> ## Documentation Index > Fetch the complete documentation index at: https://docs.moonbounce.ai/llms.txt > Use this file to discover all available pages before exploring further. # Export a decision's feedback as training data > Lists feedback with its claims and evaluated cases; reads only. The raw feedback on a decision, for a producer's training data (calibration, weights, suggested versions): one item per submission, oldest first, with its claims, the evaluation it judged (version hashes, engine profile, cut points, score, outcome, each leaf's answer and distribution, and the content when the decision stores it) and its group's split. Holdout feedback is never listed, not its claims, its evaluation or its content; it is only counted in `withheld.holdout`, together with feedback whose content's text belongs to a held-out group, so suggestions can still be tested on feedback their producer never saw. `withheld` counts submission rows, including withdrawal events, not distinct evaluations. It is counted on the first page only (no `cursor`) and is null on later pages. Counts per `structure` or `weights` filter can tell you that a version gained held-out feedback, never which item. It is raw, not a consensus: superseded submissions are listed and marked `superseded_by`, and how submissions become training data is the reader's call. `withdrawn_at` marks a claimless withdrawal event: resolve chain heads across pages, exclude withdrawn chains without reviving ancestors, and invalidate cached answers. On a spot check, `after_reveal` is false on the reviewer's first, blind answer and true on each answer that changed it after Moonbounce's call was revealed: pick the blind answer to measure accuracy, the latest to train. `channel` says where each submission came from; Slack's terms bar training an LLM on `slack` feedback. Scoped to the key's organization and environment, and feedback on deleted evaluations is left out. A decision with no feedback this key can see answers 200 with no items. `structure` and `weights` keep only evaluations on those hashes. `split.group` is an opaque id shared by submissions on near-identical items in the key's environment; it never names an item. Page forward by passing `next_cursor` as `cursor` while `has_more` is true. `next_cursor` is also set on the last page, so keep it: reading from it later returns what was written after it, apart from the following. Items are in creation order, not commit order, and a submission's creation time is the start of the transaction that wrote it: any write that commits after a later one was read (a seed batch, or any concurrent submission) can land behind a cursor already handed out, and a holdout group that is burned later appears under its submissions' original creation time. A reason added later to a submission already read (a reviewer's `POST /v1/review/items/{item_id}/reason`) does not move it either. So a forward reader re-reads from the start now and then. Not an MCP tool, so bulk content stays out of coding agents' context. The web session may not read it (403). ## OpenAPI ````yaml /openapi.json get /v1/decisions/{decision_id}/feedback openapi: 3.1.0 info: title: Moonbounce Decisions API version: 0.0.0 description: >- Moonbounce decisions: evaluate items against a decision spec (features combined by aggregators), record feedback and seed examples, review a ranked queue, and improve the spec through suggestions that ship as code diffs, each with a preview (backtest) and an evidence label. **Authentication.** Send an API key as `Authorization: Bearer ` on every operation except `GET /health` and `POST /v1/specs/hash`. Call the API from a server, never a browser: a key must not reach one, and the API sends no CORS headers. A key belongs to one organization and one environment: production, or test for a key with the `env:test` permission. Test traffic never counts as production traffic. **Requests.** JSON bodies (`Content-Type: application/json`) of at most 1 MB, parsed strictly: duplicate keys, lone surrogates, integers outside ±(2^53 - 1), `__proto__` keys, U+0000 and nesting deeper than 64 are a 400. **Retries.** Send an `Idempotency-Key` on every POST under /v1 that takes one (all but `POST /v1/specs/hash`). Keep it for the request's retries. The same key with the same request returns the first stored result, marked `Idempotent-Replayed: true`, for 24 hours (410 if that result was since deleted); with another request it is 422 IDEMPOTENCY_KEY_REUSED. Keys are scoped to the organization and environment across every operation, so use a fresh key per request. A request that failed before its result was stored runs again when retried. 429 and 503 under /v1 may carry `Retry-After` (seconds); a daily QUOTA_EXCEEDED cap instead names its reset time. **Errors.** Every error under /v1 is `{"error": {"type", "message", "reason"?, "param"?, "details"?