--- title: "Candidate Score - Evaluate Pre-Extracted Pair" method: POST path: "/candidate-score/evaluate" tags: ["Exato - Candidate Score"] --- # Candidate Score - Evaluate Pre-Extracted Pair `POST /candidate-score/evaluate` Scores an already-extracted candidate against an already-extracted job position. Mode 1 (with jd_extraction_json): 0-100 score. Mode 2 (without): dual heatmap matrices. PDF: send render_html=true and the report PDF is stored at d.exa.to/{uid}.pdf. It is produced AFTER the response by default, so allow a few seconds on a sync call; on an async call it already exists once the poll reports the run complete. While a render is still in flight the PDF endpoint answers success_with_remarks rather than the bytes. Send options.generate_pdf_sync=true to make a sync call wait for it instead. Fetching a stored result: pass uid + exec_mode=result, and send include_html_report=false when you only read the structured fields — the report string is ~90% of a full body and omitting it changes nothing about the stored row or the PDF. ## Query parameters - `test_mode` boolean - `lightweight` boolean ## Response `200` Success - CandidateScoreEvaluateResponse — CandidateScoreEvaluateResponse - `ApiResultType` 'Success' | 'SuccessWithRemarks' | 'InvalidInput' | 'DataUnavailable' | 'NotFound' | 'InvalidParameters' | 'UnsupportedCombination' | 'RemoteUnavailable' | 'Timeout' | 'RetryLimitReached' | 'RemoteError' | 'InProgress' | 'InsufficientCredits' | 'ConcurrentLimitReached' | 'TransactionUnavailable' | 'AccessDenied' | 'TransactionCancelled' | 'Created' | 'Updated' | 'Deleted' | 'InternalError' | 'ValidationFailed' | 'Conflict' | 'ResourceLocked' | 'PendingApproval' | 'Approved' | 'Rejected' | 'Cancelled' | 'Paused' | 'Resumed' — Novo tipo de resultado unificado para queries e operações administrativas. Substitui TransactionResultType com melhor semântica. - `BalanceInBrl` number, double — Saldo atual em BRL. - `BalanceInCredits` integer — Saldo atual em créditos. - `DataSourceCategory` string — Categoria da fonte de dados - `Date` string, date-time — Data da transação. - `ElapsedTimeInMilliseconds` integer — Tempo de execução (em milisegundos). - `HasPdf` boolean — Indica que existe PDF de comprovante do resultado. - `Message` string — Mensagem. - `OriginalFilesUrl` string — Url para download dos arquivos originais da transação (quando disponíveis). - `OutdatedResult` boolean — Indica que o resultado é datado. - `PdfUrl` string — Url para download do arquivo de comprovante em PDF da transação (quando disponível). - `ResultSubtype` string — Subtipo específico do resultado para fornecer contexto adicional (ex: 'duplicate_name', 'invalid_cpf'). - `ResultTypeCode` integer — Código numérico do resultado (compatível com TransactionResultTypeCode mas extensível). - `TotalCost` number, double — Internal use only. - `TotalCostInCredits` integer — Custo total da transação em créditos. - `TransactionResultType` 'Success' | 'SuccessWithRemarks' | 'InvalidInputData' | 'UnavailableData' | 'EntityNotFound' | 'InvalidParameters' | 'ParametersNotSupported' | 'RemoteSystemUnavailable' | 'Timeout' | 'AttemptsLimitReached' | 'RemoteSystemError' | 'AsyncExecutionInProgress' | 'AwaitingExecution' | 'BirthDateRequired' | 'InsufficientBalance' | 'SimultaneousTransactionsLimitReached' | 'TransactionUnavailable' | 'AccessDenied' | 'TransactionCancelled' | 'InternalError' —
TransactionResultType CodeTransactionResultTypeDefinitivoFaturavelErro
1SucessoVerdadeiroVerdadeiroFalso
2Sucesso parcial ou com observaçõesVerdadeiroVerdadeiroFalso
3Entrada ou documento inválidoVerdadeiroFalsoVerdadeiro
4Dados não disponíveisVerdadeiro (no dia)FalsoFalso
5Entidade ou documento inexistente/não encontradoVerdadeiroVerdadeiroFalso
6Parâmetros de entrada inválidosVerdadeiroFalsoVerdadeiro
7Combinação de parâmetros não suportadosVerdadeiroFalsoVerdadeiro
8Fonte de dados ou sistema remoto indisponívelFalsoFalsoVerdadeiro
9Limite de tempo esgotadoFalsoFalsoVerdadeiro
10Limite de tentativas atingidoFalsoFalsoVerdadeiro
11Erro na fonte de dados ou sistema remotoFalsoFalsoVerdadeiro
12Transação em execuçãoFalsoFalsoFalso
101Saldo em créditos insuficiente, entre em contatoFalsoFalsoVerdadeiro
102Limite de transações simultâneas atingido, entre em contatoFalsoFalsoVerdadeiro
