Quality

Trust in ingest begins with stable value laws and ends with a retrieval output whose identity, configuration, and citations can be inspected. Green interface tests cannot compensate for drift in normalization, chunk identity, ordering, or embedding semantics.

Evidence chain

flowchart LR
    values["value and result laws"]
    stages["stage invariants"]
    composition["pipeline and stream properties"]
    boundaries["CLI, HTTP, persistence"]
    evaluation["offline retrieval evidence"]
    limits["known limits and residual risk"]

    values --> stages --> composition --> boundaries --> evaluation --> limits

Claims and proof

Trust claim	Focused evidence	Residual limit
records reject invalid shape	core, result, and public-model unit tests	validity does not establish source truth
cleaning and chunking are deterministic	processing tests plus identity/span invariants	caller-supplied stages join the determinism boundary
lazy execution preserves order and termination	streaming, scheduling, backpressure, and property tests	materialized observations can scale with corpus size
deduplication is stable	rule and pipeline tests over structural keys	semantic duplicates remain distinct
resilience is bounded and visible	retry, breaker, resource, cache, and effect tests	policies work only where explicitly composed
public surfaces preserve domain meaning	strict-model, serialization, CLI, HTTP, and schema tests	deployment state and security remain application concerns
local retrieval behaves as declared	persisted-index and deterministic evaluation tests	hash embeddings do not prove semantic quality
answers retain usable citations	truthfulness gate and answer-path tests	citation presence does not prove source authority

High-risk changes

Changes to normalization, spans, chunk identity, ordering, deduplication, embedding dimensions, persisted index format, citation linkage, or error translation require evidence at the owning layer and at every public boundary they cross. External-model changes also require model-specific evaluation; the deterministic hash profile is not a proxy.

Accept a preparation result

Review a prepared corpus in custody order. Later evidence cannot repair an earlier missing identity:

Review record	Accept when	Refuse or qualify when
source inventory	every intended input has a stable identity and disposition	inputs are silently omitted or identifiers are reused
effective configuration	normalized cleaning, safeguard, and chunk settings are retained	only a mutable configuration path or defaults are known
transformation record	outputs and typed failures account for the inventory	a successful-record count hides rejected, retried, or truncated inputs
chunk set	each chunk names its prepared parent and valid normalized-text offsets	ordering, overlap, tail handling, or parent custody is ambiguous
persisted artifact	records, index state, format, and digest belong to one versioned unit	files can be mixed across runs or overwritten without detection
retrieval observation	query, backend, index, candidates, scores, and citations remain linked	answer text survives without ranked records and citation identities

This review establishes repeatable preparation under the recorded rules. It does not establish source truth, corpus completeness, semantic embedding quality, or downstream claim validity. Record those as separate claims at their owning boundaries.

Evidence routes

Need	Guide
Understand the suite by ownership layer	Test strategy
Review non-negotiable value and pipeline laws	Invariants
Select evidence for a proposed change	Change validation
Review a change consistently	Review checklist
Decide whether work is complete	Definition of done
Evaluate optional and core dependencies	Dependency governance
Understand claims the package cannot make	Known limitations
Inspect unresolved failure modes	Risk register
Interpret preparation artifacts without overstating their guarantee	Interpreting preparation evidence

The appropriate proof is proportional and local: begin with the invariant that owns the behavior, add the crossed boundary, and use corpus evaluation only for retrieval-quality claims.