Five records, five observation points
A complete matrix separates period, unit, collection method, access model, direct observation, inference, and known gap across legal process, OpenAI's company report, Hugging Face host telemetry, the independent investigation, and vendor remediation.
The five evidence lanes answer different questions and carry different access and attribution boundaries.
| Row | Column | Record | State |
|---|---|---|---|
| Access model | Hugging Face host telemetry | Direct access to Hugging Face systems and records within the host review | Verified |
| Access model | Independent investigation | Independent investigation with host-controlled data access | Caution |
| Access model | Legal process | State inquiry with compulsory requests for records | Caution |
| Access model | OpenAI company report | First-party access to OpenAI systems and records | Caution |
| Access model | Vendor remediation | Vendor-authored remediation and advisory records | Caution |
| Collection method | Hugging Face host telemetry | Affected-host logs and forensic reconstruction | Verified |
| Collection method | Independent investigation | Analysis of datasets OpenAI supplied, with requests for additional material | Caution |
| Collection method | Legal process | Official announcement, subpoena, and multistate letter | Caution |
| Collection method | OpenAI company report | First-party internal investigation and evaluation | Caution |
| Collection method | Vendor remediation | Vendor publication and maintained advisory index | Caution |
| Known gap | Hugging Face host telemetry | Recovered records may be incomplete; sensitive indicators are generalized or redacted | Caution |
| Known gap | Independent investigation | No direct OpenAI infrastructure query or primary-model run access; known loss and analysis errors | Caution |
| Known gap | Legal process | No adjudicated violation, liability, or consumer harm in the cited materials | Caution |
| Known gap | OpenAI company report | Preliminary evaluations lack public sample sizes, intervals, and independent replication | Caution |
| Known gap | Vendor remediation | The record does not establish deployed patch coverage or remediation effectiveness | Caution |
| Inference | Hugging Face host telemetry | Parts of the motivational account, including an effort to obtain test solutions | Caution |
| Inference | Independent investigation | Behavioral analysis, including origin in broader scorer-tampering research | Caution |
| Inference | Legal process | Consumer-protection implications remain allegations under investigation | Caution |
| Inference | OpenAI company report | Attribution and remediation effectiveness remain first-party interpretations | Caution |
| Inference | Vendor remediation | No independent incident attribution is inferred from the advisory lane | Unavailable or historical |
| Direct observation | Hugging Face host telemetry | Recovered host activity, affected assets, and operational search metadata | Verified |
| Direct observation | Independent investigation | Patterns in supplied transcripts, board data, and artifacts | Caution |
| Direct observation | Legal process | Existence, scope, requests, and deadlines of the legal process | Verified |
| Direct observation | OpenAI company report | Internal activity, system configuration, impact assessment, and remediation tests | Caution |
| Direct observation | Vendor remediation | Published collaboration statement and advisory entries | Caution |
| Period | Hugging Face host telemetry | Observed July 9 at 02:28 UTC through July 13 at 14:14 UTC | Verified |
| Period | Independent investigation | Scope June 26 through July 13, focused mostly on July 7 through July 13 | Caution |
| Period | Legal process | Subpoena served August 20; investigation announced August 24; response commanded by September 14 | Caution |
| Period | OpenAI company report | Internal chronology through the August 26 reports | Caution |
| Period | Vendor remediation | Statement published July 27 and updated August 5; advisories current through August 25 in the ledger | Caution |
| Unit | Hugging Face host telemetry | Recovered actions, clusters, systems, and affected customer datasets | Verified |
| Unit | Independent investigation | Agents, entries, board messages and files, standalone messages, and transcripts | Caution |
| Unit | Legal process | Legal questions and 16 requests for information or documents | Caution |
| Unit | OpenAI company report | Models, workers, systems, credentials, repositories, tasks, controls, and evaluations | Caution |
| Unit | Vendor remediation | Vendor statement, advisories, CVE context, and fixed-version context | Caution |
- Sources
- Retrieved
- Units
- qualitative scope fields
- Transformations
- Keep named units inside their source cells and do not aggregate across columns.
- Preserve each source lane's terminology and attribution without assigning a score.
- Render every missing or bounded detail as an explicit gap rather than an inferred absence.
- Uncertainty
- Cells summarize the reviewed source ledger. Source access and publication dates differ, and MR-2 mirrors the joint report rather than adding another investigation.
- What this figure does not prove
- The matrix does not rank credibility, merge incompatible units, resolve disputed interpretations, or convert allegations into findings.