Public-sector source-packet workflow

Civic Staffer AI

Civic-Staffer-AI turns research/source packets into decision-grade public-sector strategy outputs.

Upload source packets, choose the outputs and reasoning/cost profile, then generate and review the selected package.

Start source-packet workflow Verify live Gemini on Google CloudMock/manual mode works without API keys.

Public-sector consulting workflow

Source packets in, decision package out

The main path is intentionally simple: provide the source material, select the outputs that matter, then run one package-generation action.

Inputs

Research/source packets from ChatGPT, Gemini, budget files, local reports, or other curated public-sector materials.

Outputs

Raw report, decision memo, executive brief, decision card, quality review packet, and final export-ready guidance.

Next action

Use the single workflow button to generate only the selected output package with visible progress and status.

Strategy packet map

From executive recommendation to audit trail

Executive briefRecommended allocationCurrent-vs-proposed spendRed-team critiqueImplementation roadmapVerification gatesDecision memoAudit log

Quality Review source-packet workflow

Civic-Staffer-AI source-packet workflow

Civic-Staffer-AI turns research/source packets into decision-grade public-sector strategy outputs. Upload or paste source packets, choose the output package and reasoning/cost profile, then generate and review the selected package.

main workflow

Generate from research/source packets

Provide one or more source packets, choose the outputs you want, then run one selected workflow action.

Quality Review from source packets uses locally imported PDFs or pasted curated source packets plus the Decision Memo Quality Gate. Files are processed in this browser and are not uploaded or stored. This is not live automated Deep Research or live web ingestion.

decision question

Decision question / problem statement

State the question you want answered. Source packets below are the evidence, documents, and pasted facts the answer should use.

Decision question / problem statement

input

Upload research/source packet

Upload one or more PDFs, or paste manually generated ChatGPT/Gemini, budget, or local-report packet text in the editor below.

Upload research/source packet

PDF files are processed locally in this browser and are not uploaded or stored.

ChatGPT Deep Research packetGemini Deep Research packetOfficial budgetLocal report

Packet 1

Research/source packet

Source metadata: Other

Packet tools

Packet title

Packet text

Paste source-packet text here, or use the single upload control above.

PDF files are processed locally in this browser and are not uploaded or stored.

Packet text is empty. Paste a curated source packet before Quality Review testing.

Source Library / Research Library

local import

Source Library

Manage persistent local-dev research records. For each run, all usable Source Library files are used by default; pasted/uploaded packets add context, and selecting files narrows only the Source Library portion.

No source-library records yet.

Imported files are read locally and only extracted text is stored in ignored local source-library data. This panel uses full stored text as source material for generation; it does not expose local file paths and does not run chunked retrieval.

execution

Choose execution mode

Pick deterministic local app code or a live API route before selecting outputs and generating the package.

Deterministic app codeNo paid API callDeterministic app code means no paid LLM/API call. The app uses repeatable local code for previews, formatting, readiness checks, and fallback/demo outputs. Use API mode for real model synthesis.API / Live model: OpenAIChecking API configUse the configured OpenAI route for real model synthesis. Final status will show the attempted provider, model, and applied reasoning.API / Live model: GeminiChecking API configUse the configured Gemini route for real model synthesis. Final status will show the attempted provider, model, and applied reasoning.

Selected execution mode: Deterministic app code
Selected provider: app-code
API key/config: No API key is needed for deterministic app-code mode.
Will Generate call a live model?: No
Expected live calls: 0
Expected architecture: deterministic app code

Deterministic app code means no paid LLM/API call. The app uses repeatable local code for previews, formatting, readiness checks, and fallback/demo outputs. Use API mode for real model synthesis.

Manual prompt mode is available in the manual research prompts section below; it is not run by Generate.

output package

Select outputs

Select only the outputs you want generated. Excluding the 12-18 page raw report can reduce time and cost.

Raw report / evidence synthesis, 12-18 pagesSource-heavy evidence synthesis for the deepest review path.Executive memo / source-limited internal reviewDecision-grade synthesis for public-sector leaders.Detailed business case memo / decision support packet, 15-20 pagesLong-form client decision support packet preserving the detailed business-case depth.2-page executive briefShort executive summary and tradeoff brief.1-page decision cardOne-page action, options, fiscal impact, risk, and decision line.Source QA audit packetSource packet, quality gates, audit preview, and cross-packet comparison.Executive internal-review packageBuilds the short Bryan-facing executive internal-review package.Always included.

