Which setup should your client use?
What changes when the evidence does?
Bring test results and their costs. Compare what actually passed, write the recommendation, and keep the result connected to its assumptions. When an input changes, the desk flags the affected work for recomputation and review.
No sign-up or installation. This independent Second Run build demonstrates the workflow with fictional tasks, outputs and costs.
1,847 real public records you can load and inspect.
The Public tail, 2026-09-29 packet holds 433 sources and 1,414 claims from public pages: 85 ClusterMAX review pages with 1,273 verbatim claims, 343 SemiAnalysis newsletter posts, 3 captioned videos with 74 quotes, the TCO calculator's assumptions (55 line items, 7 MTBF presets) and 5 July 2026 GPU price medians from OpenComputePrices. Every source carries its URL, SHA-256 and retrieval time so it can be re-fetched and re-hashed.
Claims start as open and remain separate from acceptance until reviewed. Sources are recorded as observed, not verified; the packet has no calculation records. Its checksum and event chain detect changes. Loading replays every event and takes about a minute. Rebuild and check it with packets/build_packet.js and packets/validate_packet.js.
A correction changes the recommendation.
A client requires at least 90% of tasks to be correct and on time. Two configurations process the same 24 fictional invoices. Configuration A costs half as much, but the original grader rejects its explanatory text around otherwise correct JSON.
A: 18 / 24. B: 22 / 24.
Only B clears the acceptance floor.
Recommend B within this sample.A: 22 / 24. B: 22 / 24.
Both qualify; A has half the supplied cost.
A becomes the lower-cost eligible option.The desk flags the old calculation and note, records a new calculation, and requires another review before freezing the revised report. The original delivery remains intact. A correct answer that arrived late still fails.
Try it in five minutes.
- Open the example. In Client delivery, enter a reviewer name and rationale, record the review and freeze the baseline report.
- Return to Start here. Choose Change the contract to extraction. The app opens the comparison and marks the earlier calculation and client note stale.
- Choose Record recomputation, then Draft a successor client note.
- In Client delivery, review and freeze the revised note. Download its HTML report. Both versions remain available.
- Choose Save packet to keep your work. On another visit, use Open packet to restore it.
Use it for one real decision.
1. Bring the evidence.
Start a blank workspace. Add permissioned sources, claims and task-run results in the documented input format. Set quality, deadline and cost rules before comparing runs.
2. Inspect the conclusion.
Keep failed attempts in the bill, compare compatible runs, and review what the evidence supports. Revise a source or rule to see which dependent records need attention.
3. Deliver and retain it.
Export a reviewed client report and save the workspace. A later correction creates a new version rather than silently replacing what the client originally received.
A separate Technical diligence view applies the same approach to supplied contracts, pilots, MOUs, cash receipts and operating-dependency evidence. Inspect its fictional company report.
What the desk does, and what you still do.
The desk organizes supplied evidence, checks its own contracts, computes comparisons, traces recorded dependencies and exports reports. You choose and obtain the evidence, judge source reliability, review conclusions and decide what to tell the client.
The desk records sources, claims and task-run results in documented formats. Model execution, live retrieval, system audits and external publication remain outside this workflow; existing exports may need an adapter.
He can keep the application.
The offline copy is one HTML file containing the code and public example. It opens in a current desktop browser without an installer. The source and supported contracts are available alongside it; no email attachment bundle is needed.
App v1.0.0 is preserved byte-for-byte from the delivered standalone. This page is the simpler public handoff. Sample reports establish software behavior, not measured GPU performance, verified commercial revenue, investment suitability or BEP adoption.