Ralph Loop Note: Confidence Hydration Refresh

Context

research/latest_state.json is Ralph’s first handoff surface, but legacy files can carry an older confidence schema that lacks the newer replay-quality fields:

• pair_eventful_metadata_share
• large_move_pair_eventful_metadata_share
• recent_replay_rows
• recent_replay_share
• recent_pair_eventful_metadata_rows
• recent_pair_eventful_metadata_share
• recent_large_move_pair_eventful_metadata_rows
• recent_large_move_pair_eventful_metadata_share

That forced operators to infer replay-quality context from status output instead of the persisted handoff file.

Change

Updated research/status.py so status hydration also refreshes the persisted confidence block when the stored handoff is missing those newer replay-coverage fields. The refreshed block is built from the current replay snapshot plus the current research-loop confidence policy.

Also documented the behavior in research/README.md.

Validation

Ran:

python3 research/status.py --json >/tmp/pam_status.json
python3 -m py_compile research/status.py research/continuous_loop.py research/candidate_review.py research/prepare_replay.py

Observed in the hydrated status payload:

• confidence.pair_eventful_metadata_share = 0.0363
• confidence.recent_replay_rows = 0
• confidence.recent_replay_share = 0.0
• confidence.recent_pair_eventful_metadata_share = 0.0
• confidence.reasons now includes pair_eventful_metadata_share<0.05

Observed after persisting the refreshed handoff:

• research/latest_state.json now contains the newer confidence replay-coverage fields instead of only the older minimal block.

Operator Impact

Ralph can now read low replay-quality causes directly from latest_state.json without waiting for a fresh loop cycle or reconstructing the confidence gate from status output.