}}` (ErrorBody); `GET /health` answers its own shape. Every response carries a `Request-Id` header. Rate limits apply per key and per client IP, counted per Cloudflare location. servers: - url: https://api.moonbounce.ai description: Production security: [] paths: /v1/decisions/{decision_id}/feedback: get: tags: - feedback summary: Export a decision's feedback as training data description: >- Lists feedback with its claims and evaluated cases; reads only. The raw feedback on a decision, for a producer's training data (calibration, weights, suggested versions): one item per submission, oldest first, with its claims, the evaluation it judged (version hashes, engine profile, cut points, score, outcome, each leaf's answer and distribution, and the content when the decision stores it) and its group's split. Holdout feedback is never listed, not its claims, its evaluation or its content; it is only counted in `withheld.holdout`, together with feedback whose content's text belongs to a held-out group, so suggestions can still be tested on feedback their producer never saw. `withheld` counts submission rows, including withdrawal events, not distinct evaluations. It is counted on the first page only (no `cursor`) and is null on later pages. Counts per `structure` or `weights` filter can tell you that a version gained held-out feedback, never which item. It is raw, not a consensus: superseded submissions are listed and marked `superseded_by`, and how submissions become training data is the reader's call. `withdrawn_at` marks a claimless withdrawal event: resolve chain heads across pages, exclude withdrawn chains without reviving ancestors, and invalidate cached answers. On a spot check, `after_reveal` is false on the reviewer's first, blind answer and true on each answer that changed it after Moonbounce's call was revealed: pick the blind answer to measure accuracy, the latest to train. `channel` says where each submission came from; Slack's terms bar training an LLM on `slack` feedback. Scoped to the key's organization and environment, and feedback on deleted evaluations is left out. A decision with no feedback this key can see answers 200 with no items. `structure` and `weights` keep only evaluations on those hashes. `split.group` is an opaque id shared by submissions on near-identical items in the key's environment; it never names an item. Page forward by passing `next_cursor` as `cursor` while `has_more` is true. `next_cursor` is also set on the last page, so keep it: reading from it later returns what was written after it, apart from the following. Items are in creation order, not commit order, and a submission's creation time is the start of the transaction that wrote it: any write that commits after a later one was read (a seed batch, or any concurrent submission) can land behind a cursor already handed out, and a holdout group that is burned later appears under its submissions' original creation time. A reason added later to a submission already read (a reviewer's `POST /v1/review/items/{item_id}/reason`) does not move it either. So a forward reader re-reads from the start now and then. Not an MCP tool, so bulk content stays out of coding agents' context. The web session may not read it (403). operationId: listFeedback parameters: - schema: $ref: '#/components/schemas/DecisionId' required: true name: decision_id in: path - schema: type: string pattern: ^sha256:[0-9a-f]{64}$ description: Only feedback on evaluations of this structure hash. required: false description: Only feedback on evaluations of this structure hash. name: structure in: query - schema: type: string pattern: ^sha256:[0-9a-f]{64}$ description: Only feedback on evaluations with this weights hash. required: false description: Only feedback on evaluations with this weights hash. name: weights in: query - schema: type: integer minimum: 1 maximum: 100 default: 50 required: false name: limit in: query - schema: type: string minLength: 1 maxLength: 1024 description: The previous page's `next_cursor`. required: false description: The previous page's `next_cursor`. name: cursor in: query responses: '200': description: A page of feedback, oldest first content: application/json: schema: $ref: '#/components/schemas/FeedbackExport' '400': description: 'Invalid request: `param` names the field at fault when one applies' content: application/json: schema: $ref: '#/components/schemas/ErrorBody' '401': description: Missing or invalid API key (reason KEY_INVALID) content: application/json: schema: $ref: '#/components/schemas/ErrorBody' '403': description: The key may not do this content: application/json: schema: $ref: '#/components/schemas/ErrorBody' '429': description: >- RATE_LIMITED: retry after Retry-After seconds; QUOTA_EXCEEDED: details.reset_at gives the daily reset time (no Retry-After), or details.too_large asks for a smaller