Tests: internal/infra/inference

internal/infra/inference · 72 tests

Each test is named for the property it holds, and runs against a real deployment: there is no mock of the database and no arm that skips when it is absent. The sentence is the test's name read back; the name is what go test -run takes.

config_test.go

• No message content leaves for an unlisted host — TestNoMessageContentLeavesForAnUnlistedHost. ── The egress guard ────────────────────────────────────────────────────────────────────────── Extraction sends a customer's message text to a model.
• A trailing slash on the endpoint is trimmed — TestATrailingSlashOnTheEndpointIsTrimmed. A trailing slash on the endpoint must not produce a doubled one in the path.
• Plaintext is permitted only to this machine — TestPlaintextIsPermittedOnlyToThisMachine. Plaintext is permitted to the machine the process is on, and to nothing else.
• The allowlist is host and port rather than AURL — TestTheAllowlistIsHostAndPortRatherThanAURL. The allowlist is matched on host and port, so an entry carrying a scheme matches nothing.
• Embedding can go somewhere else and still has to be listed — TestEmbeddingCanGoSomewhereElseAndStillHasToBeListed
• One provider needs one endpoint — TestOneProviderNeedsOneEndpoint. One provider is the ordinary case and needs one set of variables.
• A second endpoint obeys every rule the first one does — TestASecondEndpointObeysEveryRuleTheFirstOneDoes. A second endpoint on a host nobody listed is refused for the same reason the first one is: it is an egress of exactly the data this product exists to hold, and the list is what makes permitting it a deliberate act.

corpus_test.go

• The extraction corpus — TestTheExtractionCorpus

embedder_test.go

• Vectors come back against the input they are for — TestVectorsComeBackAgainstTheInputTheyAreFor. The order is a promise.
• An embedding reply that cannot be trusted is refused — TestAnEmbeddingReplyThatCannotBeTrustedIsRefused. Every way the reply can leave a caller matching a vector onto the wrong thing.
• An unconfigured embedder refuses rather than returning zeros — TestAnUnconfiguredEmbedderRefusesRatherThanReturningZeros. A deployment that extracts and never embeds is a working deployment.
• An empty input is not sent — TestAnEmptyInputIsNotSent. An empty surface embeds to whatever the model does with nothing, which is a vector that matches things for no reason.
• Similarity is an angle rather than a length — TestSimilarityIsAnAngleRatherThanALength. Normalised, because a dot product used without normalising rewards long vectors, and whether a model returns normalised vectors is a property of the model rather than of the interface.
• A meaningless comparison is an error rather than zero — TestAMeaninglessComparisonIsAnErrorRatherThanZero. A comparison that is meaningless says so rather than returning zero, which would read as "unrelated" — and a caller thresholding on it would quietly treat every mismatch as a rejection.

embeddingbounds_test.go

• Embedding rejects a zero direction before returning vectors — TestEmbeddingRejectsAZeroDirectionBeforeReturningVectors
• Cosine rejects non finite components instead of returning na n — TestCosineRejectsNonFiniteComponentsInsteadOfReturningNaN
• Embedding input limits refuse before network and admit boundaries — TestEmbeddingInputLimitsRefuseBeforeNetworkAndAdmitBoundaries
• Embedding dimension bounds and finite extreme cosines — TestEmbeddingDimensionBoundsAndFiniteExtremeCosines
• Embedding reply overflow is not accepted as a valid prefix — TestEmbeddingReplyOverflowIsNotAcceptedAsAValidPrefix. A complete JSON value followed by whitespace is still a valid prefix.

embeddingconfig_test.go

• Embedding only configuration binds its own allowed endpoint — TestEmbeddingOnlyConfigurationBindsItsOwnAllowedEndpoint. A read-only service validates only the endpoint it uses, without borrowing a different provider's credential or weakening the generation process's independent validation.

golden_capture_test.go

• Capture golden — TestCaptureGolden. Regenerates the golden after a DELIBERATE wording change, run with CAPTURE=1.

identity_test.go

• Extraction identity changes with model origin prompt and implementation but never keys — TestExtractionIdentityChangesWithModelOriginPromptAndImplementationButNeverKeys
• Report identity changes with model origin prompt and implementation but never keys — TestReportIdentityChangesWithModelOriginPromptAndImplementationButNeverKeys. A report's identity moves with the model, the origin, the prompt and the compiled code that composes and parses — and with nothing else.

