Anthropic CCA-F Questions & Answers

Consumers cannot know which contract applies. A union of undocumented old and new shapes leaves consumers unable to apply the right validation. A new field becomes mandatory but old consumers still deserialize the prior schema. Instead, publish a versioned schema contract and provide a compatible migration path to consumers. Making a new schema field mandatory changes the contract consumed by deployed readers.

Old deserializers may fail on the new required field. The new producer could emit a field older deserializers reject or misinterpret. A new field becomes mandatory but old consumers still deserialize the prior schema. Instead, publish a versioned schema contract and provide a compatible migration path to consumers. Versioning and a compatibility migration let old applications continue decoding while new producers and consumers coordinate.

Fabrication undermines provenance and business semantics. Default text masquerades as an observed value when the new field has no source. Versioning and a compatibility migration let old applications continue decoding while new producers and consumers coordinate. A new field becomes mandatory but old consumers still deserialize the prior schema.

The field should reflect actual optional notes only. Footer boilerplate is not equivalent to the explicit notes the field is meant to capture. An optional notes field can be omitted entirely when the underlying document contains no notes. Instead, permit absent notes rather than inventing filler text.

Mandatory filler can be mistaken for source content. A textual N/A may be mistaken for an actual note authored in the source. A tax form has an optional notes field and some records have no notes at all. Instead, permit absent notes rather than inventing filler text.

Valid records should not be dropped for nonrequired text. Rejecting forms without notes contradicts the field's optional contract. An optional notes field can be omitted entirely when the underlying document contains no notes. A tax form has an optional notes field and some records have no notes at all.

The model does not learn which date representation failed. An unchanged request repeats the same date formatting failure without showing the expected representation. An invoice extractor returns invalid YYYY/MM/DD format and silently repeats the same prompt. Instead, send the failed field plus expected ISO format in the retry context.

A new value would be invented rather than normalized. The machine clock does not represent the date printed on the invoice. A useful extraction retry carries the failing field path, observed value and required ISO date format. An invoice extractor returns invalid YYYY/MM/DD format and silently repeats the same prompt.

The required output contract is still violated. Accepting YYYY/MM/DD undermines consumers that require ISO calendar-date strings. A useful extraction retry carries the failing field path, observed value and required ISO date format. Instead, send the failed field plus expected ISO format in the retry context.

Schema design pressures the model to hallucinate. An always-required string shape forces fabricated information when the source genuinely omits a field. The source contract has no effective date and three retries hallucinate different dates. Instead, record missing evidence and request additional source instead of repeating generation.

An invented date is still not sourced. A low confidence label does not turn an invented date into documentary evidence. The source contract has no effective date and three retries hallucinate different dates. Instead, record missing evidence and request additional source instead of repeating generation.

No missing source evidence will be created by sampling. Changing sampling randomness changes guesses but does not add a missing contractual date. The source contract has no effective date and three retries hallucinate different dates. Instead, record missing evidence and request additional source instead of repeating generation.

Syntactic correctness cannot reconcile line-item arithmetic. A JSON parser can succeed on numbers whose arithmetic is incompatible with the document. An invoice may be well formed while its line-item arithmetic conflicts with the reported total. A structured invoice object passes schema checks but its line-item sum differs. Instead, calculate the line-item aggregate and compare it to the stated source total with discrepancy metadata.

Deleting evidence would corrupt the document interpretation. Deleting line items to force agreement alters the source accounting evidence. A structured invoice object passes schema checks but its line-item sum differs. Instead, calculate the line-item aggregate and compare it to the stated source total with discrepancy metadata. A deterministic sum plus discrepancy metadata distinguishes calculation failure from a transcription or source-data error.

The mismatch must remain traceable for adjudication. Silently rewriting the total eliminates the original discrepancy needed by a reviewer. A deterministic sum plus discrepancy metadata distinguishes calculation failure from a transcription or source-data error. A structured invoice object passes schema checks but its line-item sum differs.

The team cannot measure the noisy pattern reliably. Without a typed detected pattern, repeated false positives cannot be quantified or tuned. Developers reject many alerts about generated code but the bot cannot group them. Instead, include detected_pattern metadata to analyze false positives.

