Budgeted Review

Motivation

Budgeting prevents a review system from becoming an uncontrolled token sink. Exhausted work is recorded, not hidden.

Theory

Evidence risk review treats an answer as a set of claims C={c1,c2,...,cn}C = \{c_1, c_2, ..., c_n\} and sources S={s1,s2,...,sm}S = \{s_1, s_2, ..., s_m\}. Each source receives a tier rank r(s)∈[0,100]r(s) \in [0,100]. A claim with assertiveness a(c)a(c) has a profile-specific minimum rank t(a,c)t(a,c). A deterministic gap is:

g(c)=max⁡(0,t(a,c)−max⁡s∈Scr(s)) g(c) = \max(0, t(a,c) - \max_{s \in S_c} r(s))

The package then combines evidence gaps, profile checks, and verdict precedence into a stable ReviewResult.

Design + Diagram

flowchart LR A[Artifact] --> B[Evidence tier labeling] B --> C[Risk sweep] C --> D[Verdict reduction] D --> E[ReviewResult] C --> F[(Review log)] G[Domain profile] --> C H[Budget meter] --> C

Data Model / Contract

ReviewBudget caps LLM calls, tokens, heavy checks, and wall time.

Field Meaning
artifact_id Host-defined stable identifier.
claims Discrete claims extracted by the host.
sources Cited evidence with optional tier hints.
profile_key Risk profile used for thresholds and checks.
findings Structured review findings emitted by checks.

ADR

Problem: adapters tend to grow business logic

Decision: keep PHP facade, Artisan, HTTP, and MCP as thin adapters over ReviewEngine.

Consequences: there is one behavior to test, but adapters must validate input carefully.

Problem: external review can become expensive

Decision: run cheap deterministic checks first and spend LLM budget only when enabled and useful.

Consequences: default installs have zero token cost, but hosts must bind an LLM contract for semantic review.

Worked Example

'budget' => ['max_llm_calls' => 3, 'max_tokens' => 6000],

Gotcha / Limits

The package reviews evidence strength and risk boundaries; it does not retrieve sources, extract claims automatically, or replace human review for regulated decisions.

previous Risk Sweep next Review Logs