# Follow the work. Inspect the decision. A self-guided Firm Capacity walkthrough · Build n Bloom · version 1.0 · 26 September 2026. Northbank is fictional. The records and reviewer decisions below are authored examples. This page runs no AI model and connects to no client system. Its calculator performs local arithmetic only. No live workflow test or client result is shown. ## 01. Start with the whole engagement Northbank sells an operating-model diagnostic. It has more work to deliver, but senior people keep rebuilding context and correcting first versions. A complete case uses assumed 18 senior, 36 analyst and four operations hours, including allocated selling/scoping, preparation, review, correction and coordination. W1 includes one senior, one analyst and 0.2 operations hours of allocated pursuit/scope effort per completed case, unchanged by the proposed automation. Forty cases use 720 senior hours. Waiting is separate. The design covers one path: enquiry/scope → usable evidence brief → reviewed report → client acceptance → finance handoff → collection record and permitted learning. Three improvements connect along that path; the first possible installation is the report/review capability. No savings from the other two are included in its model. Open sample-blueprint.html for the complete design. Open synthetic-effort-ledger.csv to check the baseline. The numbers describe this teaching case only. ## 02. Open the inputs before the output The approved-scope specimen excludes staffing recommendations. Source S-01 says the engagement lead prepares recommendations, the practice reviewer approves material conclusions and the client sponsor decides actions. S-02 records one handoff sent Monday and received Wednesday; another has no receipt date. Neither records workload, staffing needs or the cause of the delay. Method M-01 requires a source and locator for each material observation, separates interpretation from recommendation and routes missing evidence to a reviewer. Inspect working-records.html for the exact brief, source register, passages and method. The records are fictional and visible, so you can assess whether the proposed conclusion follows. ## 03. A fluent conclusion does not create its evidence **Authored draft:** 'The client can safely reduce its leadership team by two roles.' It cites S-02. The source records a handoff, not staffing capacity. **Fictional reviewer return X-01:** reject the staffing inference as unsupported and outside the agreed scope; retain the supported observation; return any staffing request to the managing partner. An authorised scope change is required before gathering staffing evidence or making that recommendation. **Revised authored result:** 'The sampled approval records show an unresolved handoff; leadership structure cannot be concluded from this evidence.' State: revised draft. No real professional approval, client issue or receipt exists. The proposed installed capability would display original passages beside the finding and require a named review decision. It must be tested on actual authorised cases; this authored return does not prove a production model obeyed that boundary. ## Prepared work beyond a single finding Open prepared-report.html and report-inputs.html: eight source records become a decision brief, three source-linked findings, reviewed-inference examples and an owner handover. These are substantial authored specimens, not a live AI run or professional approval. The original rough note is included so the change in reviewability is visible without inventing a time saving. ## 04. Carry useful work to the person who decides W1 produces a versioned scope/evidence brief. W2 can prepare source-linked findings, apply mechanical checks, assemble a report and route gaps. The qualified reviewer decides whether the advice is adequate. A source change invalidates affected findings; an absent reviewer blocks issue; a permission refusal stops that source; a replay cannot create another accepted version. These are proposed requirements. The commissioning scope needs case-limited identities, enforceable action ceilings, actual source locators, receiving receipts, exception owners and manual recovery. It also needs current provider terms and live permission checks. The client receives the complete design and can commission another competent implementer. ## 05. Use the simplest close-out that works W3 matches the approved report, customer acceptance, scope milestone and completion evidence for finance. Existing fields, approval workflow and a deterministic checklist can carry this job. The recommendation is no custom AI close-out agent. Finance approves invoicing and records actual collection. A practice lead separately decides whether a method can be reused. Removing names does not automatically confer rights. Conditional acceptance, absent rights and an accounting outage keep their relevant record pending. ## 06. Change the assumption; change the decision First-install W2 targets are nine senior, 22 analyst and 1.5 operations hours per case, compared with 12, 28 and one. On 40 existing cases this releases 120 senior and 240 analyst hours while adding 20 operations hours. Reserve first-year setup of 18/32/12 hours and annual oversight of 8/12/12. Assume 80 spare operations hours and six additional comparable cases of demand. The resulting first-year caps are six senior, six analyst, eight operations and six demand: six added cases. Total effort at 46 cases is 716 senior, 1,424 analyst and 231 operations hours. Assumed availability is 720/1,440/240. No fixed-payroll cash saving is claimed. With collected fee 18,000 and incremental cash cost 8,000, each additional case contributes 10,000 before annual non-labour system cost of 6,000. Six cases yield 54,000 before unknown one-time investment K; three yield 24,000; zero yield -6,000. All financial inputs are synthetic, AUD excluding tax, not quotes or forecasts. This package-level model compares the proposed change with the unrepaired baseline. Incremental AI value against the repaired-method alternative is unmeasured; the model does not justify an AI installation price. If W2 still needs 10.5 senior hours, senior capacity allows only two added cases: 14,000 before K. If the repaired method can produce the same accepted work with less burden, prefer it. Open buyer-approval-pack.html for the forwardable decision and synthetic-model-inputs.json for the assumptions. ## 07. Inspect a separate real source review Our public-source review is a different evidence class: completed work with public originals. It examines a statement about government AI pilots, finds that the cited survey cannot establish a pilot success rate and narrows the conclusion. Inspect the original, calculation and reasoning at https://www.buildnbloom.io/downloads/proof/public-source-review.html. It demonstrates source interpretation. It does not supply Northbank's fictional figures, prove advisory-client outcomes or certify a deployed system. Keeping these evidence types separate makes each claim easier to check. ## 08. Take the decision to your team Share the buyer approval pack. Ask partners whether this is the binding constraint, whether the simpler route wins, whether reviewers and operations have capacity, whether sources are permitted and what the downside does to the investment. The next step can be the free 25-minute Firm Capacity Fit Call. No confidential files are required. A paid Blueprint, if appropriate, maps your revenue path and fully designs three connected opportunities. A first installation, stabilisation and any continuity have their own scope, price and authority.