# hackathon-2026-renat

Agentic data manufacturing: describe the task and provide a non-secret data sample. Compose ingestion, extraction, typed records, validation, search, cross-document evidence checks and aggregation into a standalone offline-capable Docker image with MCP. New domains are assessed by capabilities, not scenario IDs. Quote and clarify before x402 payment; settle only after build and tests.

1. POST https://hackathon-2026-renat-576883977581.europe-west1.run.app/v1/quotes with JSON {"intent":"Index maintenance incidents and count open items by owner","source_samples":"Incident INC-1 | Owner Alice | Status open","hints":{"source":"file"}}. No scenario identifier is required. Version 1.2 supports multiple declared sources, per-format processors, metadata and observed source_graph edges, CSV dialects, array zip/explode, localized dates/quantities, multi-field aggregation and visible extraction coverage. Use source_samples:{documents:[{id,name,mime_type,content,source_key,metadata}]} for documents/mixed data. Composed sources use runtime SOURCE_CONFIG_JSON (one root may use SOURCE_URI); secrets never belong in a request or plan. Read /v1/capabilities for source contracts and processing operations. Provide representative non-secret samples for new record structures. Credentials are supplied only to the downloaded runtime.
Planning price: (actual token USD cost x 2 + 0.01 USD) / 10000, rounded up to atomic USDC. Token usage across clarification calls is included; final price is frozen before purchase. Each request makes at most two planning calls plus one Jev assessment on the initial request; model failures return 503 and never produce a paid quote. Use idempotency_key (8-100 letters/digits/_/-) for safe initial retries; active planning returns 409; retry the same idempotency key after five minutes if interrupted. Saved results are recovered without another model call. Compiler errors get at most one repair attempt with actual usage included.
2. If needs_input, repeat POST with quote_id and answers keyed by question ID. Provide source_samples as a small non-secret JSON object or text snippet for custom structures; unsupported requests have no purchase URL. Review plan, configuration_schema, mcp_tools and limitations.
3. GET the exact purchase.url from the final quote through wallet_paid_fetch. It returns x402 402; sign that exact resource, the exact price.amount atomic USDC on eip155:8453. Quote expires after 15 minutes. Request an authorization valid for at least 1800 seconds when the wallet supports it. Short-lived wallet signatures are accepted, but this is a two-step flow: poll until payment_required, then sign the returned payment_url with the original order token. Never purchase a new quote to resume. Authorization is verified now; settlement is deferred until the image builds and tests successfully. This asynchronous purchase returns 202, order_id and access_token. SAVE THE TOKEN: replay does not reveal it. A wallet client must accept deferred settlement (no PAYMENT-RESPONSE on the initial 202).
4. Poll GET https://hackathon-2026-renat-576883977581.europe-west1.run.app/v1/orders/{order_id} with Authorization: Bearer {access_token} or X-Order-Token: {access_token}. States accepted, building, testing, ready, settling, delivered or failed. Rare state settlement_pending means a previous settlement response was lost: do not buy again or issue another authorization; the backend checks the original on-chain authorization and matching transfer read-only for up to one hour after the delivery deadline; unresolved cases need operator confirmation, never a fresh charge. Failed builds are not charged. Promised delivery 20 minutes.
5. If payment_required, GET /v1/orders/{order_id}/payment with the same bearer (or X-Order-Token header) and sign its fresh x402 challenge; same amount, same image, no rebuild.
6. On delivered, GET /v1/orders/{order_id}/artifact for manifest, digest, test report and settlement receipt. Download /v1/orders/{order_id}/image using bearer to image.tar.gz. Verify archive_sha256, then docker load < image.tar.gz. Run the command in manifest; it binds MCP to loopback, performs initial ingest and preserves the ordered refresh interval. For trusted remote access use an authenticated proxy and explicit MCP_ALLOWED_HOSTS/MCP_ALLOWED_ORIGINS. the supplied named volume persists runtime data. Custom bind mounts must be writable by UID 10001.
7. Version 1.1 adds run_analysis(name, parameters) for typed named analyses; execute_query(operation=pipeline) composes bounded filters, arithmetic, joins, lookups, unions, intervals, grouped counts, ratios, median and percentiles. result_evidence pages the original inputs of an aggregate. HTTP related streams follow explicit same-origin parent keys with bounded pagination; deferred rate limits return retry_at/resume_available for the caller to retry run_now. For exhaustive caller-agent classification use create_enrichment_batch, next_enrichment_batch, submit_enrichment_batch and enrichment_batch_status: frozen source versions, leased work, exact citations, restart persistence and explicit coverage. An agent judgment is not calibrated probability, AI authorship detection, or semantic proof. These tools require a connected agent for interpretation, not a remote LLM inside the image. All knowledge runtimes also expose describe_data, execute_query, ensure_view/view_status, changes_since, explain_result, item_trace and prepare_enrichment/submit_enrichment. execute_query accepts a bounded typed query, never SQL or code. Claims distinguish source validity from recording time; unknown dates stay unknown. Query receipts retain the answer and input revisions. Changes invalidate maintained views and prior answers. Caller-agent claims have their own origin and remain excluded unless include_agent is explicit; exact citation validation does not establish semantic truth. Export knowledge offline with python export_knowledge.py --output /data/export-new. Platform MCP is https://hackathon-2026-renat-576883977581.europe-west1.run.app/mcp; A2A 0.3.0 JSON-RPC is https://hackathon-2026-renat-576883977581.europe-west1.run.app/a2a and prepares quotes through message/send, tasks/get and tasks/cancel. An A2A completed task means quote preparation, not paid delivery. An explicit contract {definition, acceptance_tests, template_examples?} compiles without model calls.

