Action Outcome SQL Replay 006 - Raw Research Artifact

Generated at: 2026-06-18T02:14:23Z Fixtures:

• docs/research/fixtures/action_outcome_episodes.json
• docs/research/fixtures/action_outcome_distractor_episodes.json
• Quality labels: docs/research/fixtures/action_outcome_retrieval_quality_labels.json

SQL Table Counts

Table	Rows
action_outcome_episodes	32
action_outcome_episode_metrics	101
action_outcome_retrieval_quality_labels	26
action_outcome_replay_recommendations	18
action_outcome_gate_suppressions	21

Replay Summary

Variant	Source	Top-3 Hit Rate	Failed-Action Avoidance	Cross-Family Top3	Weak Cross-Family Top3	Gate Suppressions	Labeled Top3 Quality
full_guarded	SQL	1.00	0.67	0	0	0	useful:8, superficial:5, misleading:5
context_gated	SQL	1.00	1.00	0	0	21	useful:8, superficial:4, misleading:6
context_gated	JSON control	1.00	1.00	0	0	21	useful:8, superficial:4, misleading:6

SQL vs JSON Control

• Match status: pass
• Top-action mismatches: 0

Per-Query SQL Context-Gated Actions

Query	SQL Top Action	JSON Top Action	Match	SQL Top Actions	Avoids Failed Repeat
aoe_checkout_002	rollback	rollback	yes	rollback:1.03, scale_out:0.31, config_revert:0.16	yes
aoe_payment_002	traffic_drain	traffic_drain	yes	traffic_drain:0.67, restart:0.16, scale_out:0.16	yes
aoe_inventory_002	config_revert	config_revert	yes	config_revert:0.36, traffic_drain:0.36, scale_out:0.16	yes
aoe_cache_002	cache_purge	cache_purge	yes	cache_purge:0.59, restart:0.16, rollback:0.16	yes
aoe_queue_003	scale_out	scale_out	yes	scale_out:0.54, traffic_drain:0.17, restart:0.16	yes
aoe_search_003	rollback	rollback	yes	rollback:0.56, restart:0.36, traffic_drain:0.07	yes

Example SQL Queries

SELECT episode_id, incident_type, action_class, human_outcome
FROM action_outcome_episodes
WHERE incident_type = 'order_queue_backlog'
ORDER BY action_ts_ns;

SELECT episode_id, metric_name, metric_value
FROM action_outcome_episode_metrics
WHERE metric_name IN ('cpu_pct', 'db_conn_used_pct', 'consumer_error_pct')
ORDER BY episode_id, metric_name;

SELECT query_id, candidate_id, action_class, reasons
FROM action_outcome_gate_suppressions
WHERE multiplier_micros < 1000000
ORDER BY query_id, candidate_id;

Interpretation

Experiment 006 validates the first SQL-backed research contract. The replay logic no longer consumes raw JSON fixtures directly; it reconstructs episodes from SQL tables and matches the JSON control result.

This is still a local SQL harness, not a live ZeptoDB server run. The next step is to execute the generated SQL through ZeptoDB’s HTTP SQL endpoint and compare live query results against this report.

Next Steps

1. Run the generated SQL seed against a live ZeptoDB HTTP server.
2. Add live SQL query checks for top actions and suppression rows.
3. Promote the table schema into a design note if the live replay matches.