oddspark
← back to the button

How does this work?

One button starts a bounded Evidence-to-Render pipeline. It can return a qualified Opportunity Brief or an approved house Brief, while keeping rejected work and private material out of the page.

A published seed is transparent input to generation. Model output is nondeterministic, so the seed is not a promise that another run will produce identical words. The current receipt references the approved Evidence and binds the committed artifact; it does not claim that the public can recreate or independently prove the model run.

1 · Evidence to Render

Each stage has one job. Only a Candidate that passes deterministic local checks and an independent Judge can reach the authoritative commit. A rejected Candidate is never rendered.

  1. Evidence. Assemble an immutable, allowlisted grounding bundle.
  2. Generate. Ask the active generation role for one Candidate.
  3. Local Gate. Check schema, grounding, privacy, names, and number provenance without another model call.
  4. Judge. Independently assess all nine gates, tone, and claims.
  5. Commit. Ask COORD to authoritatively accept one qualified Brief.
  6. Render. Show only the committed Brief and its permitted receipt details.

2 · Evidence and privacy

Evidence can use approved local priors and public feed inputs. A submitted website may be read only when that capability is active. Private sources, personal information, raw pages, and rejected Candidates are excluded before persistence or model use. The published seed is separate input, not Evidence.

  1. Use allowlisted inputs. Build Evidence only from approved local priors and permitted public observations.
  2. Keep the boundary. Exclude private sources, personal information, raw pages, and rejected Candidates before persistence or model use.
  3. Separate the seed. Pass the published seed to Generate as transparent input, not as Evidence or a reproduction guarantee.
  4. Count Evidence calls. Any metered Evidence request consumes the same shared call ledger used by generation and judging.

3 · Bounded attempts and house fallback

Evidence, generation, and judging share one cap of six model calls. The orchestrator starts only a complete Generate-and-Judge pair that fits in the remaining budget and deadline. There can be at most three such pairs when Evidence used no calls, and fewer when it did.

  1. Share six calls. Evidence, Generate, and Judge debit one request-level ledger capped at six model calls.
  2. Start complete pairs only. A Candidate attempt begins only when both its Generate and Judge calls fit inside the budget and deadline.
  3. Use a qualified result. A Candidate proceeds only after both the Local Gate and Judge pass it.
  4. Fall back safely. On failure or exhaustion, select an approved house Brief only when its catalog authority is valid and COORD can safely commit it. A house Brief is curated content, not another model.

4 · Authority and receipt honesty

COORD is the authority for claiming work, reading current state, committing the Brief, and counting served outcomes. KV is a projection for compatible reads, not commit authority. If COORD is uncertain, oddspark does not render a new Brief.

  1. Trust COORD. COORD alone claims, reads, commits, and counts served outcomes.
  2. Treat KV as projection-only. KV may support compatible reads but cannot authorize a commit.
  3. Read the receipt narrowly. It references the approved Evidence and binds the committed artifact; receipt-proof claims remain off until an active ReceiptClaimManifest authorizes exact wording.
  4. Expect nondeterminism. The same published seed can lead a model to different valid wording on another run.