Correct the completed workflow-causation investigation

[nothingnesses]

This issue records a proposal. Implementation requires a selected work brief and separate execution approval.

Background and problem

D-096 records a completed but uncorrected causation investigation. Its independent critique disputes:

• Grades.
• Counts.
• Causal ranking.

The critique identifies ancestry selection and model confounding.

Neither the original ranking nor a proposed replacement establishes causation.

Proposal

Examine the critique against preserved evidence. Incorporate supported corrections into one sanitised final report.

Distinguish:

• Observations.
• Plausible explanations.
• Unresolved causal claims.
  Keep reset-era controls separate from later controller and planning-loop decisions.

Scope and exclusions

This is the unresolved causation follow-up, not another preservation delivery for the reports merged through PR #9.

This proposal excludes:

• New experiments.
• Raw session copies.
• Private evidence in Git.
• Changes to unrelated audit records.

The work does not select itself in .agents/work.toml.

Acceptance criteria

• Each material correction identifies its supporting evidence.
• Unsupported corrections remain unresolved.
• Corrected counts and grades have a reproducible basis.
• The report states ancestry and model confounders.
• Later choices do not appear as controls available during the historical period.
• The final report preserves uncertainty and contains no private execution data.

Alternatives and open questions

Retain disputed claims with explicit qualifications when evidence cannot resolve them. Prefer supported correction over a replacement causal ranking under Ground decisions in evidence.

References

• Disputed evidence and D-096.
• D-096.