live_test.go

• A real models claims are all citable — TestARealModelsClaimsAreAllCitable. ── What extraction refused is kept, against a model free to paraphrase ─────────────────────── The unit tests prove the span rule holds for proposals a stub was told to produce.

livedescription_test.go

• Whether a described entity separates where an assembled one did not — TestWhetherADescribedEntitySeparatesWhereAnAssembledOneDidNot

liveneighbourhood_test.go

• Whether an entitys neighbourhood is what distinguishes it — TestWhetherAnEntitysNeighbourhoodIsWhatDistinguishesIt

livereport_test.go

• A live report says what the material said and nothing else — TestALiveReportSaysWhatTheMaterialSaidAndNothingElse

liveresolution_test.go

• What a semantic surface resolves and what it would merge — TestWhatASemanticSurfaceResolvesAndWhatItWouldMerge

livespeaker_test.go

• Live speakers are bound across languages and corrections — TestLiveSpeakersAreBoundAcrossLanguagesAndCorrections. This runs the production vocabulary gate and formation path, not only the model's proposal parser.

livesurfacecontract_test.go

• Each embedding surface wins only its own retrieval question — TestEachEmbeddingSurfaceWinsOnlyItsOwnRetrievalQuestion

livethemes_test.go

• Whether a thematic question finds the right subject — TestWhetherAThematicQuestionFindsTheRightSubject

parse_test.go

• A replied document is found behind a prefix the model invented — TestARepliedDocumentIsFoundBehindAPrefixTheModelInvented. The exact reply that made a correct extraction fail.
• A claim without its envelope is an error rather than an empty result — TestAClaimWithoutItsEnvelopeIsAnErrorRatherThanAnEmptyResult. A single claim emitted without its envelope must NOT be read as "nothing was asserted".
• An empty claim list is a result and not a failure — TestAnEmptyClaimListIsAResultAndNotAFailure. An empty list means the message asserts nothing, which is the ordinary case and must stay distinguishable from a reply that could not be read.
• Commentary after the document is ignored — TestCommentaryAfterTheDocumentIsIgnored. Prose after the document does not fail the parse, because the decoder reads one value and stops.
• A fenced document is still read — TestAFencedDocumentIsStillRead. A fenced document still works, which is the case that was already handled and must not regress.
• A reply with no envelope is an error — TestAReplyWithNoEnvelopeIsAnError. A reply with no envelope anywhere is an error, not an empty result.
• The error quotes the reply without letting it run away — TestTheErrorQuotesTheReplyWithoutLettingItRunAway. The error carries the reply, bounded.

prompt_test.go

• The message is fenced with a value it could not contain — TestTheMessageIsFencedWithAValueItCouldNotContain. ── Directive framing, the part that needs no ruling ────────────────────────────────────────── A message arrives as data inside a fence the message could not have contained.
• The fence is not reused between calls — TestTheFenceIsNotReusedBetweenCalls. The fence is different every time, so it cannot be learned from one call and used in the next.
• A fence that occurs in the content is regenerated — TestAFenceThatOccursInTheContentIsRegenerated. A fence that somehow occurs in the content is regenerated rather than used.
• Every instruction precedes the content — TestEveryInstructionPrecedesTheContent. The instructions end before the content begins.

promptfile_test.go

• The composed prompt is byte identical to the text it replaced — TestTheComposedPromptIsByteIdenticalToTheTextItReplaced
• The vocabulary is rendered where the file marks it — TestTheVocabularyIsRenderedWhereTheFileMarksIt. The vocabulary is rendered from the table, in the table's order, and nowhere else.
• A file missing a section is refused — TestAFileMissingASectionIsRefused. ── What the file is not allowed to do ──────────────────────────────────────────────────────── The file holds wording.
• An empty section is refused — TestAnEmptySectionIsRefused. A section present and blank is the same defect as a section absent, and is easier to produce by accident — a paragraph deleted from a file that still parses.
• The vocabulary is marked exactly once — TestTheVocabularyIsMarkedExactlyOnce. None means the model is told the vocabulary is closed and never shown it, so every extraction is unmapped.
• The request uses the model contract from the file — TestTheRequestUsesTheModelContractFromTheFile. The request carries what the file says it carries.
• A file that is not YAML is refused — TestAFileThatIsNotYAMLIsRefused
• A contract without a response format is refused — TestAContractWithoutAResponseFormatIsRefused

redirect_test.go

• A model endpoint cannot redirect content to a host nobody listed — TestAModelEndpointCannotRedirectContentToAHostNobodyListed. ── The allowlist names where content goes, not where it goes first ─────────────────────────── A listed endpoint that answers with a redirect is asking for the request body — message content, a report's facts, a history to summarise — to be sent again to a host nobody listed.

