Fictional study · deterministic browser demonstration. Not live AI, a deployed agent or a client result. All participants, notes and review decisions are invented.
A decision brief the client can use
Study question: How do users prepare a reporting product’s output for client discussions?
Ready to start — Start with the client question and inspect the first teaching draft. Tailoring is optional.
Which reporting step should we examine first?
Preview of the intended result: clarify metric definitions with finance, compare one completed client report with the built-in export, or defer a larger change until evidence and ownership are clear.
The analyst prepares the evidence, the professional reviewer challenges its meaning, and the reporting owner decides the next check.
Open the complete worked engagementTailor this study
Try a source problem (I-02)
Six synthetic interview notes are permitted sources for this example. Starting prepares the first source-linked findings for review.
Export this record
Draft · unreviewed — the download below matches the record on this page exactly.
Read the export preview
Unreviewed draft · fictional demonstration
What makes a build useful?
Find the evidence. Follow a change. Use your method. Run and own the work. Explore how these design choices could help the people preparing, reviewing and receiving it.
RAG brings retrieved source passages into the context used to prepare a draft.
Context graphs connect evidence, versions and decisions so you can inspect what depends on a change.
GraphRAG uses graph relationships in retrieval. It is an option to compare when relationship questions justify it.
Proposed installation designs. The browser examples use authored records and rules; they do not execute RAG or GraphRAG. A small stable case may need only an approved pack and linked registers.
Explore all 26 build advantages
Find the evidence
Find the relevant evidence
Permission-aware evidence retrieval can reduce repeated source hunting.
Interaction or proposed walkthrough: Ask one supplied case question; inspect the permitted passage and why it is relevant.
How it could be built: Start with explicit fixed-pack passage selection. A deterministic selector is labelled a retrieval-design simulation; real indexed retrieval is a separate candidate.
Inspect the complete worked exampleGround drafts with RAG
RAG can ground generated work in relevant allowed sources.
Interaction or proposed walkthrough: Trace a drafted statement to the excerpts that would be supplied to a generation step, then compare with the source.
How it could be built: Show selected context and existing authored output side by side. Do not call this a live RAG run. Actual RAG requires a real generation/retrieval path and representative retrieval/grounding tests.
Inspect the complete worked exampleUse the current permitted version
Correct version, purpose and revocation checks can prevent stale or unauthorised material entering a draft.
Interaction or proposed walkthrough: Revoke or change a source; see it leave the eligible context and identify the affected output.
How it could be built: Source eligibility fields and pure filtering in a synthetic pack; no protected data. Production must enforce before retrieval/context and before release, including index/cache/derived-copy handling.
Inspect the complete worked exampleInspect the basis of a finding
Exact source lineage can make review faster to navigate and easier to defend.
Interaction or proposed walkthrough: Click a sentence, open the exact quoted span/version, return to the same sentence; distinguish fact, report, inference and unknown.
How it could be built: Reuse current citation navigation/export; extend to the CV07 claim map. Citation fidelity does not establish that the inference is professionally sound.
Inspect the complete worked exampleExplore GraphRAG where it helps
GraphRAG may help when relationship-based retrieval is necessary.
Interaction or proposed walkthrough: Open a worked relationship question, such as which eligible current method and source support this claim, and inspect the permitted traversal.
How it could be built: Authored traversal over the synthetic logical graph labelled “GraphRAG design example”. Only justify actual graph-assisted retrieval if representative questions outperform simpler retrieval at acceptable cost/permission control.
Inspect the complete worked exampleKeep private notes out of public searches
Separate public research can add context without copying private case material into searches.
Interaction or proposed walkthrough: Compare an invented private question with a separately approved sanitised public question and its proposed source register.
How it could be built: Static controlled example, no live web call needed. No claim of current public-research execution; actual outbound question would require its task scope and permissions.
Inspect the complete worked exampleFollow the work and its decisions
Complete useful work
Complete work, not isolated snippets: scope, analysis, report and finance preparation can use the same accepted question.
Interaction or proposed walkthrough: Open the actual CV07 report and compare its recommendation with the brief and allowed alternatives.