request headers: Retry-After: schema: type: string description: Seconds to wait before retrying, when waiting helps. required: false description: Seconds to wait before retrying, when waiting helps. content: application/json: schema: $ref: '#/components/schemas/ErrorBody' '503': description: >- Temporarily unavailable, with Retry-After (for example reason AUTH_UNAVAILABLE, ENGINE_UNAVAILABLE or STORE_UNAVAILABLE): retry with the same Idempotency-Key headers: Retry-After: schema: type: string description: Seconds to wait before retrying, when waiting helps. required: true description: Seconds to wait before retrying, when waiting helps. content: application/json: schema: $ref: '#/components/schemas/ErrorBody' security: - bearerAuth: - decisions:develop components: schemas: DecisionId: type: string pattern: ^[a-z][a-z0-9_.-]{0,127}$ FeedbackExport: type: object properties: environment: $ref: '#/components/schemas/Environment' items: type: array items: $ref: '#/components/schemas/FeedbackExportItem' next_cursor: type: - string - 'null' description: >- Where to resume: pass it as `cursor`. Set whenever the page has items (the last item's position) and, on an empty page read with a cursor, that cursor again, so a reader at the end keeps its place. Reading from it later returns what was written after it, apart from writes that commit behind it and reasons added later to submissions already read (see the endpoint), so re-read from the start now and then. Null only on an empty first page. has_more: type: boolean description: True when more items follow this page now. withheld: anyOf: - type: object properties: holdout: type: integer minimum: 0 description: >- Submissions in held-out groups, and submissions whose content's text belongs to a held-out group: counted, never listed. unassigned: type: integer minimum: 0 description: Submissions whose item has no split yet. required: - holdout - unassigned additionalProperties: false - type: 'null' description: >- Submission rows, including withdrawal events (not distinct evaluations), in scope (this decision, the key's environment, the same version filters) that the export leaves out of every page. Counted on the first page only (no `cursor`); null on later pages. Counts per filter can tell you that a version gained held-out feedback, never which item. required: - environment - items - next_cursor - has_more - withheld additionalProperties: false ErrorBody: type: object properties: error: type: object properties: type: $ref: '#/components/schemas/ErrorType' reason: type: string pattern: ^[A-Z][A-Z0-9_]+$ description: >- Set where a caller can act on it, one of the reasons in the ErrorReason schema. New reasons may be added: treat an unknown one by the status and `type`. message: type: string description: For people; don't parse it. param: type: string description: >- The request field at fault, as a dotted path (`subject.0.text`, `claims.1.value`). details: type: object additionalProperties: {} required: - type - message additionalProperties: false required: - error additionalProperties: false Environment: type: string enum: - production - test description: >- The environment of the key that made the call: `production` or `test`. Evaluations, feedback, reviews and suggestions are separate per environment; registered structures are shared. FeedbackExportItem: type: object properties: id: $ref: '#/components/schemas/FbkId' source: $ref: '#/components/schemas/FeedbackSource' channel: type: string enum: - slack - web - api - mcp - import description: >- The surface the submission came through, stamped by the server. Slack's terms bar training an LLM on Slack data: leave out `slack` when you fit a model. actor: type: string description: >- Who submitted it, as recorded (an `ext:` prefix marks an actor a key named). reason: type: string reason_pending: type: boolean description: >- True when a reason was promised and has not arrived yet. On a superseded submission (`superseded_by` set) it means the reason was never added before the answer was replaced, and none will be: read the chain head. reason_added_at: type: string format: date-time description: >- When a pending reason was filled in. A reason added after the answer was revealed is commentary, not blind evidence. unsure: type: boolean withdrawn_at: type: string format: date-time description: >- When this claimless submission withdrew its chain. Resolve chain heads before excluding withdrawn chains; ancestors never revive. after_reveal: type: boolean description: >- Spot checks only. `false` on the reviewer's first answer, given blind; `true` on an answer that changed it after Moonbounce's call was revealed. Accuracy is measured on the blind answer. Omitted on other items, other sources and withdrawal events. supersedes: type: string pattern: >- ^fbk_[0-9a-f]{8}-[0-9a-f]{4}-7[0-9a-f]{3}-[89ab][0-9a-f]{3}-[0-9a-f]{12}$ description: The submission this one replaces. superseded_by: type: string pattern: >- ^fbk_[0-9a-f]{8}-[0-9a-f]{4}-7[0-9a-f]{3}-[89ab][0-9a-f]{3}-[0-9a-f]{12}$ description: >- Set when a later submission replaced this one. This one no longer counts. created_at: $ref: '#/components/schemas/Timestamp' claims: type: array items: $ref: '#/components/schemas/ExportedClaim' description: The claims, in the order they were sent. evaluation: $ref: '#/components/schemas/ExportedEvaluation' split: type: object properties: split: $ref: '#/components/schemas/ExportedSplit' group: type: string pattern: ^grp_[0-9a-f]{32}$ description: >- An opaque, stable id of the item's group of near-identical items in the key's environment: submissions with the same `group` share one. It never names an item. required: - split - group additionalProperties: false description: >- The item's group and its side of the training line: `seed` (a starting example), `train`, or `burned` (a holdout group whose cases were shown in a preview, that was merged with a near-duplicate group, or whose content was later imported as a seed example; it counts as training). required: - id - source - channel - actor - reason_pending - unsure - created_at - claims - evaluation - split additionalProperties: false ErrorType: type: string enum: - invalid_request - authentication - permission_denied - not_found - conflict - rate_limited - engine_failed - unavailable - internal FbkId: type: string pattern: >- ^fbk_[0-9a-f]{8}-[0-9a-f]{4}-7[0-9a-f]{3}-[89ab][0-9a-f]{3}-[0-9a-f]{12}$ FeedbackSource: type: string enum: - app_native - app_resolution - reviewer - owner - import - synthetic_review Timestamp: type: string format: date-time ExportedClaim: type: object properties: claim: $ref: '#/components/schemas/Claim' node_hash: type: string pattern: ^sha256:[0-9a-f]{64}$ description: >- Feature claims only: the node hash of the prompt that was judged. A feature whose prompt has since changed has a different hash, and the claim does not apply to it. shown: type: object properties: structure: $ref: '#/components/schemas/VersionHash' weights: $ref: '#/components/schemas/VersionHash' required: - structure - weights additionalProperties: false description: The structure and weights whose answers the person saw. required: - claim - shown additionalProperties: false ExportedEvaluation: type: object properties: id: $ref: '#/components/schemas/EvlId' item_id: type: string assignment_mode: $ref: '#/components/schemas/AssignmentMode' created_at: $ref: '#/components/schemas/Timestamp' version: type: object properties: structure: $ref: '#/components/schemas/VersionHash' weights: $ref: '#/components/schemas/VersionHash' weights_source: type: string enum: - pinned - default binding: type: integer minimum: 1 engine_profile: type: string required: - structure - weights - weights_source - binding - engine_profile additionalProperties: false outcomes: type: object properties: below: type: string pattern: ^[a-z][a-z0-9_]{0,63}$ description: The outcome for a score under the first cut. cuts: type: array items: type: object properties: at: type: number outcome: $ref: '#/components/schemas/Name' required: - at - outcome additionalProperties: false minItems: 1 maxItems: 9 abstain: $ref: '#/components/schemas/Name' required: - below - cuts additionalProperties: false description: The cut points as the caller sent them. score: type: number outcome: $ref: '#/components/schemas/Name' features: type: object additionalProperties: $ref: '#/components/schemas/ExportedFeature' description: Each leaf's answer, by feature name. content: anyOf: - $ref: '#/components/schemas/ExportedContent' - type: 'null' description: The content judged; null when the decision does not store content. required: - id - item_id - assignment_mode - created_at - version - features - content additionalProperties: false ExportedSplit: type: string enum: - seed - train - burned Claim: oneOf: - type: object properties: kind: type: string enum: - target value: type: boolean description: Whether the spec's target statement is true of this item. required: - kind - value additionalProperties: false description: >- Whether the target statement is true. Only for specs that declare a target. - type: object properties: kind: type: string enum: - outcome value: type: string minLength: 1 maxLength: 64 description: >- The right outcome: one of this evaluation's outcomes (`below` or a cut's outcome); an `abstain` name counts only where it reuses one of them. required: - kind - value additionalProperties: false description: >- The outcome this item should have had. On a cut score, the lowest outcome also says the target is false and the highest says it is true; middle outcomes say nothing about the target. - type: object properties: kind: type: string enum: - verdict value: type: string enum: - approve - reject description: 