reporter_test.go

• A report request carries the contract and the material — TestAReportRequestCarriesTheContractAndTheMaterial
• A model that refuses is a refusal rather than an empty report — TestAModelThatRefusesIsARefusalRatherThanAnEmptyReport
• An enormous reply does not decide how much memory is spent — TestAnEnormousReplyDoesNotDecideHowMuchMemoryIsSpent. A reply is read up to a bound, because a remote system does not decide how much memory this process spends on one answer.
• A cancelled report stops — TestACancelledReportStops. A cancelled request stops rather than running to completion somewhere nobody is waiting.
• Report input distinguishes equal speaker names without changing quotes — TestReportInputDistinguishesEqualSpeakerNamesWithoutChangingQuotes

reportprompt_test.go

• A report prompt missing a section is refused — TestAReportPromptMissingASectionIsRefused. The file holds wording.
• A blank report section is refused — TestABlankReportSectionIsRefused. A section present and blank is the same defect as one absent, and easier to produce by accident.
• A report prompt without a response format is refused — TestAReportPromptWithoutAResponseFormatIsRefused
• Reports are not a creative task — TestReportsAreNotACreativeTask. Two rebuilds of one community should say the same thing about what survived.
• The report boundary is the last thing the model is told — TestTheReportBoundaryIsTheLastThingTheModelIsTold. The boundary is last, because its job is to be the final word before somebody else's text arrives.
• The fence is absent from every part of the material — TestTheFenceIsAbsentFromEveryPartOfTheMaterial. The material is fenced with a value it cannot contain, chosen after the material was assembled — including the material inside child reports, which is text from a model that read somebody's quotes.
• The same context renders the same material — TestTheSameContextRendersTheSameMaterial. Two renderings of one context are identical, so two rebuilds of a community ask the same question.
• A report is found inside whatever the model sent — TestAReportIsFoundInsideWhateverTheModelSent. A model in JSON mode still wraps its answer in prose or a code fence often enough that a plain Unmarshal turns a good report into a failure.
• A reply with no report is an error rather than an empty report — TestAReplyWithNoReportIsAnErrorRatherThanAnEmptyReport. An object that is not a report is skipped rather than returned as a report with everything empty — which would then be refused for having no title, reporting the wrong problem.
• The decode error does not quote the whole reply — TestTheDecodeErrorDoesNotQuoteTheWholeReply. The error carries a bounded amount of the reply, because the reply can contain somebody's words.

reportrevision_test.go

• Report source revisions never enter model material — TestReportSourceRevisionsNeverEnterModelMaterial. Source revisions authorize persistence; adding them must not change what leaves for inference.

summariser_test.go

• The compaction prompt loads and ends on the boundary — TestTheCompactionPromptLoadsAndEndsOnTheBoundary. The compaction prompt loads with the report prompt's loader, and its system prompt ends on the data boundary: whatever the material says, the last instruction the model reads is that the material is data.
• Material is rendered fenced and in order — TestMaterialIsRenderedFencedAndInOrder. The material is rendered fenced, oldest first, with who said what; the fence is chosen against every byte of the material so the material cannot close it.
• The summary is located in the reply and an empty one is refused — TestTheSummaryIsLocatedInTheReplyAndAnEmptyOneIsRefused. The reply is located inside whatever the model wrapped it in, and an empty summary is refused: a segment that says nothing would stand in for turns and lose them.
• The summariser asks the model under the files contract — TestTheSummariserAsksTheModelUnderTheFilesContract. The request carries the file's model contract and the fenced material, and the provider's answer becomes the summary; a provider that fails is an error, not an empty segment.

timeout_test.go

• The model client timeout follows the configured budget — TestTheModelClientTimeoutFollowsTheConfiguredBudget. The client's own timeout must follow the configured budget: if it stayed at a lower constant, the deadline that fired would not be the one the operator set, and the failure would misreport it.