Interpreting Preparation Evidence

An ingest result is evidence of a declared transformation, not certification of the source or the downstream decision. Interpret it by following the retained source identity, configuration, transformations, output records, and terminal observations together.

flowchart LR
    source["source descriptor + digest"]
    config["preparation configuration"]
    stages["clean + chunk + embed + deduplicate"]
    records["prepared records + identities"]
    observations["counts + failures + metrics"]

    source --> config --> stages --> records --> observations

Read One Preparation Result

Review question	Evidence to retain	Failure if missing
Which input was admitted?	source URI or record key, content digest, reader and adapter identity	a later run cannot distinguish source drift from transformation drift
Which text was transformed?	normalized-text digest and cleaning configuration	chunk spans cannot be related to the exact normalized text
How were chunks created?	chunking parameters, tail policy, order, offsets, chunk identities	matching prose may conceal different segmentation
Which vectors were produced?	embedder, model or implementation revision, dimension, device and parameters	vector compatibility and reproducibility are unknown
What was removed?	deduplication rule, input/output counts, rejection or observation records	absence can be mistaken for source absence
What was published?	saved-index envelope or output artifact, schema version, fingerprint, terminal status	partially written state can look complete

Bounded Claim Vocabulary

Claim	Required evidence	Bound on the claim
deterministic cleaning	identical normalized text for the same input, implementation, and configuration	caller-supplied or version-drifting cleaners join the boundary
stable chunk identity	matching document identity, offsets, text, parameters, and digest	any identity-bearing input can change the result
ordered preparation	stable document and span order through the selected pipeline	does not imply distributed exactly-once delivery
validated embedding	vector shape and dimension match the declared specification	does not establish semantic usefulness
replayable preparation	source, configuration, adapter identity, and retained outputs remain available	external services and mutable source systems can still drift
retrieval baseline	metrics reproduce for the named corpus, query set, and model profile	does not generalize to another corpus or production model
citation retained	answer output resolves to a stored chunk and span	does not establish source authority or entailment

Interpret Identity Correctly

A digest detects identity or change; it does not prove correctness. A schema defines an accepted representation; it does not prove that an HTTP service is deployed. A deterministic hash embedder is a repeatable reference profile; it does not measure semantic retrieval quality. A citation resolves an answer to a chunk; it does not prove that the chunk supports the answer.

Offsets refer to normalized Python strings unless an artifact explicitly retains original byte positions. Structural deduplication compares the selected record equality or key; it does not find paraphrases or copied ideas. BM25 and NumPy cosine scores belong to different ranking models and must not be compared as a shared confidence scale.

Optional Integrations Widen The Boundary

When a model, source reader, storage adapter, clock, caller-supplied stage, or application service enters the path, retain its name, version, configuration, failure behavior, and relevant external state. If any of those identities is absent, reproducibility is unknown rather than implied by the package's local deterministic stages.

Continue with invariants for machine-enforced laws, known limitations for boundaries that remain after those laws pass, and the risk register for failure signals and controls.