103Transação não disponível, entre em contatoFalsoFalsoVerdadeiro
104Credenciais inválidas ou acesso negado, entre em contatoFalsoFalsoVerdadeiro
105Transação cancelada, entre em contatoFalsoFalsoVerdadeiro
255*Erro interno não esperado, entre em contatoVerdadeiro (no dia)FalsoVerdadeiro
* Consulta em manutenção, tente refazer mais tarde ou entre em contato com a Direct Digital para notificação do problema
- `TransactionResultTypeCode` integer — Código do tipo de resultado da transação. - `UniqueIdentifier` string — Identificador único da transação. Recomendado armazenar para controle e depuração. - `Result` CandidateScoreResult — CandidateScoreResult - `ScoringMode` string — WithJd or WithoutJd - `DataOrigin` string — LinkedIn or Resume - `TotalScore` integer — Overall score 0-100 (Mode 1) - `ExperienceScore` integer — Experience fit 0-100 - `SkillsScore` integer — Skills match 0-100 - `EducationScore` integer — Education match 0-100 - `CertificationsScore` integer — Certifications match 0-100 - `LocationScore` integer — Location match 0-100 - `LanguagesScore` integer — Languages proficiency match 0-100 - `VirtualRecruiterAssessment` string — LLM-generated qualitative assessment (Mode 1) - `Differentiators` DifferentiatorResult[] — Unstated-but-present candidate attributes (Mode 1). 0-5 entries: language/cert/education/experience the candidate exhibits that the JD did NOT ask for but may matter to the role. - `Text` string — One-sentence pt-BR description, ≤ 180 chars - `Category` string — Idioma | Formação | Certificação | Experiência | Outro - `Overqualified` OverqualifiedAssessmentResult — OverqualifiedAssessmentResult - `IsOverqualified` boolean — True when the candidate is materially over-qualified for this role (seniority gap ≥ 2 levels, tenure ≥ 1.5× JD baseline, or terminal education materially exceeds JD). - `Reason` string — pt-BR rationale (≤ 180 chars, one sentence) naming the trigger that fired. Empty when is_overqualified=false. - `RequirementsGate` RequirementsGateResult — RequirementsGateResult - `Status` string — met | not_verified | not_met | no_must_requirements. 'partial' on a must does not trip the gate (evidence exists with gaps; the score already prices it). not_verified means a must could not be proven from the documents (absence of evidence), never that the candidate lacks it. - `MustTotal` integer — Number of must-priority JD requirements evaluated. - `MustMet` integer — Musts judged met. - `MustPartial` integer — Musts judged partially met. - `MustNotFound` integer — Musts with no evidence either way in the candidate documents. - `MustContradicted` integer — Musts with explicit contrary evidence in the candidate documents. - `Unmet` UnmetRequirementResult[] — Every must with status not_found or contradicted, in JD order. - `Text` string — Requirement text as stated in the JD. - `Status` string — not_found | contradicted. - `Requirements` RequirementResult[] — Judged JD requirements checklist the gate is derived from (Mode 1). Each entry carries text, priority (must/should/could; null when the text didn't join to the JD or the run predates priority stamping) and status (met | partial | not_found | contradicted; legacy runs may carry 'miss'; pre-filter gated pairs carry 'not_evaluated'). Null when the run carries no checklist (Mode 2). - `Text` string — Requirement text as stated in the JD. - `Priority` string — must | should | could. Null when the requirement text didn't join to the JD's key requirements or the run predates priority stamping. - `Status` string — met | partial | not_found | contradicted ('miss' on legacy runs; 'not_evaluated' on pre-filter gated pairs). - `FunctionalAreaMatrix` MatrixResult — MatrixResult - `RowLabels` string[] — Row axis labels (functional areas or industry verticals) - `ColumnLabels` string[] — Column axis labels (seniority levels) - `Cells` MatrixCellResult[] — Matrix cells with scores. Cells scoring below the matrix visibility threshold are omitted. - `Row` integer - `Col` integer - `RowLabel` string - `ColLabel` string - `Score` integer - `CrossReferenceTags` string[] - `IndustryVerticalMatrix` MatrixResult — MatrixResult - `RowLabels` string[] — Row axis labels (functional areas or industry verticals) - `ColumnLabels` string[] — Column