How it could be built: Versioned synthetic case manifest and existing artifact views. A complete authored report demonstrates shape and coherence, not autonomous production.
Inspect the complete worked exampleFollow decisions in a context graph
A logical context graph answers “what depends on this?”
Interaction or proposed walkthrough: Select Source N06 → finding → report section → review decision; change N06 and inspect the affected records/owners.
How it could be built: Typed stable IDs and relationships in ordinary JSON or existing records, derived from the case contract. No graph database required; no invented links.
Inspect the complete worked exampleCarry the same case basis
Shared versioned contracts preserve meaning across jobs.
Interaction or proposed walkthrough: Change the accepted scope or report version and see which downstream brief/acceptance/milestone checks reopen.
How it could be built: Explicit case, scope, report, acceptance and milestone references. Do not merge independent existing fixtures just because their screens sit together.
Inspect the complete worked exampleReview the exact version
Exact review/version binding keeps a changed document from inheriting approval.
Interaction or proposed walkthrough: Review v1, edit the basis, then see the prior review marked superseded and the new draft queued.
How it could be built: Reuse current basis-bound research reducer; extend exact report/method/receiver binding. A demo approval remains fictional.
Inspect the complete worked exampleMake corrections actionable
Human decisions and owned return paths make correction actionable.
Interaction or proposed walkthrough: Reviewer returns a specific statement with reason; operator fixes it; receiving owner requests clarification or defers.
How it could be built: Typed return/receiver decisions with basis, reason, owner and next action. Preparation, professional issue and client acceptance stay distinct.
Inspect the complete worked exampleCheck calculations and open balances
Deterministic calculations and reconciliation reduce avoidable numerical ambiguity.
Interaction or proposed walkthrough: Follow CV07 cost arithmetic and M2 partial payment; retry allocation without changing the balance twice.
How it could be built: Pure calculation and stable fixture event IDs. A spreadsheet can be the best implementation; no real money or accounting write is implied.
Inspect the complete worked exampleUse and protect your method
Use your approved method
The firm's approved method can shape analysis, rubric and output.
Interaction or proposed walkthrough: Choose two explicitly authored method/template profiles; see changed required sections, evidence checks and reviewer questions.
How it could be built: Versioned method/configuration profile consumed by existing engine. Formatting customisation does not change professional rules; method changes require their owner's review.
Inspect the complete worked exampleReuse expertise with permission
Rights-aware method reuse can reduce repeated setup while protecting client material.
Interaction or proposed walkthrough: Nominate the generic assessment method, exclude case-specific interview text, then inspect applicability and permission for a second fictional case.
How it could be built: Reuse manifest with method version, owner, rights, expiry and explicit no-reuse result; reuse existing safe-method-reuse engine patterns. Redaction never creates rights; CV07 interviews are case-only.
Inspect the complete worked exampleGive each role the access it needs
Scoped identities and least privilege can reduce unnecessary exposure.
Interaction or proposed walkthrough: Switch between operator and finance perspectives; finance sees accepted report/milestone references rather than raw interview text.
How it could be built: Synthetic role projections derived from the same manifest. This is not real authentication or proof of server-side isolation; production enforcement must live outside the model.
Inspect the complete worked exampleKeep source text separate from authority
Source-instruction and tool boundaries can keep evidence from becoming authority.
Interaction or proposed walkthrough: Inspect a harmless marked source containing an instruction to change the job; see it excluded as an instruction and leave the proposed external-action queue unchanged.
How it could be built: Authored threat/reaction fixture and allowlisted action representation. Do not claim general prompt-injection resistance from one canned example; RAG does not remove this risk.
Inspect the complete worked exampleAgree the evidence lifecycle
Data lifecycle and provider choices can keep processing within agreed boundaries.
Interaction or proposed walkthrough: Inspect an example record's allowed purpose, expiry, retention and handling owner; see a revocation queue and a separately retained financial-history obligation.
How it could be built: Synthetic lifecycle view with explicit unresolved provider/region/training/retention fields. No certification, blanket deletion or real tenant check claimed.
