Action Outcome Vendor SQL Replay 010 - Raw Research Artifact

Generated at: 2026-06-18T15:03:38Z Endpoint: http://127.0.0.1:19343/ SQL seed: docs/research/results/action_outcome_sql_replay_006.sql

Load Summary

Seed statements attempted: 203
Seed statements succeeded: 203
Seed statements failed: 0
Seed row-count status: pass
Vendor table row-count status: pass
Failed-repeat JOIN status: pass
Context top-action JOIN status: pass
Suppression JOIN status: pass
Misleading retrieval JOIN status: pass
Window ROW_NUMBER status: pass
Window LAG status: pass
Overall SQL/JOIN/window status: pass

Seed Row Counts

Table	Expected Inserts	Live Rows	Status
`action_outcome_episode_metrics`	101	101	pass
`action_outcome_episodes`	32	32	pass
`action_outcome_gate_suppressions`	21	21	pass
`action_outcome_replay_recommendations`	18	18	pass
`action_outcome_retrieval_quality_labels`	26	26	pass

Vendor Replay Table Counts

Table	Expected Rows	Live Rows	Status
`action_outcome_vendor_queries_010`	6	6	pass
`action_outcome_vendor_recommendations_010`	72	72	pass
`action_outcome_vendor_retrieval_010`	72	72	pass
`action_outcome_vendor_suppressions_010`	21	21	pass

Failed-Repeat JOIN

This query joins top recommendations to query episodes and uses alias-qualified WHERE predicates on the joined result:

SELECT r.variant, r.query_id, r.action_class, q.true_failed_action
FROM action_outcome_vendor_recommendations_010 r
JOIN action_outcome_vendor_queries_010 q ON r.query_id = q.query_id
WHERE r.recommendation_rank = 1 AND r.avoids_failed = 0

Status: pass

Variant	Query	Recommended Action	Historical Failed Action
`reflection_only_memory`	`aoe_cache_002`	`restart`	`restart`
`runbook_action_prior`	`aoe_cache_002`	`restart`	`restart`
`runbook_action_prior`	`aoe_queue_003`	`restart`	`restart`
`similar_incident`	`aoe_cache_002`	`restart`	`restart`
`similar_incident`	`aoe_payment_002`	`scale_out`	`scale_out`

Context-Gated Top Actions

Status: pass

Query	Context-Gated Top Action	Historical Failed Action
`aoe_cache_002`	`cache_purge`	`restart`
`aoe_checkout_002`	`rollback`	`restart`
`aoe_inventory_002`	`config_revert`	`scale_out`
`aoe_payment_002`	`traffic_drain`	`scale_out`
`aoe_queue_003`	`scale_out`	`restart`
`aoe_search_003`	`rollback`	`scale_out`

JOIN/Window Acceptance

Suppression JOIN rows: 21
Suppression JOIN status: pass
Misleading retrieval JOIN rows: 23
Misleading retrieval JOIN status: pass
ROW_NUMBER rows: 72
ROW_NUMBER status: pass
LAG rows: 72
LAG status: pass

Interpretation

Experiment 010 now has a live ZeptoDB SQL acceptance path. The comparison is no longer only a Python fixture report: recommendations, retrieval evidence, query controls, and context suppressions are materialized into declared ZeptoDB tables and audited with native hash JOIN plus window queries.

The key product result remains intact: vendor-inspired baselines have failed-repeat top actions, while context_gated_action_outcome returns six safe top actions and uses suppressions to keep mismatched historical outcomes from dominating the ranking.

Next Best Step

Port this SQL/JOIN/window replay into the two-node live harness to record which parts are single-node SQL-complete and which parts still depend on distributed JOIN/window planner work.