axis labels (seniority levels) - `Cells` MatrixCellResult[] — Matrix cells with scores. Cells scoring below the matrix visibility threshold are omitted. - `Row` integer - `Col` integer - `RowLabel` string - `ColLabel` string - `Score` integer - `CrossReferenceTags` string[] - `CandidateName` string — Candidate name - `TotalCareerYears` number, double — Total career years - `ExperienceCount` integer — Number of experience entries - `HtmlReport` string — Self-contained A4 HTML report. PDF input, not a web page: render it with a Chromium-based engine. A browser or wkhtmltopdf ignores the @page rules and flattens it into one continuous column. - `ElapsedTimeMs` integer — Processing time in ms - `LlmCalls` integer — Number of LLM API calls - `TotalInputTokens` integer — Total input tokens across all LLM calls - `TotalOutputTokens` integer — Total output tokens across all LLM calls - `ExtractionElapsedMs` integer — Extraction phase time in ms - `ScoringElapsedMs` integer — Scoring phase time in ms - `ExtractionProvider` string — Provider used for extraction (Haiku/Mercury/Haiku (fallback)) - `ExtractionModel` string — Model ID used for extraction - `ScoringProvider` string — Provider that actually performed scoring, with a " (fallback)" suffix when a higher-priority provider failed first. Read this and scoring_model to tell a primary answer from a degraded one: both are resolved per attempt, unlike precision_preset, which echoes what was requested. - `ScoringModel` string — Model ID that ANSWERED the scoring call, resolved per attempt — so on a fallback it names the rung that replied, not the preset's primary. It is not an echo of the request; precision_preset is. Pair it with scoring_provider, whose " (fallback)" suffix says a descent happened. - `TaxonomyMode` string — Taxonomy mode (Full/Compact) - `PrecisionPreset` string — Precision preset this result was produced under: Turbo | Fast | Balanced | Precise | Maximum | Bulk | BulkSelfHosted | BulkGlobal | BulkFast | BalancedGoogle | Express. Read scoring_model for the model that actually answered, which differs from the preset's primary when a fallback fired. - `CacheCreationTokens` integer — Total cache creation tokens (Anthropic) - `CacheReadTokens` integer — Total cache read tokens (Anthropic) - `EducationsExtracted` integer — Number of education entries extracted - `CodesNormalized` integer — Taxonomy codes normalized by post-processing - `ClassificationsGapFilled` integer — Industry verticals gap-filled by CNPJ/catalog lookup - `ExtractionJsonParseStrategy` string — JSON parse strategy for extraction (direct/code_block/brace_extraction) - `ScoringJsonParseStrategy` string — JSON parse strategy for scoring (direct/code_block/brace_extraction) - `InconsistencyFindings` InconsistencyFindingResult[] — Inconsistency findings (null when SUB 19105 didn't run) - `RuleId` integer — Stable numeric ID of the rule (matches InconsistencyRuleCode enum value) - `RuleCode` string — Symbolic rule code (e.g. A1_ExperienceVsAge) - `Category` string — Tier classification (InternalLogical/CrossSource/DossierCrossValidation/DocumentIntegrity) - `Severity` string — Severity tier (Info/Warning/Critical) - `Headline` string — One-line pt-BR summary - `Detail` string — Detailed pt-BR explanation - `EvidenceExcerpt` string — Optional pt-BR excerpt from the source - `RecommendedAction` string — Optional pt-BR recommended action for the recruiter - `Metadata` DictionaryStringString — Dictionary - `InconsistencyAggregateRag` string — Aggregate RAG indicator (Green/Amber/Red). Null when SUB 19105 didn't run. - `InconsistencyCriticalCount` integer — Number of Critical findings - `InconsistencyWarningCount` integer — Number of Warning findings - `InconsistencyInfoCount` integer — Number of Info findings - `InconsistencyDetectionElapsedMs` integer — Inconsistency detection phase time in ms - `FeedbackLetterPtBr` string — Validated pt-BR feedback letter (40-100 words). Null when SUB 19106 didn't run or failed validation. - `FeedbackLetterWordCount` integer — Word count of the rendered feedback letter (40-100 when Success=true; the prompt targets 70-100 with a strength-recognition sentence, 50-75 without one). - `FeedbackLetterSuccess` boolean — True when SUB 19106 generated AND validated the