Inspect the complete worked exampleRun and own it
Recover without duplicating work
Recovery, idempotency and manual fallback preserve useful work through interruption.
Interaction or proposed walkthrough: Interrupt after preparation; inspect last authorised checkpoint, retry the same event, resume only with the pending review intact, or use the manual packet.
How it could be built: Existing local recovery controller plus explicit checkpoint/event fixture. Does not prove provider recovery, backup restore or receiving-system idempotency.
Inspect the complete worked exampleSee what needs attention
Monitoring and ownership expose blocked work before it becomes invisible.
Interaction or proposed walkthrough: Open “What needs attention?”: case/version, reason, age, owner, next decision, last successful step and permitted cost summary.
How it could be built: Derived queue from local case events and fixed specimen clock. No implication of live monitoring, alerts, 24-hour service or installed management console.
Inspect the complete worked exampleSet operating limits
Budget, runtime and retry caps make operating exposure discussable.
Interaction or proposed walkthrough: Set an illustrative work budget; exceed it and see the case pause with preserved draft and next owner.
How it could be built: Deterministic usage fixture clearly marked assumed, with no model-provider price asserted. Real costs depend on quoted providers, usage, review and support.
Inspect the complete worked exampleChange the system deliberately
Controlled changes can keep improvements from silently breaking accepted work.
Interaction or proposed walkthrough: Switch an authored method/model-configuration version and inspect the affected regression checklist and rollback reference.
How it could be built: Versioned release metadata plus proposed checks. A green simulated checklist must never imply tests actually ran; actual tests/results are recorded separately.
Inspect the complete worked exampleWork with familiar tools
Existing-tool fit can reduce adoption and migration burden.
Interaction or proposed walkthrough: View the same brief as a web workpaper and a readable export for an existing case folder; inspect a small finance-field handoff.
How it could be built: HTML/JSON/CSV plus manual export/readback. Word/case folders/finance are fictional scenario touchpoints; SharePoint, Teams, email, licences and APIs remain unverified buyer assumptions.
Inspect the complete worked exampleKeep a portable handover
Portability and handover let the buyer operate the capability or change implementer.
Interaction or proposed walkthrough: Open a sample handover folder: dictionary, source/method/config versions, roles, tests, runbook, known limits and export manifest.
How it could be built: Small downloadable synthetic example bundle generated from current data. Agreed use rights and third-party licences govern; no blanket promise to own all vendor software.
Inspect the complete worked exampleChoose how it is operated
Client-owned, managed and hybrid operation can match available team capacity.
Interaction or proposed walkthrough: Select an operating model and see responsibility/coverage/recurring-cost categories change while professional authority stays with the firm.
How it could be built: One role-responsibility matrix and operating-model configuration. Quote-specific scope, no automatic retainer or 24-hour coverage.
Inspect the complete worked exampleCompare the complete cost of work
Whole-case economics can distinguish useful capacity from attractive but incomplete drafting metrics.
Interaction or proposed walkthrough: Compare current method, repaired manual method and proposed system across preparation, review, correction, operation, accepted quality and cost; unknowns remain blank.
How it could be built: Reuse existing deterministic capacity concepts and a clearly hypothetical comparison. Do not add savings from each workflow to the same whole-case hours, invent current pricing, or equate released hours with cash.
Inspect the complete worked exampleChoose a simpler method when it fits
A simpler/no-build route makes the recommendation proportionate.
Interaction or proposed walkthrough: Choose a fixed pack, template, spreadsheet or deferment and receive a useful next-action note rather than a failed demo state.
How it could be built: Branch in the decision artifact with stated reasons. Build only if evidence justifies extra retrieval, graph, integration and operating burden.
Inspect the complete worked exampleSwap the pieces: your transcripts for these notes, your reviewer for the partner, the output your client pays for in place of these findings. If the job still looks familiar, that is the one worth a Fit Call.
Compare total preparation, review, correction, coordination and recurring operating effort at comparable accepted quality. Check a glossary, template or existing product before a custom build. No savings were measured here. A paid Blueprint maps one revenue path and fully designs three connected opportunities chosen from actual evidence. Installation is a separate purchase with its own tests and a quote-specific fee.