the-concord/concord-index-traces

Concord Index: Model Responses and Company Commitments

How do models' answers to policy trade-offs compare with companies' published commitments?

Concord Index links repeated answers from twelve models to cited interpretations of company commitments. Follow public excerpts through editorial annotations, model responses and comparison arithmetic, including missing evidence, failures and corrections. Models answered without company context.

The traces are reconstructed prompt-and-response sessions from saved benchmark records. The release supports inspection of response sensitivity and the assumptions behind company/model comparisons. The comparison scale has not been independently validated; predictive value for organizational decisions remains unestablished.

Release v0.1.2, prepared 2026-10-16. Measurements: 2026-09-21. Company research: 2026-09-22 to 2026-09-30, with two later corrections. Authenticated full-download, loading and notebook checks passed on 2026-10-16. Verification record · Verification history.

Component	Coverage	Records
Company policy	48 designs × 2 wordings × 3 repeats × 12 models	3,456 responses
Hard Questions	10 designs × 1 wording × 3 repeats × 12 models	360 responses
Company cohort	31 scored, 14 insufficient public record	45 reports
Company dimensions	261 supported, 277 unknown, 2 contested	540 positions
Evidence	459 distinct excerpt/source-hash pairs	726 citation attachments
Sources	1,000 distinct URLs	1,004 source references

The 3,816 records come from 58 question designs. Repeats, paraphrases and reused citations are not independent observations. The current cohort contains 45 selected reports; a separate ledger documents the full sixty-report inventory.

Notebook · Methods · Field guide · Research assessment · Rights · Index website

Download, verify and load

Install the tested optional packages from requirements-optional.txt. The verifier and core analysis use only Python's standard library.

import subprocess
import sys
from huggingface_hub import HfApi, snapshot_download
from datasets import load_dataset

REPO = "the-concord/concord-index-traces"
REVISION = HfApi(token=False).dataset_info(REPO).sha  # Resolve once, then pin this full commit.
print("Hub revision:", REVISION)  # Retain this with your results and citation.
root = snapshot_download(REPO, repo_type="dataset", revision=REVISION, token=False)
for script in ("verify.py", "analysis.py", "check-loading.py"):
    subprocess.run([sys.executable, f"{root}/{script}", root], check=True)

responses = load_dataset(root, "company-policy", split="test")
dimensions = load_dataset(root, "company-model-dimensions", split="test")
print(len(responses), len(dimensions))  # 3456, 6480

Open reproduce.ipynb from the downloaded directory for a worked walkthrough. The included scripts reproduce analysis of saved records. The original collection harness is private, and a complete executable collection environment is not included.

Research task	Hub configuration or file
Answers and repetition	company-policy, hard-questions
Public commitments and evidence	companies, company-positions, citations, sources
Exact questions, scale and model provenance	questions, scales, models, question-diagnostics
Positions and comparisons	model-profiles, company-comparisons, company-model-dimensions
Baselines and sensitivity	baselines, dimension-removal-sensitivity
Reconstructed sessions	traces.zip, trace-sample-*.jsonl

One path from evidence to score

Salesforce report 00a75e76-a9d6-4500-ba2b-19d98f96fc0b contains this excerpt:

Will it build trust with all of our stakeholders - employees, customers, company, and community?

Source: Salesforce Code of Conduct. Citation f518b833f27d7afdd386b3a0 joins to company dimension purpose. Concord maps it to anchor purpose.stakeholder_decisions, value 75, with the rationale that the decision framework directs employees to consider trust with several stakeholder groups. That numeric mapping is editorial.

For the same dimension, purpose-1 asks whether major decisions should consider documented effects on employees, customers, communities and the environment rather than financial return alone. GPT-6 Astra answers yes in all six calls. Across the three included questions (purpose-1, purpose-2, purpose-3), its keyed fractions are each 1.0, so its model position is 100. purpose-4 is excluded by the frozen common-question mask.

The retained dimension agreement is 100 - abs(75 - 100) = 75. It is a distance on the editorial scale; it does not mean a 75% chance of compliance. The overall report averages its eight comparable dimensions and retains six models in its close group. The notebook executes these joins with analysis.evidence_to_score_example(root) and retains all response IDs.

What these measurements show

All valid answers agree on 37 of 48 company-policy designs. Four dimensions have identical aggregate positions across the panel. All 31 scored companies have close groups of four to eleven models, with no uniquely separated model under the retained grouping rule.

Each point is one model's aggregate position under the frozen common-question mask. Equal positions do not establish identical behavior.

The best observed model exceeds the in-sample panel-mean baseline by a median 3.0 points across scored companies (range 1.0 to 7.7). This is a descriptive comparison with the observed panel, not a held-out baseline or significance test. A dimension-removal diagnostic has 157 cases across twenty companies retaining at least six dimensions; sixteen companies have a removal that displaces every original point-score leader. The other 66 cases fall below that coverage floor and are labeled. These diagnostics do not change frozen scores or close groups.

Unknown and contested dimensions do not enter scores. Different company coverage limits comparisons across companies.

Provenance and limitations

Questions were mostly AI-generated. Team Concord reports using Claude Opus 5.5, GPT Astra and Grok 4.7 for development, with AI-assisted review and human approval by Team Concord. Exact development endpoints, prompts, revision histories and item-level review records were not retained. No independent external question or annotation review was performed. Evaluated model IDs and known serving settings are retained separately in models and the frozen configurations.

Company positions interpret published commitments. The company anchors and fractions of model answers are not a validated common interval scale. For example, representation goals can coexist with uniform hiring standards: a numeric gap on “equality” need not demonstrate a policy disagreement. Research notes work through that example. Annotation confidence is an uncalibrated pipeline value, not human agreement or a probability; Fields explains defaults.

The benchmark uses English forced choices, a curated panel and three repeats. It has no held-out generalization test, psychometric validation or demonstrated prediction of real company behavior. Two retained wordings have documented construction issues. The test split denotes evaluation records, not unseen problems. These records do not establish internal practice, model beliefs, operational quality, safety, certification or employee fit.

The 3,804 reconstructed sessions contain saved instructions, questions and model outputs, including two ambiguous returns. Structured outputs were often parsed and reserialized by the harness. Twelve unavailable calls remain separate harness failures. Full provider logs, hidden reasoning and complete research-agent conversations are absent. Limited attributed company excerpts and source metadata are included; full scraped pages and contact submissions are excluded. Citation text-consistency checks do not independently validate the interpretations.

Rights, corrections and citation

Mixed rights apply. Concord-created questions, labels, scores, metadata, schemas and documentation use CC BY 4.0 under LICENSE.md. The three utility scripts use MIT. Model outputs remain provider-dependent, and company excerpts retain their original publishers' rights. This evaluation and audit release grants no unrestricted model-training, fine-tuning, distillation or competing-model development rights. Rights by component and provider conditions define those boundaries.

Frozen records retain their values. Corrections link withdrawn and replacement reports; historical objects are marked separately. To report an issue, open a Discussion with the Hub revision, report/record IDs and supporting evidence, or use Contact.

Maintainer: Team Concord. Cite package v0.1.2, the resolved Hub commit and its manifest. The resolved commit identifies the complete package; dataRevision in the manifest identifies an earlier verified package with identical scientific records.

@misc{concordindex2026,
  author = {{Team Concord}},
  title = {Concord Index: Model Responses and Company Commitments},
  year = {2026},
  version = {0.1.2},
  url = {https://huggingface.co/datasets/the-concord/concord-index-traces},
  note = {Package v0.1.2. Model measurements collected 2026-09-21. Record the resolved Hub commit.}
}