letter. False when cosmetic failure (validator/chain). Null when SUB 19106 didn't run. - `FeedbackLetterFailureReason` string — Categorized failure when Success=false (chain_exhausted / validator_word_count / validator_banned_phrase / validator_exclamation / empty_response / missing_candidate_name / missing_gap_anchors / no_client_available / exception / cancelled). Null when Success=true or SUB didn't run. - `FeedbackLetterRoleNamedInGreeting` string — Exact role title used in the greeting; null when include_role_in_greeting=false or no role available. - `FeedbackLetterStrengthAnchorUsed` string — Strength anchor referenced in the recognition sentence; null when omitted (request flag or graceful omission). - `FeedbackLetterGapAnchorsUsed` string[] — Gap anchors (1-2) referenced in the gap sentence; empty array when SUB didn't run. - `FeedbackLetterProvider` string — Provider that produced the letter (display label, includes '(fallback)' suffix when a fallback hop succeeded). Null when SUB didn't run. - `FeedbackLetterModel` string — Model ID that produced the letter (e.g. claude-haiku-4-5). Null when SUB didn't run. - `FeedbackLetterElapsedMs` integer — SUB 19106 phase time in ms. Null when SUB didn't run. - `FeedbackLetterInputTokens` integer — Letter LLM call input tokens. Null when SUB didn't run. - `FeedbackLetterOutputTokens` integer — Letter LLM call output tokens. Null when SUB didn't run. - `PreFilterEvaluated` boolean — Whether the pre-filter stage ran on this pair. - `PreFilterGated` boolean — Whether the pre-filter short-circuited LLM scoring + cosmetic SUBs. - `PreFilterEstimatedInternalScore` number, double — Calibrated cosine projection of internal_score (0.0-1.0). - `PreFilterThreshold` number, double — Threshold applied to the estimate (0.0-1.0). - `PreFilterMatcherVersion` string — Funnel document embedder tag (e.g. 'embeddinggemma-300m-onnx-int8-sts-v1'). - `PreFilterFormulaVersion` string — Funnel calibrator version (e.g. 'funnel-cosine-isotonic-gbm-stage1-2026-06-03'). - `PreFilterElapsedMs` integer — Pre-filter call time in ms. Null when the stage didn't run. - `CandidateExtractionLlm` LlmStepResult — LlmStepResult - `Model` string — Model ID sent in the API request (e.g. claude-haiku-4-5). - `ProviderLabel` string — Display label of the resolved provider (includes '(fallback)' suffix when a fallback hop succeeded). - `DisplayName` string — Friendly display name from the client. - `VendorName` string — Company that created the model (Anthropic, OpenAI, Inception, DeepInfra). - `BackendName` string — API platform the model is called through (Anthropic, Bedrock, OpenAI). - `InputTokens` integer — Input tokens for this step. - `OutputTokens` integer — Output tokens for this step. - `ReasoningTokens` integer — Reasoning tokens (GPT-5 / Gemini thinking / DeepSeek V4 think-mode). Counted toward billable output. Zero for non-reasoning calls. - `CacheCreationTokens` integer — Tokens written to prompt cache (Anthropic cache_creation_input_tokens). Zero elsewhere. - `CacheReadTokens` integer — Tokens read from prompt cache (Anthropic cache_read_input_tokens or OpenAI cached_tokens). - `ElapsedMs` integer — Wall-clock time spent in this SUB, including post-LLM normalization. - `JsonParseStrategy` string — JSON parse strategy used (direct / code_block / brace_extraction). Null for free-form text steps (SUB 19106). - `UsedStructuredOutputs` boolean — True when the call used response_format=json_schema/json_object. Null when not tracked by the SUB. - `DerivedCostUsd` number, double — Per-call cost in USD (uncached worst-case). Null when the rate card has no entry for the (vendor, backend, model) tuple. - `JdExtractionLlm` LlmStepResult — LlmStepResult - `Model` string — Model ID sent in the API request (e.g. claude-haiku-4-5). - `ProviderLabel` string — Display label of the resolved provider (includes '(fallback)' suffix when a fallback hop succeeded). - `DisplayName` string — Friendly display name from the client. - `VendorName` string — Company that created the model (Anthropic, OpenAI, Inception, DeepInfra). - `BackendName` string — API platform the model is called through (Anthropic, Bedrock, OpenAI). - `InputTokens` integer — Input tokens for this step. - `OutputTokens` integer — Output tokens for