feat(consistency): add cross-run trend tracking template and methodology

[nanookclaw]

Summary

The consistency.md dimension asks the right question — does the system behave similarly across repeated runs? — but the current experiment-log.md template captures only a single snapshot. There is no mechanism to track consistency trends across multiple experiment log entries over time.

The Gap

experiment-log.md records one run:

- Date:
- Consistency result:

To answer whether the system is getting more or less consistent over time, you need to compare across runs. Right now that comparison is manual and informal — teams typically don't do it systematically.

Proposal

Two additions:

1. templates/baseline-comparison.md

A structured template for comparing current experiment logs against a known-good baseline:

# Baseline Comparison Log

## Baseline (v___)
- Date established:
- System version:
- Consistency result (baseline):
- Correctness result (baseline):

## Comparison Runs

| Date | System version | Consistency Δ | Correctness Δ | Notes |
|------|---------------|--------------|--------------|-------|
| | | | | |

## Trend Assessment
- Direction (improving / stable / degrading):
- Sessions since last significant change:
- Action threshold: flag when any dimension Δ > 0.10 vs baseline

2. Extension to docs/consistency.md

Add a section on longitudinal consistency — the difference between:

• Within-run stability (current scope): same agent, same scenario, repeated N times in one session
• Cross-run consistency (proposed): same agent, same scenario, measured across sessions over days/weeks

A system that is consistent within a single evaluation run can still drift significantly across runs separated by days or model updates. The cross-run view is what matters for production reliability.

Why This Matters

Practitioners building on top of this framework will run evaluations repeatedly over time (before/after model updates, after prompt changes, after infrastructure changes). Without a structured cross-run comparison mechanism, they will either:

1. Manually eyeball past logs (error-prone), or
2. Only look at the most recent run in isolation (missing drift entirely)

The baseline comparison template makes cross-run consistency a first-class evaluation artifact.

Related Work

The PDR (Periodic Delivery Reliability) framework (DOI: 10.5281/zenodo.19339987) tracks this as delivery_score and calibration_delta across sessions — the same cross-session consistency measurement problem applied to autonomous agent workflows. The delta-from-baseline approach is validated in that context.

Happy to draft both the template and the consistency.md extension if this direction looks right.