'approve: the shown outcome was right. reject: it was wrong.' required: - kind - value additionalProperties: false description: >- Approve or reject the outcome that was shown. Needs `about.evaluation_id`, and an evaluation whose outcome was shown to someone: primary traffic that decided (it has a score). An evaluation that abstained has only a fallback outcome, whatever it is called. - type: object properties: kind: type: string enum: - feature feature: type: string minLength: 1 maxLength: 64 description: A boolean feature of the evaluation's spec. value: type: boolean description: What the feature's answer should have been. required: - kind - feature - value additionalProperties: false description: >- The expected answer of one boolean feature. It is stored against that feature's prompt as it was judged, so a reworded prompt does not inherit it. - type: object properties: kind: type: string enum: - input_broken problem: type: string enum: - truncated - wrong_language - bad_extraction - unreadable - other required: - kind - problem additionalProperties: false description: >- The content could not be judged. The evaluation is kept out of fits and previews. - type: object properties: kind: type: string enum: - action value: type: string minLength: 1 maxLength: 128 description: >- What your app did, in your own words (for example `auto_approved`). logic_version: type: string maxLength: 128 description: Which version of your own logic acted. required: - kind - value additionalProperties: false description: >- What your app did with the outcome. Observed behavior, not a statement of what is true. VersionHash: type: string pattern: ^sha256:[0-9a-f]{64}$ EvlId: type: string pattern: >- ^evl_[0-9a-f]{8}-[0-9a-f]{4}-7[0-9a-f]{3}-[89ab][0-9a-f]{3}-[0-9a-f]{12}$ AssignmentMode: type: string enum: - primary - shadow - backtest - seed Name: type: string pattern: ^[a-z][a-z0-9_]{0,63}$ ExportedFeature: type: object properties: node_hash: type: string pattern: ^sha256:[0-9a-f]{64}$ description: The hash of the prompt as it was asked. status: type: string enum: - answered - abstained - failed value: type: number minimum: 0 maximum: 1 description: The answer's probability of true, when answered. distribution: type: object additionalProperties: type: number description: >- The raw per-answer distribution as the engine returned it, never re-rounded. engine: type: string model: type: string description: The model the binding asked for. resolved_model: type: string description: The versioned model id that answered. adapter: type: string description: The adapter package and version. cached: type: boolean description: True when the answer came from the feature cache. required: - node_hash - status - cached additionalProperties: false ExportedContent: type: object properties: subject: type: array items: $ref: '#/components/schemas/ContentPart' context: type: array items: $ref: '#/components/schemas/ContentPart' required: - subject additionalProperties: false ContentPart: oneOf: - $ref: '#/components/schemas/TextPart' - $ref: '#/components/schemas/JsonPart' discriminator: propertyName: type mapping: text: '#/components/schemas/TextPart' json: '#/components/schemas/JsonPart' TextPart: type: object properties: type: type: string enum: - text text: type: string minLength: 1 name: type: string pattern: ^[a-z][a-z0-9_]{0,63}$ description: >- Names the part, so a feature's `inputs` can pick it. In `subject`, avoid "context" when the request also sends `context`. sensitive: type: boolean description: >- Display policy only: reviewers see the part blurred or behind a link. The engine still reads it and it is stored like any other part. required: - type - text additionalProperties: false JsonPart: type: object properties: type: type: string enum: - json data: anyOf: - type: object additionalProperties: {} - type: array items: {} name: type: string pattern: ^[a-z][a-z0-9_]{0,63}$ description: >- Names the part, so a feature's `inputs` can pick it. In `subject`, avoid "context" when the request also sends `context`. sensitive: type: boolean description: >- Display policy only: reviewers see the part blurred or behind a link. The engine still reads it and it is stored like any other part. required: - type - data additionalProperties: false securitySchemes: bearerAuth: type: http scheme: bearer description: >- An API key, sent as `Authorization: Bearer `. Each operation lists the scope its key needs: `decisions:evaluate` (evaluate and read evaluations), `decisions:feedback` (your app's feedback) or `decisions:develop` (everything else, including owner feedback and the review queue). An operation listing no scope takes any scoped key. A key with no scope, or without the operation's, gets 403 INSUFFICIENT_SCOPE. ```` This documentation is built and hosted on [Mintlify](https://mintlify.com), a developer documentation platform.