Reasoning / cost profile

Standard: deterministic app-code mode does not attempt a live model call; common live model resolves after API provider selection. Separate reasoning parameters are only applied when supported; unsupported reasoning is shown as not applied.

Actual route if generated now: deterministic app code only. The planned four-setting matrix is still shown as policy preview; no live model is called in deterministic mode.

Show planned per-agent routing policy

Planned per-agent routing policy

This table is policy metadata before a run. API/live mode expects five staged provider calls by default; combined compatibility mode is shown separately when explicitly forced.

Agent / Task	Economy	Standard	Deep	Max
Problem Intake / Decision ArchitectureproblemIntakecritical	App codepreview	App codenormal	App codehigh	App codehighest / highest
Research DiligenceresearchDiligencejudgment	App codepreview	App codenormal	App codehigh	App codehigh
Variable ExtractionvariableExtractionroutine	App codedeterministic	App codedeterministic	App codenormal	App codenormal
Assumption RegisterassumptionRegisterroutine	App codedeterministic	App codedeterministic	App codenormal	App codenormal
Issue Tree / HypothesisissueTreeHypothesisjudgment	App codepreview	App codenormal	App codehigh	App codehigh
Diagnosis / Root CausediagnosisRootCausejudgment	App codepreview	App codenormal	App codehigh	App codehigh
Option GenerationoptionGenerationjudgment	App codepreview	App codenormal	App codehigh	App codehigh
Option Scoring / Tradeoff JudgmentallocationScorecritical	App codepreview	App codenormal	App codehigh	App codehighest / highest
Calculation ModelcalculationModelroutine	App codedeterministic	App codedeterministic	App codedeterministic	App codedeterministic
Sensitivity AnalysissensitivityAnalysisroutine	App codedeterministic	App codedeterministic	App codedeterministic	App codedeterministic
Portfolio OptimizerportfolioOptimizerroutine	App codedeterministic	App codedeterministic	App codedeterministic	App codedeterministic
Recommendation Selectionrecommendationjudgment	App codedeterministic	App codedeterministic	App codehigh	App codehigh
Red Team / Failure-Mode CritiqueredTeamcritical	App codepreview	App codenormal	App codehigh	App codehighest / highest
Implementation RoadmapimplementationRoadmaproutine	App codedeterministic	App codenormal	App codehigh	App codehigh
Current vs Proposed SpendcurrentVsProposedSpendroutine	App codedeterministic	App codedeterministic	App codedeterministic	App codedeterministic
Executive BriefexecutiveBriefroutine	App codepreview	App codenormal	App codehigh	App codehigh
Final Memo Synthesis / Recommendation DefensedecisionMemocritical	App codenormal	App codenormal	App codehigh	App codehighest / highest
Detailed Business Case Memo / Decision Support PacketdetailedBusinessCasejudgment	App codenormal	App codenormal	App codehigh	App codehigh
Source Packet ExtractionsourcePacketExtractionroutine	App codedeterministic	App codedeterministic	App codedeterministic	App codedeterministic
Source Readiness / Quality GatesourceReadinessQualityGateroutine	App codedeterministic	App codedeterministic	App codedeterministic	App codedeterministic
Basic Formatting / Status WorkformattingStatusroutine	App codedeterministic	App codedeterministic	App codedeterministic	App codedeterministic
Raw Report / Evidence SynthesisrawReportjudgment	App codepreview	App codenormal	App codehigh	App codehigh
Decision CarddecisionCardroutine	App codepreview	App codenormal	App codenormal	App codehigh
Source QA Audit PacketreviewerPacketroutine	App codedeterministic	App codedeterministic	App codedeterministic	App codedeterministic
Review PDF PackagefinalExportReadyPackageroutine	App codedeterministic	App codedeterministic	App codedeterministic	App codedeterministic

Current package: Executive memo / source-limited internal review; 2-page executive brief; 1-page decision card; Executive internal-review package. Excluded outputs: Raw report / evidence synthesis, 12-18 pages; Detailed business case memo / decision support packet, 15-20 pages; Source QA audit packet.

