cudabenchmarktest/opaque-agent-calibration-1000-baseline

Opaque Agent Calibration 1000 — Public Baseline

This is an operator-authorized public baseline containing 1,000 selected synthetic question/answer records generated through a blind worker API. It is exported in standard conversational fine-tuning JSONL.

Status and limitations

• Corpus class: calibration.
• Review state: awaiting_independent_review.
• Publication: explicit operator override of the stack's default non-release rule for calibration corpora.
• Production eligibility: no. This corpus is far below the stack's 24,000-record production minimum and was not passed through independent row-by-row review.
• Quality claim: pipeline and deterministic-answer baseline only. Publication does not establish convergence, improved intelligence, generalization, or safety.
• Prohibited use: high-stakes automated decisions or representation as human-authored production data.

The tasks are deterministic calibration exercises across ten categories. They are useful for testing ingestion, supervised fine-tuning plumbing, response formatting, and unchanged-base comparisons. They should not be treated as a substitute for a diverse production reasoning corpus.

Composition

Split	Records
Train	910
Validation	48
Test	42
Total	1,000

Each of the ten categories contributes exactly 100 selected records:

• deductive logic;
• dependency planning;
• financial arithmetic;
• inventory reconciliation;
• linear equations;
• program tracing;
• proportional allocation;
• robust statistics;
• set accounting;
• structured aggregation.

Format

Each data row contains only a two-message conversation:

{"messages":[{"role":"user","content":"..."},{"role":"assistant","content":"Derivation:\n1. ...\n\nFinal answer:\n..."}]}

Governance metadata is kept in parallel files under metadata/; it is not injected into the trainer-visible conversation. evidence/ contains the immutable generation manifest and rejection ledger. SHA256SUMS covers the uploaded artifact set.

Blind-worker observation

The answering subagent received only the loopback worker endpoint, a temporary worker token, and the lease/submit schema. Lease responses exposed task_id, lease_token, expires_at_unix, and messages; successful receipts exposed only status: recorded. No correctness, reward, provenance, source, selection, or review metadata was returned to it.

This was context-blind by API and instruction, not an adversarial OS-isolation proof: the agent runtime shared the host but was prohibited from reading project files, databases, process state, environment variables, or the trusted control plane.

Run outcome

• Submitted candidates: 3,000.
• Evaluated candidates: 3,000.
• Validation passes: 3,000.
• Independent answer-verifier passes: 3,000.
• Contamination passes: 3,000.
• Selected records and unique lineages: 1,000.
• Forge stage: awaiting_review.

The first immutable attempt is retained locally as failed evidence. Its answers were nearly all correct, but several derivation templates contained individual steps shorter than the response contract's 24-character minimum. A new run corrected only that declared format requirement and then passed finalization.

Loading

from datasets import load_dataset

dataset = load_dataset("cudabenchmarktest/opaque-agent-calibration-1000-baseline")

For training, use only messages from the selected train split. Keep the test split sealed until training and hyperparameter choices are complete.

Evaluation requirement

Follow EVALUATION-PROTOCOL.md: compare otherwise identical tuned runs against the unchanged base using seeds 17, 29, and 47, report variance and regressions, and do not convert a release/gate pass into an unsupported intelligence claim.