this step. - `ReasoningTokens` integer — Reasoning tokens (GPT-5 / Gemini thinking / DeepSeek V4 think-mode). Counted toward billable output. Zero for non-reasoning calls. - `CacheCreationTokens` integer — Tokens written to prompt cache (Anthropic cache_creation_input_tokens). Zero elsewhere. - `CacheReadTokens` integer — Tokens read from prompt cache (Anthropic cache_read_input_tokens or OpenAI cached_tokens). - `ElapsedMs` integer — Wall-clock time spent in this SUB, including post-LLM normalization. - `JsonParseStrategy` string — JSON parse strategy used (direct / code_block / brace_extraction). Null for free-form text steps (SUB 19106). - `UsedStructuredOutputs` boolean — True when the call used response_format=json_schema/json_object. Null when not tracked by the SUB. - `DerivedCostUsd` number, double — Per-call cost in USD (uncached worst-case). Null when the rate card has no entry for the (vendor, backend, model) tuple. - `ScoringLlm` LlmStepResult — LlmStepResult - `Model` string — Model ID sent in the API request (e.g. claude-haiku-4-5). - `ProviderLabel` string — Display label of the resolved provider (includes '(fallback)' suffix when a fallback hop succeeded). - `DisplayName` string — Friendly display name from the client. - `VendorName` string — Company that created the model (Anthropic, OpenAI, Inception, DeepInfra). - `BackendName` string — API platform the model is called through (Anthropic, Bedrock, OpenAI). - `InputTokens` integer — Input tokens for this step. - `OutputTokens` integer — Output tokens for this step. - `ReasoningTokens` integer — Reasoning tokens (GPT-5 / Gemini thinking / DeepSeek V4 think-mode). Counted toward billable output. Zero for non-reasoning calls. - `CacheCreationTokens` integer — Tokens written to prompt cache (Anthropic cache_creation_input_tokens). Zero elsewhere. - `CacheReadTokens` integer — Tokens read from prompt cache (Anthropic cache_read_input_tokens or OpenAI cached_tokens). - `ElapsedMs` integer — Wall-clock time spent in this SUB, including post-LLM normalization. - `JsonParseStrategy` string — JSON parse strategy used (direct / code_block / brace_extraction). Null for free-form text steps (SUB 19106). - `UsedStructuredOutputs` boolean — True when the call used response_format=json_schema/json_object. Null when not tracked by the SUB. - `DerivedCostUsd` number, double — Per-call cost in USD (uncached worst-case). Null when the rate card has no entry for the (vendor, backend, model) tuple. - `FeedbackLetterLlm` LlmStepResult — LlmStepResult - `Model` string — Model ID sent in the API request (e.g. claude-haiku-4-5). - `ProviderLabel` string — Display label of the resolved provider (includes '(fallback)' suffix when a fallback hop succeeded). - `DisplayName` string — Friendly display name from the client. - `VendorName` string — Company that created the model (Anthropic, OpenAI, Inception, DeepInfra). - `BackendName` string — API platform the model is called through (Anthropic, Bedrock, OpenAI). - `InputTokens` integer — Input tokens for this step. - `OutputTokens` integer — Output tokens for this step. - `ReasoningTokens` integer — Reasoning tokens (GPT-5 / Gemini thinking / DeepSeek V4 think-mode). Counted toward billable output. Zero for non-reasoning calls. - `CacheCreationTokens` integer — Tokens written to prompt cache (Anthropic cache_creation_input_tokens). Zero elsewhere. - `CacheReadTokens` integer — Tokens read from prompt cache (Anthropic cache_read_input_tokens or OpenAI cached_tokens). - `ElapsedMs` integer — Wall-clock time spent in this SUB, including post-LLM normalization. - `JsonParseStrategy` string — JSON parse strategy used (direct / code_block / brace_extraction). Null for free-form text steps (SUB 19106). - `UsedStructuredOutputs` boolean — True when the call used response_format=json_schema/json_object. Null when not tracked by the SUB. - `DerivedCostUsd` number, double — Per-call cost in USD (uncached worst-case). Null when the rate card has no entry for the (vendor, backend, model) tuple. - `TotalLlmCostUsd` number, double — Sum of derived per-stage costs in USD using the uncached-worst-case formula. Null when no stage has a rated backend. - `InterviewChecklistSuccess` boolean — Whether SUB 19107 (interview checklist split