Review package status: clean in-app PDF downloads and optional print/export views are available after generation.

Prompt guidance and quality checks

These controls guide output selection, prompt instructions, and configured model-tier selection for live quality-review artifact drafting. Unsupported reasoning effort is shown as requested but not applied.

Output package quality check

decision requested
quantified options
fiscal impact
what money buys/frees/moves/risks
public harm/external cost
independent targets
contestable pool logic
studies needed
decision line

next action

Generate selected output package

Runs the source-packet workflow using the decision question, source packets, checked outputs, and selected reasoning/cost profile.

Missing source material: paste/upload a source packet or import a Source Library text/PDF file before generating.

Source readiness

Source size / run risk

low

Source mode: none
Source packets/docs: 0 packets / 0 Source Library docs
Approx. characters: 0
Approx. words: 0

Token pressure is approximate because no tokenizer is used.

Deterministic mode avoids paid API calls.

Gold/Silver Memo Architect Partner EditorDefault off; runs only after client-ready preflight passes and falls back on drift.

Percent complete0%

Current phaseReady

Estimated time remainingWaiting for source packet or Source Library file

Reasoning / costStandard

Selected outputs4

Missing source material: paste/upload a source packet or import a Source Library text/PDF file before generating.

No workflow has run yet.

Results will appear below the workflow controls.

Selected outputs: Executive memo / source-limited internal review; 2-page executive brief; 1-page decision card; Executive internal-review package. Excluded outputs: Raw report / evidence synthesis, 12-18 pages; Detailed business case memo / decision support packet, 15-20 pages; Source QA audit packet.

expected call

Actual API call for this run

Before generation, this shows the route the next run is expected to use if the current selections stay the same.

not attempted

Execution style: Deterministic app code
Requested provider: mock
Actual provider used: mock deterministic app code
Actual model sent to API: deterministic-app-code
Separate reasoning parameter: none; no live provider request
Temperature behavior: none; no live provider request
Execution scope: no live provider call; no per-agent API calls
Fallback/status: deterministic app code; no fallback needed
Outputs requested: Executive memo / source-limited internal review; 2-page executive brief; 1-page decision card; Executive internal-review package

Selected support outputs print/PDF

Opens one support view with each selected memo output once. Gold/Silver memos use document tables in this view; Bryan's executive review package and the source-verification audit packet stay separate.

Advanced: secondary actions

Provider selection uses configured model routing metadata for live quality-review artifacts. It does not add live Deep Research, source ingestion, PDF generation, or unsupported reasoning parameters.

Manual research prompts for ChatGPT/Gemini

manual external research

Create source packets

These prompts are not run by the app. Generate research packets manually, then paste the outputs below.

ChatGPT Deep Research source-packet prompt

Use ChatGPT Deep Research externally to prepare a source-backed public-sector strategy packet for this civic problem:

[Paste problem, geography, budget context, time horizon, and any current-spend notes.]

Return a concise Markdown source packet for later deterministic app review. Include these sections:
- Source inventory: source title, organization, date, URL or file label when available, and why it matters.
- Key facts: factual claims with the source signal beside each claim.
- Cost/budget figures: current spend, flexible/restricted funds, capital/operating split, program/category, fiscal year, and source signal.
- Current performance: inflow, exits, durable exits / retention at 6/12/24 months, returns, length of time homeless, public harm indicators, and cost per durable outcome where source-backed.
- Durable-impact evidence: cost per prevented entry, cost per durable housing exit retained at 6/12/24 months, cost per return to homelessness avoided, cost per day/month reduction in length of time homeless, cost per retained PSH placement, cost per unit of public harm reduced, and net durable progress.
- Full-spectrum cost table inputs: prevention/emergency rental assistance, eviction prevention/legal aid, diversion, traditional emergency shelter, tiny house/non-congregate shelter, rapid rehousing, Health Through Housing/hotel acquisition, permanent supportive housing, and affordable housing production/preservation.
- $10M translation: source-backed service-unit estimates where available, and clearly labeled illustrative planning assumptions where not. State these are service units, not verified impact units.
- Public harm / external cost: EMS/fire/police, sanitation/cleanup, parks/libraries/transit/public-space impacts, ED/jail/crisis-system use, overdose/deaths, business/community impacts, and public-space loss.
- Public-cost framing: distinguish homelessness response spending, public external cost, and total social/public cost. Do not invent a total external-cost number.
- Benchmark and target evidence: research, comparable jurisdictions, local constraints, and any support for 1-year, 3-year, and 5-year targets.
- Intervention evidence: what is known about prevention, diversion, retention support, rapid rehousing, shelter-to-housing transitions, supportive housing, acquisition/preservation, system coordination, and performance governance.
- Tradeoffs: fiscal, equity, feasibility, provider capacity, acuity/risk adjustment, public-risk, and opportunity-cost tradeoffs.
- Gaps/conflicts: missing data, conflicting sources, weak evidence, and claims needing human verification.
- Assumptions: planning assumptions separated from source-backed facts.
- Source-backed claims: only claims tied to inspected sources.
- Do-not-overclaim: what the packet cannot conclude from the available evidence.
- Studies needed: spend map, cost-per-durable-outcome, public-cost / external-harm, contract performance, benchmark, system-flow, prevention/diversion targeting, and fund-source flexibility studies.

