Polish decision-model examples

Three live Jev requests on synthetic architecture examples, 2026-10-01. These demonstrate integration, not general recommendation accuracy.

The numbers below are uncalibrated model choice probabilities. A value of 1 does not mean guaranteed correctness. Statistical confidence intervals are unavailable: no independent labeled evaluation set was supplied.

connection-pools

Preserve eight application instances and the existing database connection limit. Prefer changing the application connection-pool setting if it satisfies the declared workload.

smaller-pools: Keep eight application instances and reduce each pool from 50 to 30 connections; reduces pool headroom. — eligible
fewer-instances: Reduce application instances from eight to four, retaining pools of 50; reduces modeled compute headroom. — eligible
unchanged: Keep the failing eight instances with pools of 50. — rejected by simulator

Selected: smaller-pools. Model: jev-1.13.0. Calls: 1.

{
  "fewer-instances": 0,
  "smaller-pools": 1
}

Original proposal: FAIL (introduced). Both eligible alternatives pass the declared scenarios. No changes were applied.

Full execution evidence (JSON)

cpu-capacity

Satisfy the declared workload while retaining the existing two-vCPU host size. Prefer adding an instance over changing host size.

more-instances: Use two instances, preserving the current two-vCPU host size. — eligible
larger-host: Use one four-vCPU host; preserves instance count but changes host size. — eligible
unchanged: Keep the overloaded single two-vCPU instance. — rejected by simulator

Selected: more-instances. Model: jev-1.13.0. Calls: 1.

{
  "larger-host": 0,
  "more-instances": 1
}

Original proposal: FAIL (introduced). Both eligible alternatives pass the declared scenarios. No changes were applied.

Full execution evidence (JSON)

stream-retention

Cover the declared 21-second lag, choosing the smallest offered passing retention allocation. Do not assume memory requirements are verified.

retain-240: Retain 240 messages (30 seconds at eight messages/sec); smaller retention allocation than 480. Memory bytes are not modeled here. — eligible
retain-480: Retain 480 messages (60 seconds at eight messages/sec); more lag headroom, with greater unmeasured memory requirements. — eligible
unchanged: Retain 160 messages, which cannot cover the declared 21-second lag. — rejected by simulator

Selected: retain-240. Model: typesafe/jev-1.13-20260917. Calls: 1.

{
  "retain-240": 1,
  "retain-480": 0
}

Original proposal: FAIL (introduced). Both eligible alternatives pass the declared scenarios. No changes were applied.

Full execution evidence (JSON)

Local Laya: connection pools

CPU inference with outbound socket connections blocked completed with a 180-second limit after a 90-second startup timeout.

Laya selected fewer-instances (0.8544), versus smaller-pools (0.1456). This is a passing modeled alternative but a poor match for the objective to preserve eight instances. Integration succeeded; this example does not demonstrate good goal alignment.

Full local Laya evidence

Laya checkpoint comparison

The browser and general English checkpoints received identical objective, candidate descriptions, and compact state. Both ran locally on CPU.

Checkpoint	Smaller pools	Fewer instances
Browser	14.56%	85.44% (selected)
General English	39.41%	60.59% (selected)

Both selections conflict with the objective to preserve eight instances. A greater probability for the preferred choice did not improve the selected answer. One case does not establish general model quality.

General checkpoint evidence · Comparison and fingerprints