Make each check answer a specific question.

The following is an engineering contract for the workflow we are developing. Each check should produce a report that identifies the record, rule, source version, and failure reason. Passing a later review should not erase a known limitation in an earlier stage.

SHACL provides a useful precedent for declared graph constraints and validation reports. The existence of a formal rule does not prove that the rule captures the intended meaning. Semantic disagreements need examples, review, and sometimes a change to the schema.

“Document 7, paragraph 3” is useful to a reader, but unstable if extraction changes paragraph boundaries. Character offsets also become ambiguous unless the system specifies whether it counts bytes, Unicode code points, or UTF-16 units, and which normalized text version those offsets address.

1. Source version

   Preserve an authorized snapshot or resolvable version reference, content hash, and publication metadata.

2. Extraction version

   Record OCR or parsing tools, normalization rules, page mapping, and the exact text used for offsets.

3. Claim binding

   Store the excerpt, its coordinates, attribution, and the review decision attached to that interpretation.

A content hash can detect a mismatch with an expected byte sequence. It does not prove who authored the document, whether the document is genuine, or whether its claims are true. Those questions require separate controls and evidence. PROV-O offers a vocabulary for recording derivation and responsibility; it does not authenticate the assertions entered into that vocabulary.

Source access also matters. A capsule handed to another person may contain a citation they cannot open. A useful handover should identify those access limits rather than imply that copying the metadata transferred the underlying evidence.

For a missing date, retain “unknown” with the source expression rather than inventing a value. For conflicting identity evidence, keep candidate matches separate. For an invalid source pointer, return the broken reference; a suggested replacement must remain a suggestion until checked.

A retry may fix formatting while changing meaning. Preserve the original proposal, the validation report, and the retry output so an evaluator can detect that tradeoff. Applying a schema migration should be an explicit operation with its own tests, not an invisible side effect of opening an older capsule.

Deterministic processing has a narrower benefit: with pinned inputs and transformations, a reviewer can compare or reproduce particular artifacts. It does not make model generation deterministic, establish semantic equivalence, or certify the result as true.

Context assembly has a selection problem. A record can be perfectly valid and still irrelevant; a short sentence containing an exception can be essential. We propose tagging answer-critical passages before testing selection, then checking whether the assembled context retains them.

• Coverage: include evidence for each required part of the answer, not only the easiest subquestion.
• Counterevidence: include material objections and later corrections even when they have a lower retrieval score.
• Dependency closure: include enough surrounding records to interpret conditions, aliases, and units.
• Boundaries: distinguish source text from instructions; imported documents must not gain authority to direct tools or change review rules.

SARA (2026) studies mixed text and compressed context, and the Graph-RAG reasoning preprint (2026) studies graph-walk selection. Neither establishes that our assembler preserves the evidence a policy decision requires. Our tests must include negation, numbers, exceptions, conflicting sources, and evidence absent from the corpus.

# Proposed experiment receipt; explanatory fields, not a CLI command.
corpus_snapshot: <content hash and access policy>
source_normalizer: <version and offset convention>
question_set: <version; group split by case>
model: <provider and exact model/version>
retriever: <configuration and index snapshot>
context_builder: <commit, schema, prompt, token budget>
review_policy: <rubric, reviewers, adjudication version>
run: <seed where supported; timestamps; retry policy>
outputs: <raw predictions, assembled context, validation reports>
costs: <index/update/query tokens, cache use, latency, review minutes>

A rerun should distinguish exact artifact reproduction from statistical replication. Provider models can change, nondeterminism can remain, and some sources cannot be redistributed. Record those limits, publish allowed fixtures, and retain raw outputs rather than promising byte-identical answers.

Our proposed perturbation tests change one property at a time: move evidence within a long prompt; remove a material source; add a duplicate report; reverse a negation; amend a deadline; or introduce an ambiguous alias. The test should specify the expected response, including when the correct behavior is to ask for more evidence.

For reviewability, give blinded reviewers the same task and measure both decision quality and time to locate the supporting passage. A faster interface that makes errors harder to notice is not an improvement. The benchmark plan separates these outcomes from graph and extraction scores.

Reviewed September 29, 2026. The pipeline and tests above are proposals; implementation and analytical performance need separate evidence.

1. SARA: Selective and Adaptive Retrieval-augmented Generation with Context Compression

   Jin et al. · ACL long paper · July 2026

   Combines retained natural-language passages with learned compressed representations and evidence reranking under token budgets. The approach includes model training and mixed representations; it is not a drop-in JSON compression result.

2. The Reasoning Bottleneck in Graph-RAG: Structured Prompting and Context Compression for Multi-Hop QA

   Zarrinkia et al. · preprint · March 18, 2026 · version 2

   Studies structured prompting and graph-walk compression on KET-RAG and LightRAG across three multi-hop QA benchmarks. Compression helps some configurations and hurts others. The results are not independently reproduced by TACITUS.

3. Shapes Constraint Language (SHACL)

   W3C Recommendation · 2017

   Defines validation of RDF graphs against declared conditions and a vocabulary for validation reports. Conformance to shapes does not establish the truth of the data or the adequacy of the shapes.

4. PROV-O: The PROV Ontology

   W3C Recommendation · 2013

   Provides a vocabulary for representing and exchanging provenance across systems. A provenance record describes derivation and attribution; the record itself still needs trustworthy capture and review.