novcor/gold-trace-agent-trace-assay-001

Gold Trace Agent Trace Assay 001

Foundational source-derived public assay

Version: 1.1.0 Release date: 2026-07-31 Publisher: Gold Trace Dataworks Contact: hello@gtdataworks.com

This is the first public source-derived assay set from Gold Trace Dataworks.

It contains five standalone technical reasoning artifacts derived from a private heterogeneous agent-trace archive. The five public questions and their reference answers are both released because this is an inspectable public set, not a secret benchmark. A separate first-ranked source candidate remains private as a true unseen holdout.

The private conversations, raw candidate archive, and holdout contents are excluded.

The four bars

Bar 01 — Public questions

Five prompt summaries for evaluation, rubric development, and method inspection:

• —data/public_questions.jsonl

Bar 02 — Public reference answers

The corresponding five transformed reference answers, with evidence posture and known gaps:

• —data/public_reference_answers.jsonl
• —data/public_assay_records.jsonl
• —data/PUBLIC_READING_EDITION.md

Because these answers are public, results against this split must not be called blind holdout performance.

Bar 03 — Assay decisions

The release preserves what was admitted and what stayed held:

• —data/selection_ledger.jsonl — five admitted transformed records;
• —data/held_candidate_ledger.jsonl — eighteen ordinary held candidates plus one designated private holdout;
• —docs/ASSAY_REPORT.md;
• —docs/HOLDOUT_AND_REFERENCE_ANSWER_POLICY.md.

Bar 04 — Evidence and validation

Schemas, hashes, manifests, validation code, and package receipts make the release inspectable and tamper-evident.

What makes this an assay

The automated ranking did not grant release. Public admission required manual prompt reconstruction, privacy review, internal-IP abstraction, editorial deduplication, explicit rights posture, known-gap documentation, and manifest sealing.

The five public records

1. 1.From architecture prototype to evidence-backed containment system
2. 2.Classify before deduplication in provenance-first artifact mining
3. 3.Forensic receipt layer for agent work
4. 4.High-grade signal mining for trace archives
5. 5.Bounded drift audit and evidence-supported repair

Intended uses

• —method inspection;
• —evaluation and rubric prototyping;
• —source-derived reasoning analysis;
• —dataset-governance training;
• —agent-system claim analysis;
• —curation workflow examples.

Not represented as

• —raw private conversations;
• —chronological multi-turn chat;
• —original prompt-response pairs;
• —SFT-ready training data;
• —a statistically representative benchmark;
• —blind holdout performance on the five public questions;
• —proof of model or runtime performance;
• —legal clearance for excluded private source material.

Validate

python tools/validate_release.py .
python examples/load_sample.py data/public_assay_records.jsonl

Licensing

Public transformed data and documentation are released under CC BY 4.0. Included Python code is released under the MIT License. The private source archive, omitted raw conversations, designated holdout, withheld candidate artifacts, third-party marks, and excluded materials are not licensed or redistributed by this package.