Downstream output package context:
Output package controls:
- Selected outputs: Executive memo / source-limited internal review; 2-page executive brief; 1-page decision card; Executive internal-review package
- Excluded outputs: Raw report / evidence synthesis, 12-18 pages; Detailed business case memo / decision support packet, 15-20 pages; Source QA audit packet
- Reasoning/cost profile: Standard - production default route for normal quality-review synthesis.
- Executive review package selected: yes - always included
- Generate only the selected outputs. Do not generate unchecked outputs.
- Optimize effort/cost based on the selected package; use Economy/Standard for cheaper drafts, Deep for stronger judgment work, and Max only as planned highest-route policy for four critical judgment stages.
- These controls can select configured provider model tiers for live quality-review artifact drafting. Unsupported reasoning effort is displayed as requested but not applied; do not claim provider capability or billing changes.
- The raw report is not selected; preserve the skeptical-performance-first decision logic inside the selected shorter outputs.
- The executive review package is the short Bryan-facing artifact. Detailed support and source QA audit packets must remain separate by default.
- Align every selected output to this chain: human need -> public harm -> current spend map -> current performance -> independent benchmarked targets -> performance-budget reset -> contestable pool -> reallocation -> validated bottleneck -> funding decision.
- Preserve the skeptical-performance-first posture, an independent Homelessness Performance & Investment Office / target-setting function, a 90-180 day performance-budget reset, a 5-10% contestable pool from flexible legacy funds, quantified option tables, Studies needed to better inform strategy and future action, and What would make this recommendation wrong.
- When the executive memo / source-limited internal review is selected, use the Silver-Level business-case architecture as an honest executive memo target: eight main-body sections, compact executive tables, consolidated body subsections, and detailed machinery in appendices rather than a compressed full business case.
- When the detailed business case memo / decision support packet is selected, preserve the 15-20 page long-form business-case depth as a separate client-facing artifact; do not replace the executive memo, and do not expose raw staged debug scaffolding in the body.
- Label source-backed facts, illustrative planning assumptions, and future study/local validation needs.
- Treat +$50M-$100M/year = $250M-$500M over five years as illustrative target-sizing only, not funded, not acquisition, and not a KCRHA bailout.

Do not invent citations. Do not present unsupported estimates as facts. Do not argue that homelessness is large, therefore spending should rise. Keep the output ready to paste into a reviewer source-packet field.

Gemini Deep Research source-packet prompt

Use Gemini Deep Research externally to create an independent source packet for this civic strategy review:

[Paste problem, geography, budget context, time horizon, target outcomes, and constraints.]

