Preparation Review

Review follows the data from source identity to downstream handoff. Begin at the changed invariant, then inspect every serialization or execution boundary the value crosses.

flowchart TD
    input[Source and configuration]
    transform[Filter, clean, chunk]
    execute[Embed and schedule]
    retain[Persist records and observations]
    consume[Index, retrieve, cite]

    input --> transform --> execute --> retain --> consume

Review one downstream record backward

Do not begin with aggregate counts. Choose one record that a consumer can retrieve or cite and reconstruct its custody chain:

Stop	Confirm	Refuse the review when
candidate or citation	rank, score, index identity, chunk identity, and cited span	the result contains text but no stable prepared-material reference
chunk	exact text, start/end offsets, parent identity, geometry, and tail disposition	offsets cannot resolve against the retained normalized parent
normalized document	canonical fields, applied rules, safeguard outcome, and clean identity	normalization behavior depends on unrecorded defaults or locale
source record	source ID, adapter, original field mapping, and input digest	two distinguishable sources collapse into one identity
run configuration	normalized configuration and injected component identities	a mutable path or process environment is the only record of behavior

Repeat the walk for a rejected source. Its path should end in a stable error with stage, position, cause, and termination status—not in an absent output.

Challenge the preparation boundary

Use adversarial cases that target the changed invariant rather than merely adding more ordinary documents:

• two sources whose text normalizes identically but whose identities must stay distinct;
• text that lands exactly on and one unit beyond a chunk boundary;
• overlap equal to chunk size, negative geometry, and each tail policy;
• one failing record between successful records under fail-fast and error-collecting execution;
• interruption while a retry, breaker, cache write, or output write is active;
• persisted state produced by an older format or different normalized configuration; and
• a citation whose parent bytes change without its span changing.

The expected result may be a typed refusal. The review fails only when the system accepts ambiguous state, loses custody, or reports a stronger outcome than the retained evidence supports.

Identity and transformation

• Is document identity derived before transformations that could erase source distinctions?
• Do normalization, chunk size, overlap, and tail policy participate in the retained configuration?
• Are spans interpreted against normalized text, and can every retained span be resolved?
• Does a change to identity or ordering include comparison against the prior behavior rather than only new happy-path fixtures?

Execution and failure

• Does lazy execution preserve order, bounded buffering, termination, and backpressure?
• Are retries, breakers, caches, and resource lifetimes explicit at the point where they change behavior?
• Can partial failure be distinguished from an empty successful corpus?
• Do injected readers, cleaners, embedders, stores, and clocks have recorded identity when reproducibility depends on them?

Public and persisted boundaries

• Do strict models reject unknown fields and incompatible vector dimensions?
• Do CLI and HTTP adapters preserve typed error meaning and nonzero failure status?
• Does persisted state reject incompatible formats and stale configuration?
• Are secrets, sensitive source text, and unbounded samples absent from logs and diagnostics?

Retrieval interpretation

• Is the hash embedder described only as a deterministic baseline?
• Does a quality claim name its corpus, model profile, metrics, and threshold?
• Do citations retain chunk identity and resolvable span data?
• Is governed backend selection and replay left to bijux-canon-index rather than absorbed into ingest?

A review is complete when the evidence explains both accepted material and rejected material. Route the result through release acceptance and check unresolved exposure in known limitations.