The prototypes below are under development. Local tests do not establish live service reliability, AI accuracy or wallet/payment readiness. Status is recorded for this checkpoint and is not updated automatically.
PROTOTYPE • LIVE EVALUATION PENDING01 — Answering questions from one source
- Question
- Can a tool summarize pasted text and explain which parts support an answer?
- Implemented
- A source-text form with an optional question, English/Hebrew output and instructions to separate source claims from unknowns. Access is restricted to enabled test wallets. It does not search the web.
- Local evidence
- Automated tests cover rejected unauthorized requests, input limits, simulated AI output and provider failures. The provider is mocked: these tests do not measure a real model's answers or quotation fidelity.
- Unresolved
- No live model accuracy evaluation is recorded. Quotations and interpretations may be incorrect and require comparison with the source.
- Next experiment
- Use a fixed set of sources and answerable/unanswerable questions in both languages. Manually compare every quotation and identify claims unsupported by the source. Report failures as well as successful answers.
View research prototype →
05 — Five-item development batch: roadmap 33–37
Question: can local workflows reuse evidence without silently charging credits, transmitting files or claiming verified facts?
Implemented locally
Sequential CSV batches with per-file errors and cancellation; explicit export of minimal Pro snapshots; optional preparation of those snapshots for source research; same-mint chronological snapshot comparison; a development register.
Recorded local results
42 automated tests pass for the cumulative application. New tests demonstrate that a malformed CSV fails individually while the next valid file runs, cancellation prevents the next file from being read, and size/count limits reject oversized work. Snapshot tests exclude account/payment fields, preserve unknown versus zero and reject different mints, reversed/equal timestamps and invalid metrics.
Failures and unresolved observations
The CSV errors above are deliberately constructed test cases, not reported user incidents. Separately, the owner’s real ELON payment showed success in an explorer screenshot but site verification rejected its association with the quote. Credits remain unresolved; the recovery UI is local and is not a proven fix. No second transfer is requested.
Limits and next experiments
No live browser or model evaluation was performed for this batch. Compare two real exported reports on mobile, test download/import/cancellation, and review source-based AI output in both languages. Imported JSON is not authenticated chain evidence; differences can reflect provider/pool changes. There is no scheduler, background monitor or autonomous transaction execution.
Open local workflows
06 — Private registry and credit-aware developer tools
Local tests exercise repeated and concurrent registration, per-wallet rating replacement, owner self-rating rejection, rating resets on revision changes, invalid URLs, capacity limits and retry after a simulated lost write acknowledgement. Paid SDK tests require explicit acceptance, preserve request IDs and expose pending/refund responses without automatic retry.
47 cumulative local tests pass; 12 server functions bundle. The storage tests are simulations. They do not prove hosted concurrency, a successful paid run or a completed marketplace. The existing live payment-credit mismatch remains unresolved. Next: private hosted acceptance, moderation/retention planning and mobile UI tests when deployment resumes.
07 — Program records without implied funding
Exact token-unit tests preserve the split total including rounding, reject over-allocation and duplicate plan IDs, and keep self-review and partnership exports explicitly unverified/unsent. Private grant tests exercise immutable terms, repeated submissions, test-wallet voting, applicant-only outcomes and retry after a simulated lost write acknowledgement.
51 cumulative local tests pass and 13 server entrypoints bundle. No real grants, transfers, independent quality reviews or partnership agreements were tested or created. Private proposals remain unpublished. Next: real mobile and hosted storage acceptance, then operating decisions before any public program.
ISSUES & LIMITATIONSWhat still needs work
- The Pro Connect button was hidden by a mobile navigation style. A local CSS fix is included; the owner subsequently demonstrated a signed-in wallet in screenshots. Full mobile acceptance remains pending.
- A live payment was reported, but verification rejected its reference and credits remain unresolved. Further hosted checks are deferred. Publication of frontend files alone does not establish that server functions are working.
- AI report and draft quality are not measured by tests that simulate provider responses. No accuracy percentage is claimed.
- Private AI prototypes require enabled test wallets. Public rollout needs usage controls and completed live validation.
We will revise these entries when new evidence is available. A proposed experiment is not a completed result.