Some generated-code changes can still contain real defects. Generated code is not automatically defect-free; suppressing all findings would hide legitimate failures. Developers reject many alerts about generated code but the bot cannot group them. Instead, include detected_pattern metadata to analyze false positives.

Escalation does not fix its false-positive mechanism. Inflating severity increases developer fatigue while leaving a weak detection heuristic untouched. Developers reject many alerts about generated code but the bot cannot group them. Instead, include detected_pattern metadata to analyze false positives.

Subjective confidence does not test identity relationships. Confidence metadata is not an independently checked correspondence between employee and employer IDs. Semantic validation examines business field identities and relationships in addition to parseable JSON structure. The JSON parses but the employee and company identifiers are swapped.

Identifiers can have valid types while occupying wrong semantic fields. A syntactically valid ID is still wrong when assigned to another entity type. The JSON parses but the employee and company identifiers are swapped. Instead, run domain/field consistency validation beyond syntax.

Late cosmetic changes preserve the underlying misattribution. Post-ingestion renaming does not restore the correct source-to-field relationship. The JSON parses but the employee and company identifiers are swapped. Instead, run domain/field consistency validation beyond syntax. Semantic validation examines business field identities and relationships in addition to parseable JSON structure.

The model cannot target the known validator issue efficiently. A source without the validation rule leaves the model guessing what failed and which format to use. An extraction retry needs the original document and precise validator feedback, not just a malformed output fragment. Instead, include original source evidence in retry context and include the precise validation failure.

The source facts needed for correction are missing. The invalid response alone lacks documentary evidence needed to reconstruct the correct field. A retry includes only the invalid JSON and not the source document. Instead, include original source evidence in retry context and include the precise validation failure. Without source context the model cannot reconstruct the correct value and may learn only to change punctuation.

A retry can repeat or create new errors. An unvalidated retry can introduce new problems even if its output is well formed. Without source context the model cannot reconstruct the correct value and may learn only to change punctuation. Instead, include original source evidence in retry context and include the precise validation failure.

The unreadable input would falsely appear processed. A clean empty object would falsely mark the document as processed instead of inaccessible. An extraction model loops on an encrypted scanned record with no usable content. Instead, bound attempts and route unreadable inputs for manual handling.

Similar-looking forms cannot establish this record's values. Similar invoices do not establish the amount or party on this particular encrypted file. An extraction model loops on an encrypted scanned record with no usable content. Instead, bound attempts and route unreadable inputs for manual handling.

No accessible content exists for the model to extract. Repeated attempts against unreadable ciphertext cannot extract information without the appropriate key. An extraction model loops on an encrypted scanned record with no usable content. Instead, bound attempts and route unreadable inputs for manual handling.

Discarding evidence conceals the inconsistency. Removing the header destroys the evidence explaining why the total-line value was disputed. Contradictory currency signals require a reviewable conflict marker with the actual header and total-line locations. A document says USD in header and EUR on the total line.

Normalization does not resolve contradictory source labels. An invented exchange rate changes financial meaning without resolving which currency was intended. A document says USD in header and EUR on the total line. Instead, set conflict_detected with source locations for review.

An inconsistent document needs source adjudication. The last occurrence could be a typo, and ordering is not an authorized currency rule. A document says USD in header and EUR on the total line. Instead, set conflict_detected with source locations for review.

The required fact remains unsupported. The required field remains unsupported and should not receive a false validation receipt. A validator requires net_amount but no such value appears in the uploaded bill. Instead, mark unrecoverable missing data instead of inserting zero.

Zero is an affirmative monetary claim, not a missing marker. A claimed zero amount may incorrectly affect payments, totals or financial controls. A validator requires net_amount but no such value appears in the uploaded bill. Instead, mark unrecoverable missing data instead of inserting zero.

The supposed calculation has no valid operands. A date has no mathematical relationship enabling derivation of an invoice's net amount. A validator requires net_amount but no such value appears in the uploaded bill. Instead, mark unrecoverable missing data instead of inserting zero.