HARMONIA + TRUTH TASK FORCE + DECISION LEDGER
One experience. Separate responsibilities. Reconstructable proof.
The platform is intentionally not one giant black box. Each layer does one job, exposes its inputs and outputs, and leaves an inspectable handoff for the next layer.
The trust stack
1 · User intentNatural-language question, claim, file, media or challenge.
2 · Harmonia PreflightClarify scope, estimate cost, predict research depth and information value, produce quote and authorization ceiling.
3 · Harmonia OrchestrationRoute models/tools, enforce budget, handle fallback/degraded success, collect actual usage and cost.
4 · Evidence ContinuityHash sources, trace roots, preserve timestamps, transformations, custody, limitations and conflicts.
5 · Truth Task Force AssessmentSeparate support from contradiction and publish Evidence Support % ± Uncertainty.
6 · Challenge LabIndependent challenge cases can add, downweight, invalidate, reclassify or contextualize evidence—but never overwrite the parent directly.
7 · Decision LedgerQuote, approval, research, evidence, score changes, human review, cost reconciliation and outcomes remain append-only.
What a complete receipt should prove
QuestionWhat exactly was asked and which clarifications were accepted.
QuoteExpected cost, safe estimate, hard cap, pricing snapshot and expiry.
AuthorizationWho approved the research and the maximum permitted spend.
RouteModels, tools, provider attempts, fallback and degraded-success state.
EvidenceSource IDs, hashes, root ancestry, acquisition time and limitations.
AssessmentEvidence Support, uncertainty, methodology version and scoring payload.
ChallengesSeparate issue IDs, research depth, materiality and parent effect.
Actual costGateway generation IDs, actual spend and variance from quote.
Revision historyWhat changed, why it changed, and what prior result was superseded.
What the system refuses to pretend
- Provenance is not automatically proof that a real-world narrative is true.
- A copied story repeated 1,000 times is not 1,000 independent sources.
- A surprising claim is not rejected merely because it sounds unusual.
- A challenge does not change a parent score merely because someone submitted it.
- A research quote is not permission to exceed the authorized maximum.
- An AI model does not determine cryptographic validity or silently rewrite canonical history.
TRY IT
Start with a question, not a workflow.
The site handles the routing, economics and evidence mechanics underneath.