WORKBENCH GUIDE / EXECUTE · OBSERVE · EVALUATE

Follow the evidence.

A practical guide to your first test, an honest comparison, and the evidence behind each result.

Open workbench

01 / YOUR FIRST TEST

One assumption.
Two independent tests.

  1. Open the workbench. Wait for verified runtime readiness.
  2. Keep the first transfer preset: 10 displayed units, with a multiplier changing from 1 to 2. The deliberately faulty reference ignores that scale.
  3. Press Run test. Inspect the observed delivery and the assertion verdict.
  4. Pin for comparison, choose Configure counterpart, then explicitly run again.
  5. Open Compare runs. Check equivalent semantic inputs, distinct run identities, and each report’s evidence.

The intended contrast is 10 requested → 20 observed for the faulty reference, then 10 → 10 for the corrected reference. Every displayed result must come from its own execution. A reference’s name never determines its verdict.

Open the workbench

02 / THREE ASSUMPTIONS

Test the boundary that matters.

01

Transfer quantity

Does a displayed-unit intent produce the correct raw transfer after scaling changes? Compare requested quantity, executable quantity, observed deltas and residual.

02

Price timing

Do quantity and price describe compatible adjustment states at the moment they are used? A guard can correctly refuse an input while valuation remains inconclusive.

03

Event attribution

Does explicit event evidence justify classification? Distinguish structural rescaling from reinvestment and check that repeated delivery applies only once.

↗

Independent evaluation

The reference proposes behavior. The runner observes actual runtime state. A separate evaluator checks the result. Missing evidence stays unknown.

Only registered, reviewed references and bounded presets are supported. The three core scenarios imply no live feed, issuer authentication, public-chain transaction, cash distribution, tax result or unrestricted application audit is implied.

03 / READ THE EVIDENCE

A completed test can fail.

Completed · pass
Every required assertion has sufficient evidence and passes.
Completed · failing test
The execution finished, and independent observations contradict a required assertion.
Completed · inconclusive
A required value or attribution cannot be established. Unknown is never converted to zero or a pass.
Operational error
The service could not establish or deliver a validated result. A stored example is never substituted.

Three useful downloads

Report explains the findings and verdict. Evidence retains observed state and provenance. Rerun configuration records semantic inputs, without expected results, credentials or source code.

Repeat supported tests in the hosted workbench. CLI reproduction requires a separately authorized private source checkout; a configuration download does not include runner source. No repository access is needed to use the public workbench.

Hashes help check consistency and integrity; they are not independent certification. Atomic balances remain exact strings. Supported scales, precision and the measured Token-2022 executable identity are retained in each report; its release version may be unknown.

04 / SESSIONS & RECOVERY

Know what the service is doing.

One test at a time
Busy means no new test was queued. Wait for cleanup, then explicitly retry.
Lost connection
An accepted test may still be running. Reconnect to the same submission intent rather than starting a replacement.
Cancellation
The worker receives a cancellation request. Cleanup must finish before another test can start. Cancellation does not undo already observed changes.
Temporary results
Hosted reports and downloads belong to your guest session. Download them within 20 minutes of completion. Guest sessions last up to two hours.
Restart or expiration
Temporary recovery may become unavailable. Previous tests are never automatically replayed, and result links are not permanent public shares.

Page loads, hover, form changes, readiness checks and downloads do not submit tests. The animated observation model is an illustration, not a runtime status indicator.

Explore labeled synthetic examples

05 / CAPTURE & REPLAY

One source observation.
A fresh test every time.

  1. Open Capture & Replay. Provider access must be available before an observation can be captured.
  2. Choose Pyth TSLA (the default), Pyth QQQ, or the PreStocks product set. Press Capture fresh observation once.
  3. Check source-age categories and the frozen evaluation clock. Retrieval time never replaces a missing provider timestamp.
  4. Run a native test, pin its result, then explicitly run the counterpart against the same snapshot.
  5. Replay fresh test creates a new no-value native position without requesting provider data again.

Pyth checks feed identity, freshness, market session and confidence. TSLA and QQQ each work as a single equity feed. Apple pairing is unavailable on the current trial and does not block either selection. PreStocks keeps token price, mark price, mark valuation and implied valuation distinct. Unknown units or timestamps stay unknown.

Benign inputs can pass in both implementations. Derived stress cases change a clearly labeled test copy and preserve the original capture; they are not claims about a market event.

Private data, useful findings

Capture downloads contain a data-minimized summary or an expiring session replay reference. They contain no raw market data, price-recoverable quantity/value pairs, dataset or runner source. Replay requires the original guest session and retained snapshot. Captures and all related reports expire together within 20 minutes of capture, or earlier when permission ends.

Open Capture & Replay