Keep Deterministic Results Machine-to-Machine

Trigger

The fixed processor produced a complete valid 135-record result, but the model-mediated MCP call submitted only 20 records. Prompt strengthening improved tool selection and fetch behavior but could not guarantee lossless serialization of a large structured payload.

Expected Behavior

Once deterministic code has created a schema-valid result, the receiving system should validate and ingest that exact machine artifact without asking a model to reconstruct the payload as tool arguments.

Actual Behavior

The result crossed an unnecessary model serialization boundary. That boundary introduced truncation, rereads and retry ambiguity even though both the source fetch and deterministic business processing had already succeeded.

Root Cause

Unclear ownership of structured-result transport
A probabilistic orchestration boundary was placed between deterministic producer and deterministic consumer

Evidence

.project-docs/50-evidence/topics/2026-07-30-superagent-fetch-oss-prompt-experiment.md
.project-docs/50-evidence/topics/2026-07-30-arr2-programmatic-pipeline.md
arr_web/programmatic.py
tests/test_arr_programmatic.py

Lesson

Use models to interpret ambiguous inputs or choose work when that is genuinely required. Do not use a model as the transport encoder for a complete, machine-generated business result. Preserve the original artifact, validate it at the system boundary, and make the database commit authoritative.

Action

Updated the architecture through ADR-004.
Added programmatic success/failure vertical slices and active-runtime deployment guards.

Promotion

Promoted to ADR-004, the ARR2.0 architecture documents, deployment contracts and regression tests. No new skill is needed unless this pattern recurs in another project.

1.8 KiB Raw Permalink Blame History