Produce a concise Markdown packet that can be pasted into a deterministic quality-review workflow. Include:
- Source inventory: source title, publisher/organization, date, URL or file label when available, and relevance.
- Key facts: local baseline facts, demand/capacity facts, outcome facts, and source signals.
- Cost/budget figures: current spend, flexible/restricted funds, discretionary/restricted status, capital/operating split, funding constraints, and source signals.
- Current performance: inflow, exits, durable exits / retention at 6/12/24 months, returns, length of time homeless, public harm indicators, and cost per durable outcome where source-backed.
- Durable-impact evidence: cost per prevented entry, cost per durable housing exit retained at 6/12/24 months, cost per return to homelessness avoided, cost per day/month reduction in length of time homeless, cost per retained PSH placement, cost per unit of public harm reduced, and net durable progress.
- Full-spectrum cost table inputs: prevention/emergency rental assistance, eviction prevention/legal aid, diversion, traditional emergency shelter, tiny house/non-congregate shelter, rapid rehousing, Health Through Housing/hotel acquisition, permanent supportive housing, and affordable housing production/preservation.
- $10M translation: source-backed service-unit estimates where available, and clearly labeled illustrative planning assumptions where not. State these are service units, not verified impact units.
- Public harm / external cost: EMS/fire/police, sanitation/cleanup, parks/libraries/transit/public-space impacts, ED/jail/crisis-system use, overdose/deaths, business/community impacts, and public-space loss.
- Public-cost framing: distinguish homelessness response spending, public external cost, and total social/public cost. Do not invent a total external-cost number.
- Benchmark and target evidence: research, comparable jurisdictions, local constraints, and any support for 1-year, 3-year, and 5-year targets.
- Intervention evidence: evidence for prevention, diversion, retention support, rapid rehousing, shelter-to-housing transitions, supportive housing, performance governance, and where evidence is mixed.
- Tradeoffs: implementation capacity, fiscal realism, equity, acuity/risk adjustment, provider capacity, public harm, political/public transparency, and speed.
- Gaps/conflicts: missing current-spend data, source disagreements, outdated figures, and weak claims.
- Assumptions: assumptions that must be verified before a professional recommendation.
- Source-backed claims: claims that can be traced to the source inventory.
- Do-not-overclaim: conclusions that should wait for better evidence or expert review.
- Studies needed: spend map, cost-per-durable-outcome, public-cost / external-harm, contract performance, benchmark, system-flow, prevention/diversion targeting, and fund-source flexibility studies.

Downstream output package context:
Output package controls:
- Selected outputs: Executive memo / source-limited internal review; 2-page executive brief; 1-page decision card; Executive internal-review package
- Excluded outputs: Raw report / evidence synthesis, 12-18 pages; Detailed business case memo / decision support packet, 15-20 pages; Source QA audit packet
- Reasoning/cost profile: Standard - production default route for normal quality-review synthesis.
- Executive review package selected: yes - always included
- Generate only the selected outputs. Do not generate unchecked outputs.
- Optimize effort/cost based on the selected package; use Economy/Standard for cheaper drafts, Deep for stronger judgment work, and Max only as planned highest-route policy for four critical judgment stages.
- These controls can select configured provider model tiers for live quality-review artifact drafting. Unsupported reasoning effort is displayed as requested but not applied; do not claim provider capability or billing changes.
- The raw report is not selected; preserve the skeptical-performance-first decision logic inside the selected shorter outputs.
- The executive review package is the short Bryan-facing artifact. Detailed support and source QA audit packets must remain separate by default.
- Align every selected output to this chain: human need -> public harm -> current spend map -> current performance -> independent benchmarked targets -> performance-budget reset -> contestable pool -> reallocation -> validated bottleneck -> funding decision.
- Preserve the skeptical-performance-first posture, an independent Homelessness Performance & Investment Office / target-setting function, a 90-180 day performance-budget reset, a 5-10% contestable pool from flexible legacy funds, quantified option tables, Studies needed to better inform strategy and future action, and What would make this recommendation wrong.
- When the executive memo / source-limited internal review is selected, use the Silver-Level business-case architecture as an honest executive memo target: eight main-body sections, compact executive tables, consolidated body subsections, and detailed machinery in appendices rather than a compressed full business case.
- When the detailed business case memo / decision support packet is selected, preserve the 15-20 page long-form business-case depth as a separate client-facing artifact; do not replace the executive memo, and do not expose raw staged debug scaffolding in the body.
- Label source-backed facts, illustrative planning assumptions, and future study/local validation needs.
- Treat +$50M-$100M/year = $250M-$500M over five years as illustrative target-sizing only, not funded, not acquisition, and not a KCRHA bailout.

