Research workbench foundations for Trueness.ai
The reference synthetic panel supplies a working interaction pattern for audience rehearsal: read the evidence, ask a synthetic audience to react, synthesize disagreements, and challenge the synthesis. Trueness.ai can develop that pattern into a maintained research service by adding independent human validation, explicit limits, tenant controls, and a commercial operating system. The existing implementation is not evidence that synthetic responses predict human outcomes.
This is an internal development assessment, dated 25 October 2026. Exact source paths are retained in the private local source-lineage log. The labels below identify the reviewed components without exposing the source organization. Source research remains in its original project; it has not been copied into this venture.
What exists
| Source | Observed capability | Lesson for Trueness.ai |
|---|---|---|
reference corpus metadata, metadata inspected |
Snapshot built 14 October 2026; 28 documents across brand research (17), consumer research (8), and audience definitions (3) | Separate source types, publication dates, bases, and methodological status rather than flattening everything into persona prose |
reference persona metadata, structure and mounted corpus metadata |
Twelve synthetic people spanning three audience segments and two markets | Useful qualitative browsing interface; twelve generated voices do not constitute twelve independent human observations |
reference corpus builder |
Assembles extracted research, audience slides, and chart readings with provenance | Preserve extraction lineage, source version, and whether a number was printed or estimated from a chart |
reference research API |
Cited document Q&A, individual conversations, panel reactions, chair, challenger, reconciliation | Reuse the workflow concept; implement empirical validation as a separate service |
reference evidence storage |
AES-256-GCM encrypted corpus, immutable versioned blobs, local development fallback, cache | Version evidence snapshots and make lineage reproducible; use private per-tenant storage and managed key controls for the venture |
reference stimulus reader |
Reads documents, decks and images with explicit size/page/image caps and warnings | The customer must know exactly which stimulus pages were processed; incomplete evidence must block comparable-score claims |
reference research interface, sections.js |
Evidence benches, source reading, audience selector, attachments, preserved panel result if synthesis fails, adversarial review | Keep useful intermediate results; allow resumable work and source inspection |
reference PDF exporter, api/lib/research-pdf-html.js |
Exports the material already reviewed, including synthesis and challenge | Export a fixed result version, evidence scope, model version, validation status and limitations together |
reference brand-page component, reference brand-pyramid copy specification |
Brand pyramid, audience, positioning, products, voice, reasons to believe, future identity guidance | Use the written structure. The reference brand's prices, visual identity and claims do not transfer to this business |
The source corpus includes qualitative, survey, segmentation, brand and synthesis work. This inventory establishes what was assembled, not a new audit of every underlying study or a license to reuse it. Brand recommendations and human observations must remain distinguishable in the derived evidence ledger.
Changes required before commercialization
- Replace model confidence with measured status. The chair currently generates a confidence number. A model-written number is not a calibrated probability. Show held-out human comparison, applicable scope, uncertainty, last validation date, and approved uses.
- Keep generated voices identifiable. The persona prompt asks the model to act as a real person and never disclose it is synthetic. Trueness.ai should label generated responses in the interface, exports, and tool outputs; immersive phrasing must not turn into fabricated human testimony.
- Separate challenge from validation. The current panel, chair and challenger use the same configured model. A different provider can expose additional disagreements, but agreement between models still does not establish human fidelity.
- Make numeric outputs deterministic. Calculate summary statistics from structured responses in code, not the chair's prose. Do not infer population proportions from the number of simulated people.
- Meter every attempt. Existing code retries truncated generations and repairs malformed JSON. A truncation retry can return only the final attempt's usage through the current helper. The venture's cost ledger must record initial, failed, repaired and retried calls as well as successful outputs.
- Separate immutable run stages. Panel, synthesis and challenge should be durable jobs with idempotency keys. A synthesis failure must not rerun and recharge a successful panel.
- Replace shared access gates. Use tenant authentication, role-based permissions, isolated storage and signed access; a shared project passcode is not a commercial tenancy system.
- Restrict file fetching. The reviewed stimulus code accepts HTTPS attachment URLs. A commercial service should accept only tenant-owned uploaded asset IDs, enforce byte and time limits before processing, and prevent internal-network or arbitrary-URL fetching.
- Turn expiration into a scheduled obligation. Source corpus versions, generated outputs and stimuli have different retention needs. Scheduled deletion and recovery evidence should not depend on a later upload triggering a cleanup routine.
- Bound model upgrades. The existing model constant is convenient for a prototype. Every commercial release must record a supported provider/model version, prompts, evidence and validation protocol, and pass regression testing before replacing a prior version.
What can transfer
Product knowledge, research-workflow experience, general orchestration patterns, lessons from failures, and an independently written implementation can inform the new venture. Confidential reports, participant statements, segmentation definitions, customer assets, code ownership and rights to use a client case study require their own provenance record. Contribution of effort alone does not establish ownership of every underlying asset.
Pilot demonstrations should use commissioned research with appropriate participant information and client permission, or data whose license permits the specific use. Do not publish the original client corpus, persona biographies, screenshots, raw quotes, or performance claims as a Trueness.ai case study without the relevant rights. A clean synthetic demonstration may illustrate workflow, but it cannot demonstrate calibration quality.
Practical migration order
First preserve the reusable workflow in a clean specification. Next build tenant/evidence/run boundaries and a blinded validation harness. Then commission the first narrow human study, record the outcome including an inconclusive result, and release only the use cases it supports. Add ongoing refresh, customer self-service and assistant integrations after this first cycle works.
Enterprise continuation status
No Trueness research application or validated audience has been built by this planning work. The published planning-document renderer is not the product platform. Rights and legal review of the proposed reuse precede the first product build and source ingestion. The enterprise design carries forward the six useful actions—Panel, Verdict, Challenge, Talk, Ask the research and Export—while adding programme/market/wave records and controlled agency access. Historical demonstrations establish workflow examples only; they do not prove adoption, decision improvement or prospective calibration.