interpreted as a pure measure of scientific reasoning capability. Shared drafts incur additional preparation cost, recorded separately in supplements/draft_preparation_cost.json.

These configurations reuse known development questions. They are not independent confirmation, and results must not be pooled by selecting the best answer from each configuration. No answer was recovered from an earlier draft for scoring. No preservation-of-capability or no-forgetting conclusion is established.

Contents

• trajectories/*.jsonl.gz: full normalized native logs, every captured model request/response, tool feedback, actual terminal results and runtime markers.
• trajectory_index.jsonl: task-level lookup with actual scores and record positions.
• task_catalog.jsonl: questions and reference labels, separately stored from generation inputs.
• results.json, results.csv: the 14 arm-level summaries, including the incomplete diagnostic.
• protocols/, plans/, provenance/: declared policies, supplied plans and shared-draft provenance.
• audits/, audit_support/: original workflow findings plus A/A, startup, source and adapter identity evidence.
• supplements/: documented failure analyses, training-source contract audit, preparation costs and broader benchmark summaries.
• reproduction/, model_provenance.json, environment/: portable setup inputs and deployment identifiers.
• verification.json, checksums.json: export validation and publication file integrity.

The broader DS-1000/SciCode supplement contains summary evidence; the complete trajectory panel in this bundle is SciBench. Model weights and credentials are not included. Upstream benchmark questions and reference labels remain subject to their original source terms.

Load one trajectory

import gzip, json
from huggingface_hub import hf_hub_download
path = hf_hub_download(
    repo_id="Corning/code-adapter-harness-evidence", repo_type="dataset",
    filename="trajectories/bounded_delivery.jsonl.gz",
)
with gzip.open(path, "rt", encoding="utf-8") as handle:
    records = (json.loads(line) for line in handle)
    row = next(r for r in records
               if r["task_id"] == "scibench_thermo_055" and r["arm"] == "code")
print(row["task_id"], row["result"])

Native logs and full request histories intentionally duplicate some text. Do not sum usage from every repeated copy; use result.budget or the aggregate result table.

This is a normalized export: local paths, credentials, calendar metadata, scheduler identifiers and wall-clock telemetry are removed or neutralized. Request order, tool-call identities, scientific programs and scored outcomes are retained subject to documented metadata normalization. The archive is not byte-identical to the private original logs; source_sha256 identifies the originals, while checksums.json verifies the published files.