We label customer results, operating history, synthetic demonstrations, modelled cases and acceptance records so you can see what each record shows.
Human-reviewed workflowThe prepared report is a fictional process example. The sample Blueprint shows scope and test structure.
Open the original inputs, a substantial prepared report and the decisions that return unsupported work. Then inspect the complete design and the investment case. The advisory case is fictional; the separate public-source review contains real public originals.
Six invented interview notes, three source-linked findings and an overstated conclusion returned for correction. No live model execution or client result is claimed.
Inspect the notes and reviewInputs, responsibilities, alternatives, economics, proposed tests and a ranked first installation.
Open the complete sampleScope, participation, data boundaries and the scenario that changes the recommendation.
Open the approval packThese are component-level observations from public datasets. They do not test an installed client workflow or establish human time saved, client capacity or financial results.
Measured public-data component tests
Observed public-data tests, not professional acceptance. Read the test record.
ContractNLI results are shown as matches out of 20 for each batch. The exact-evidence measure is a stricter subset of classification matches.
Pilot: 13/20 classification matches; 9/20 with exact evidence. Balanced labels; reference matching is not professional acceptance.
| Batch | Classification | + exact evidence | Limit |
|---|---|---|---|
| Pilot | 13/20 | 9/20 | Balanced labels; reference matching is not professional acceptance. |
| Isolated retest | 10/20 | 6/20 | A revised prompt and presentation did not improve the reference score. |
| Document-disjoint continuation | 12/20 | 4/20 | Public training contamination remains possible. |
Separate measure: FinQA pilot recorded 13/15 numerical or boolean matches. Literal quotation checks recorded 40/40 continuation outputs, which tests copying only and is not comparable with either ContractNLI measure.
Modelled teaching inputs
Fictional per-accepted-case role hours. Targets are untested and are not measured improvements. Inspect the model inputs.
Senior and analyst inputs fall in this constructed scenario, while operations rises from 4 to 4.5 hours. Capacity needs the whole accepted case, including that handoff work.
Senior: 18 existing hours and 15 target hours per accepted case. The target remains untested; operations increases in this model.
| Role | Existing input | Target input | Status |
|---|---|---|---|
| Senior | 18 h | 15 h | Synthetic; target untested |
| Analyst | 36 h | 30 h | Synthetic; target untested |
| Operations | 4 h | 4.5 h | Synthetic; target untested |
Inspect the full test record and its limits. The capacity chart uses fictional teaching inputs from the downloadable effort ledger; it does not report an improvement.
The evidence type sits beside the claim. A mechanism demonstration shows how the system behaves; a measured client record shows what happened in comparable use.
Observed, attributable, permissioned and published with its actual workflow, period and limitation.
A dated operating fact that establishes history or discipline—not performance in a different buyer market.
Realistic example data used to inspect behaviour. It is not customer work or customer performance.
An explicit scenario built from stated assumptions. It informs a decision but does not report an observed outcome.
A versioned record of tested system behaviour, exceptions, reviewer authority, defects and handover.
Build n Bloom's NDIS operating history sits in its own operating context. It does not establish the performance of an advisory, engineering or private-capital workflow.
Richard scaled a provider from zero to 26 homes over five years. That is Richard's individual operating history as an NDIS operator — not a Build n Bloom company result, and not evidence of advisory, engineering or private-capital performance.
The demonstrations run on fictional firm and pursuit data. Disconnect a source, introduce an unsupported claim, force a human decision, and inspect the record it leaves.
See which approved record supports a working claim, and where its approval state limits reuse.
Introduce an unsupported number, watch progression stop, and choose: supply evidence, rewrite, or remove.
Follow the source, the correction, the named approval and the version into the accepted record.
Workflow, metric definition, baseline, period, geography and source are recorded.
No extrapolation beyond the observed context.
The source supports the exact wording and conflicting evidence is retained.
Vendor or founder interpretation is identified.
Attribution, anonymity, client material and publication language are explicitly authorised.
No private or client-sensitive detail appears by inference.
Evidence class, date, context, source and limitations remain visible beside the claim.
Corrections change both the record and every dependent claim.
A published case record includes its baseline, test results, reviewer corrections, adoption friction and failures—not only a headline number.
Time, correction load, retrieval and exception measures use the same definition before and after.
Representative inputs, pass/fail criteria, defects, remediation and accepted version.
Exceptions, reviewer corrections, adoption friction and third-party dependencies remain visible.
Named, anonymised or private status and the exact language the client approved.
The demos are synthetic and clearly labelled. The Fit Call helps you decide whether your workflow is suitable for a paid Blueprint.
