# PDR 000: Own review policy and evidence, rent model inference

- HTML version: https://robbiepalmer.me/projects/agentic-code-review/pdrs/000-own-review-policy-and-rent-inference
- Project: Agentic Code Review
- Status: Accepted
- Opened: 2026-06-21
- Decided: 2026-07-25

## Context

Agentic development creates review demand faster than a solo maintainer or a
fixed third-party allowance can absorb it. The project tried five managed
reviewers. Their findings were useful, but quotas, product withdrawals, shared
subscription limits, and broader enterprise bundles made dependable capacity
hard to buy at the required pace.

The parts that compound are specific to this development loop: repository
policy, Pull Request history, finding dispositions, cost records, and evidence
about which models find useful problems. A model or inference provider can
change without making that history less valuable.

The custom prototype proposed on 21 June showed that a narrow multi-model
reviewer could produce useful complementary findings within the comparison
budget. By 25 July, it had produced better observed findings per dollar than
the evaluated managed services. Its remaining weakness was stateless
orchestration, not the decision to own the review layer.

## Decision

Own the review-specific product layer: repository policy, review timing, Pull
Request state, finding lifecycle, dispositions, budgets, and evaluation data.
Rent model inference behind a stable finding contract and keep both models and
providers replaceable.

The initial product serves the owner's public repositories. It remains an
advisory reviewer rather than a coding agent, merge authority, static-analysis
bundle, or general software-development platform. Managed reviewers may remain
supplementary inputs or a fallback, but no vendor-private learning store or
mutable allowance becomes the product's core.

Select models using observed downstream value, coverage, cost, latency, and
failure behavior. A new model should join through configuration and an adapter,
then earn its place through replay and live evidence. Model count, benchmark
reputation, and open-weight status have no value on their own.

## Consequences

Review capacity and accumulated evidence no longer depend on one specialist
review vendor. The product can change its model mix as quality, price, and
availability move.

The project now owns GitHub integration failures, prompt and schema changes,
provider outages, cost control, privacy boundaries, and false-positive policy.
That maintenance is justified only while the focused reviewer saves more work
or produces better evidence than buying a managed service.

The narrow initial audience delays private repositories, teams, billing,
tenant isolation, compliance controls, and general administration. Each would
need evidence and a later product decision.

## Evaluation

Keep the owned layer while it provides dependable review at the development
pace, preserves portable outcome evidence, and beats the credible managed
fallback on useful findings after cost and maintenance are counted. Reconsider
the decision if operating the service costs more attention than it saves or if
a managed reviewer offers equivalent capacity and evidence portability.

## Migration note

This record separates the product boundary first proposed in the historical
Custom Agentic Code Review ADR from its GitHub Actions and OpenRouter design.
The proposal became the active product direction on 25 July when the stateful
successor recorded the prototype's observed value and retained the owned review
layer. The historical acceptance date predates this PDR migration. ADR 000
retains the deprecated prototype architecture, and ADR 001 retains the
proposed stateful architecture.

## Related records

- Project: [Agentic Code Review](https://robbiepalmer.me/projects/agentic-code-review.md)
- Informed by ADR: [ADR 009: CodeRabbit](https://robbiepalmer.me/projects/personal-knowledge-graph/adrs/009-code-rabbit.md)
- Informed by ADR: [ADR 041: Gemini Code Assist](https://robbiepalmer.me/projects/personal-knowledge-graph/adrs/041-gemini-code-assist.md)
- Informed by ADR: [ADR 042: Greptile](https://robbiepalmer.me/projects/personal-knowledge-graph/adrs/042-greptile.md)
- Informed by ADR: [ADR 044: Codex Code Review](https://robbiepalmer.me/projects/personal-knowledge-graph/adrs/044-codex-code-review.md)
- Informed by ADR: [ADR 045: Qodo](https://robbiepalmer.me/projects/personal-knowledge-graph/adrs/045-qodo.md)
- Implemented by ADR: [ADR 000: Custom Agentic Code Review](https://robbiepalmer.me/projects/agentic-code-review/adrs/000-custom-agentic-code-review.md)
- Implemented by ADR: [ADR 001: Stateful AI Code Review](https://robbiepalmer.me/projects/agentic-code-review/adrs/001-stateful-ai-code-review.md)
- Implemented by ADR: [ADR 002: DuckDB for the AI Review Scorecard](https://robbiepalmer.me/projects/agentic-code-review/adrs/002-duckdb-ai-review-scorecard.md)
- Implemented by ADR: [ADR 006: ChatGPT Plan Models as Complementary Review Scouts](https://robbiepalmer.me/projects/agentic-code-review/adrs/006-chatgpt-plan-models-as-complementary-review-scouts.md)

---

Markdown index of this site: https://robbiepalmer.me/llms.txt