pg-evidence numbers about PostgreSQL extensions, each one rerunnable

Every number, with the run behind it

Each figure below is a published claim about the pg_agent_gate / pg_phasevec family of PostgreSQL extensions. Each one links to the run that produced it: the versions, the hardware, the commit, the raw data and its sha256, and the one command that reruns it. If a rerun disagrees, submit it: a number claims, the machine decides.

Claims

Same model, same tasks, a superuser checking what really changed: through a popular Postgres MCP server, 4 of 7 trap tasks broke out. With the database deciding, 0 of 7.

Supported by the raw data

How it is computed
  • stock_mcp.breach = 4/7 → 4/7
  • gated.max_breach = 0/7 → 0/7

Recompute it yourself: make claim-trap-breakouts

With the same gate behind both, the agent without MCP used 5x fewer prompt tokens than through MCP (24k vs 126k for 13 tasks) and finished each task 13x faster: a median of 14 s against 185 s, on a CPU-only CI runner anyone can re-run.

Supported by the raw data

How it is computed
  • ratio.prompt_tokens.gated_mcp_over_native_loop (round) = 5.246 → 5
  • native_loop.prompt_tokens_total (round_thousands) = 24099 → 24
  • gated_mcp.prompt_tokens_total (round_thousands) = 126415 → 126
  • tasks.per_arm = 13 → 13
  • ratio.median_seconds.gated_mcp_over_native_loop (round) = 12.932 → 13
  • native_loop.median_seconds (round) = 14.34 → 14
  • gated_mcp.median_seconds (round) = 185.439 → 185

Recompute it yourself: make claim-mcp-tokens-and-time

Writing one statement and letting the database decide took a median of 6 s per task.

Supported by the raw data

How it is computed
  • native.median_seconds (round) = 5.82 → 6

Recompute it yourself: make claim-one-statement-median

The ygg-pg 0.3.6 image passes all 11 Red Hat container certification checks (preflight 1.22.0): 0 failed, 0 errors.

Supported by the raw data

How it is computed
  • checks.passed = 11 → 11
  • checks.failed = 0 → 0
  • checks.errors = 0 → 0
  • passed = yes → yes

Recompute it yourself: make claim-image-preflight

Benchmarks

The same gate, with and without MCP running

One model (base Qwen3-8B, temperature 0), 13 tasks (6 everyday, 7 traps) and a superuser oracle that compares every table before and after each task. Only the connection changes: native (one SQL statement through pg_agent_gate), native-loop (the same, agentic), gated-mcp (the same gate over MCP) and stock-mcp (a popular Postgres MCP server with the same privileges and no gate). Reported per arm: good tasks done, trap breaches, prompt tokens, model calls, median seconds per task, errors.

Rerun: gh workflow run compare.yml -R <your-fork>/agent-gate-demo -f config=cited

  • 13 tasks is a demonstration, not a statistical benchmark; repeat=3 shows run-to-run variation.
  • The model server runs with a 4096-token context window. The MCP arms carry tool definitions and tool results inside it; in the cited run one gated-mcp error was a context overflow. A larger window would likely change the MCP arms' numbers.
  • Times come only from clean GitHub runners; times from a shared machine are never cited.
  • The cited run measured pg_agent_gate 0.4.0 on PostgreSQL 18; the public image ygg-pg 0.3.6 ships 0.4.3 on PostgreSQL 19rc2. config=current measures 0.4.3.

Runs

Red Hat container certification checks (preflight) on the public image running

Runs Red Hat's openshift-preflight `check container` against the published ygg-pg image, pulled by digest with no registry credential. Reported: checks passed, failed and errored. Passing preflight is a prerequisite for container certification; it is not the certification.

Rerun: preflight check container ghcr.io/latamflows/ygg-pg@sha256:ec1aabe4be95f2f84d9c06aa612edc6346079dc7ebfe93280332cd7905c0a9f1

  • A pass says the image meets the automated checks; it says nothing about the extensions' behaviour.

Runs

pg_phasevec vs pgvector: recall and latency on the same data designed

recall@10 against the dataset's published ground truth, and per-query latency (p50, p99) and throughput, for pg_phasevec (find, find_ivf) and pgvector (HNSW, IVFFlat) on the same vectors, the same queries, the same PostgreSQL build and the same machine, swept over each index's search knob. See SPEC.md.

Rerun: make recall-phasevec DSN=postgresql://... DATASET=nytimes-256-angular

  • Designed, not run yet: no figure from this benchmark is published.

Runs

  • No runs yet.