EvidencePublic artifacts/judge path
GitHub repository ↗Inspect without running the project
Start with the run. Then inspect every claim.
These are the exact public-safe outputs behind Retirement Conductor: native actions, reviewable changes, DataHub evidence, refusals, and the experiments that changed what we shipped.
- Clean path
- Real PostgreSQL drop
- Changed world
- Zero destructive actions
- Evidence policy
- No hidden caveats
The shortest judge path
The consequential run
Clean retirement
One PostgreSQL DROP COLUMN committed exactly once. The replacement Spark workload kept the same 128-row output digest, and replay was refused.
COMMITTED ONCE↗Late-consumer refusal
Fresh DataHub evidence discovered a late Spark consumer. The lease became INVALIDATED, zero destructive statements committed, and the legacy column remained present.
REFUSED BEFORE ACTION↗Static sign-off control
The point-in-time approval stayed green, the column was dropped, and the legacy Spark workload failed with SQLSTATE 42703.
NATIVE BREAKAGE↗What the agent actually changed
Exact change and validation
Exact dbt migration
The reviewable one-file patch that moves the authorized consumer from legacy_status to order_status.
1 FILE↗Native change receipt
The accepted target, source version, plan binding, and dbt parse, seed, build, and test result.
PASSED↗Agent transcript
A public-safe stage-by-stage record of the Codex run, including the external authorization pause and the later reversal.
5 STAGES↗Public PR and passing CI
The exact generated dbt change as a reviewable GitHub pull request with the semantic-dbt check attached to its head commit.
CI PASSED↗Where context changed the answer
DataHub evidence and contributions
DataHub inventory expansion
Repository inspection found one consumer. Fully paged DataHub lineage returned 31 across seven pages—a minimum expansion of 30 consumers.
31 / 31↗Unified agent campaign
The digest-bound run joining Codex, DataHub discovery, one exact PR, native validation, publication read-back, lease issuance, and late-consumer invalidation.
VERIFIED↗Lineage pagination fix
Our upstream DataHub MCP contribution fixes a GraphQL pagination boundary that could silently hide later consumers.
PR #195↗Deployment-gate diagnostics fix
A second upstream contribution corrects diagnostics that could misstate the DataHub MCP deployment gate.
PR #196↗Successes, failures, and negative results
Stress tests and product truth
Retirement Gauntlet v2
Twenty-four frozen cases: 4 READY, 6 BLOCKED, 10 UNSAFE, and 4 REVIEW. All matched, with zero false readiness and zero unexpected closure.
24 / 24↗Native breakage outcome lab
Two clean Spark/JDBC sequences prove the consequence of deleting too early: legacy_status disappears and the workload exits 42 with SQLSTATE 42703.
BREAKAGE PROVED↗Nested-model ablation
Across 135 attempts, DataHub context improved fault coverage but nested Gemini failed the value threshold. We removed the extra model layer.
REMOVE GEMINI↗Agent boundary ablation
The project skill and MCP improved completion and reduced calls and retries. Shell exposure remained, so capability containment is explicitly unclaimed.
KEEP · BOUNDEDNESS UNKNOWN↗Fresh-CI comparison
The honest counterfactual: competent fresh CI matched the frozen safety properties. Retirement Conductor's narrower value is the joined workflow and durable authority protocol.
NO MATERIAL ADVANTAGE↗