call) succeeded. Null when the SUB didn't run. - `InterviewChecklistSkippedByGate` boolean — True when the deterministic NonMatchChip gate suppressed the checklist without an LLM call (internal_score at/below the minimum-adherence threshold). - `InterviewChecklistItemCount` integer — Number of checklist items produced (0 on a gate skip). Null when the SUB didn't run. - `InterviewChecklistProvider` string — Provider that produced the checklist (display label, includes '(fallback)' suffix when a fallback hop succeeded). Null when the SUB didn't run or the gate skipped. - `InterviewChecklistModel` string — Model ID that produced the checklist. Null when the SUB didn't run or the gate skipped. - `InterviewChecklistElapsedMs` integer — SUB 19107 phase time in ms (runs parallel with the letter SUB). Null when the SUB didn't run. - `InterviewChecklistInputTokens` integer — Checklist LLM call input tokens. Null when the SUB didn't run or the gate skipped. - `InterviewChecklistOutputTokens` integer — Checklist LLM call output tokens. Null when the SUB didn't run or the gate skipped. - `SalaryScore` integer — Salary compatibility match 0-100. The seventh factor of the rubric; published from 2026-08-17, having been computed all along and dropped at the DTO boundary while its six siblings were carried. - `Factors` CandidateScoreFactor[] — The seven rubric factors (Mode 1). ALWAYS seven entries in canonical order, including factors this JD does not weigh — `score` is null rather than 0 on those, because a zero would claim the candidate scored badly on something nobody asked about. Each row carries name, key, score, reasoning, weight, role and counts. Null in Mode 2 or when the pre-filter gated the pair. - `Name` string — Canonical factor name as the scoring math keys it (Experiência, Competências, Formação, Idiomas, Localização, Certificações, Compatibilidade Salarial). - `Key` string — Stable ASCII key for matching without an accented string. Closed set: experiencia | competencias | formacao | idiomas | localizacao | certificacoes | salarial. - `Score` integer — 0-100. NULL when this factor carries no score, because the JD did not ask for it or the model omitted it. Never 0 as a stand-in: a zero is a statement about the candidate, an absent score is a statement about the vacancy. - `Reasoning` string — The model's written justification for this factor. Null on a row it did not emit — there is no rationale for a judgment never made. - `Weight` number, double — Weight this factor carries in the aggregate (0.30 experiência, 0.25 competências, 0.15 formação, 0.10 idiomas, 0.10 localização, 0.05 certificações, 0.05 salarial). Published so a caller recomputing the total does not keep its own copy. - `Role` string — required | differential | not_requested. Tri-state on purpose: a differential is a bonus axis where meeting it helps and missing it is never a penalty, so rendering it the same as not_requested tells the reader the vacancy does not care about something it treats as a plus. Null on runs predating formula 1.5.0. - `DifferentialMet` boolean — Set only when role=differential: whether the candidate demonstrably meets it. False or null withholds the bonus and never applies a penalty. - `Counts` boolean — Whether this factor entered the renormalized aggregate — the deterministic gate behind `role`. - `InterviewChecklist` CandidateScoreChecklistItem[] — The interview checklist questions (Mode 1). Null when SUB 19107 did not run or the NonMatchChip gate suppressed it — distinguishable from an empty list on purpose, so a consumer draws nothing rather than an empty control. `purpose` is RECRUITER-facing: do not render it anywhere a candidate may read it. - `Question` string — The interview question, pt-BR. - `Purpose` string — verify_strength | probe_gap | clarify_ambiguity. RECRUITER-FACING: do not render it anywhere a candidate may read it, on screen or in print. A candidate who sees 'probe_gap' beside a question feels pre-judged before the interview starts. --- [API](https://skmtc.dev/directdigital/apis/backend-apis.md) · [All operations](https://skmtc.dev/directdigital/apis/backend-apis/llms.txt) · [OpenAPI document](https://skmtc-service-production.skmtc.workers.dev/v1/apis/directdigital/backend-apis/revisions/a3305895eccd/schema)