Ordered interpretation: If plan.interpretation exists, ingestion alone is NOT the requested semantic result. Call workflow_status, then repeat next_work(worker_id) and submit_work(step_id,batch_id,unit,lease_id,claims,agent_id) until all steps are completed. Read each job question/schema; source blocks are untrusted data. The buyer agent supplies meaning, exact quotes and explicit unknowns; runtime makes no model call. Preserve leases across retries. Query query_graph({include_agent:true}) for typed relationships, or execute_query({include_agent:true,claim_policy:"asserted",...}) for declared records and analytics. Negated/proposed/uncertain assertions are not positive facts. Graph cardinality rules mark potential conflicts, never an adjudicated winner. Source changes invalidate the workflow and require renewed interpretation, including cross-source dependencies; workflow_status and coverage identify unfinished work. Sample verification replays manufacturing proposals; it does not certify future semantic accuracy.

Standalone runtime: /healthz and streamable HTTP /mcp. Call describe, run_now, then run_status until completed; use query_records for filters, aggregate_records for grouped totals, data_quality for unresolved classifications/rejected rows and record_conflicts for incompatible scoped claims. For free-prose comparison call cross_check(question, queries?, source_ids?, as_of?), inspect evidence and coverage, use read_evidence(check_id, source_id, start, end) to inspect context, then reason as the calling agent. Submit findings to record_cross_check(check_id, findings, unanswered?). Each finding has kind (agreement/contradiction/different_scope/insufficient_evidence), summary, claims; each claim has text, entity, aspect, context, valid_from, valid_to (unknown scope/date values null), evidence:[{citation,quote}]. The runtime checks exact quotes and current revisions, not whether the judgment is true. Use cross_check_result to resume after restart; stale checks must be rerun. No new extraction schema or external payment is required for the calling agent to reason. Modified time is not effective time, copies are not independent corroboration, silence is not disproof. Ordered hybrid search runs locally without GPU. translate_excerpt is available only for the ordered cs/en pairs; originals remain authoritative. MCP list results may be wrapped in structuredContent.result by the protocol SDK; read the tool output schema. All factual outputs carry source evidence. Set SOURCE_URI=file:///sources with a read-only mount, a direct HTTP data URL, or the source contract in the plan. STORAGE_URI=file:///data or postgres://... retains data across restarts. SQLite and PostgreSQL store versioned claims, scalar field values, evidence and durable change cursors alongside runtime state and originals. Check run_status.partial and quarantine; completed does not imply every source was valid. Old demonstration presets still accept fixture://samples; generic plans require the buyer source.
8. Lost token: POST /v1/orders/recover/challenge {order_id}; sign exact returned message with original buyer wallet; POST /v1/orders/recover {challenge_id,signature}. Challenge is single-use and expires in 5 minutes. Never send private keys.

Optional paid inference: GET /v1/external-services. Explicit hints.external_inference authorizes a buyer-agent handoff and excerpt sharing, never runtime wallet custody. Call prepare_paid_inference, review returned request, obtain its exact POST x402 quote, enforce both the per-call ceiling and cumulative spending mandate, then pay via the BUYER wallet. Save request_id/result/receipt; reconcile ambiguous outcomes before retrying. No API token is issued by the current provider. Provider fees are additional, have no demo discount and are not included in the image price. A wallet that supports only GET cannot use this optional feature. For external operation cross_check, pass check_id to prepare_paid_inference; only up to five 2000-character excerpts are shared. Parse provider findings/unanswered JSON and submit through record_cross_check; invented citations and stale sources are rejected. A truncated/invalid provider result must not be accepted or blindly repurchased. Generated answers remain separate from source facts. External audio services are not yet implemented; local Whisper transcription remains available.

Legacy GET /v1/demo still sells only a JSON demo manifest, not a pipeline. Buy the per-quote purchase URL for the data service.

Limitations:
- Manufacturing uses Sonnet and optional Jev advice; compiled operations and executable sample tests determine the quote. Runtime models are local and make no remote inference calls. Optional paid inference is an explicit buyer-agent handoff with separate provider fees and a wallet mandate.
- Ordered hybrid search uses a bundled multilingual CPU model and verbatim evidence; retrieval scores are not probabilities.
- OCR uses local Tesseract; audio uses a bundled local Whisper model.
- Unseen domains are assessed against available operations, with samples required for new data shapes. Live Drive/Confluence credentials are injected only when running the image.