Do not fabricate sources or citations. Flag uncertainty plainly. Do not argue that homelessness is large, therefore spending should rise. Keep the packet practical for an expert public-sector reviewer.

Advanced: deterministic readiness preview

Deterministic extraction preview

Deterministic preview only. This does not verify source truth, perform live research, or replace expert review.

Authoritative source verification and full source-to-claim mapping will be added in a later PR.

manual review required

Decision Memo Quality Gate

Skeptical-performance-first checklist for XPRIZE Decision Memos. This is a manual memo-standard gate, not automated factual validation.

Skeptical-performance first: not pro-spend, not pro-cut, and no 'homelessness is large, therefore spend more' argument.
Use the decision chain: human need -> public harm -> current spend map -> current performance -> independent benchmarked targets -> performance-budget reset -> contestable pool -> reallocation -> validated bottleneck -> funding decision.
Treat the raw report / evidence synthesis as the primary artifact: 12-18 pages, source-heavy, table-heavy, no filler, and no artificial compression before the raw report stage.
The executive memo / source-limited internal review, 2-page executive brief, and 1-page decision card must condense the raw report rather than substitute for it.
State the source-packet decision boundary: enough evidence for an interim performance-budget strategy, not enough evidence for final investment allocation by intervention.
Answer 'do we know enough now?' clearly: enough to reject blind cuts and blind unrestricted new spending; enough for controls, spend mapping, flexible-dollar review, and reallocation-first strategy; not enough to assign exact $X gains or losses by intervention.
Define the missing investment question as 'what does one unit of durable impact cost?', not only 'how much does a service cost?'
Durable-impact units include cost per prevented entry, retained durable exit at 6/12/24 months, return to homelessness avoided, shorter length of time homeless, retained PSH placement, unit of public harm reduced, and net durable progress.
Raw report covers human need, public harm, current spend map, fund restrictions, system flow, durable outcomes, cost effectiveness, benchmarks, options, studies needed, and what would make the recommendation wrong.
Raw report separates source-backed facts, illustrative planning assumptions, and future study / local validation needed.
The executive memo / source-limited internal review uses the Gold/Silver Business Case Standard as a true executive memo: eight main-body sections with detailed machinery carried in appendices, not a compressed full business case, a strict page-count promise, or a 24-section visible report.
Silver body sections are Executive Summary and Recommended Decision; Decision Logic; Situation; Complication; Options and $/Impact Logic; Recommended Path; Execution Conditions; and Next Steps: 90-180 Day Validation Gates.
Silver appendices carry detailed decision criteria, financial/value model, RACI/resourcing, legal/procurement/budget authority, data governance, provider transition/service continuity, evidence architecture, caveats/manual verification, and structural changelog.
Executive Summary is a concise narrative: preserve approve / do not approve / conditionally approve and confidence content in prose, without standalone Decision Block, Confidence, or old scaffold headings.
Options, Tradeoffs, and $/Impact Logic is the analytical core and should be more developed than the Recommended Path section in the detailed memo.
Keep stakeholder and approver mapping out of the main table of contents; fold it into governance, RACI, legal authority, budget authority, and change management.
Define the north star: reduce human harm and public harm per dollar by making homelessness rarer, briefer, and less recurring.
Current spend map separates total vs flexible spend, restricted vs discretionary funds, capital vs operating dollars, contracted dollars, and discretionary dollars.
Performance dashboard flags PIT count, unsheltered count, annual estimate, inflow, exits, durable exits, returns, duration, deaths, public-cost indicators, cost per durable outcome, and contract/audit/procurement performance.
Public harm section considers EMS/fire/police, sanitation and cleanup, parks/libraries/transit/public-space impacts, ED/jail/crisis use, overdose/deaths, business/community impacts, and public-space loss.
Target-setting and performance governance answers who sets targets, what independent function owns the roadmap, and whether targets are benchmarked or only self-reported.
Recommend creating or assigning a small independent Homelessness Performance & Investment Office / strategy function if neutrality or capability is missing.
Set proposed 1-year, 3-year, and 5-year targets from research, benchmarks, local constraints, and local validation without inventing exact targets.
Define success metrics with risk-adjusted or acuity-aware treatment so high-acuity providers are not punished.
Run a 90-180 day performance-budget reset before treating current spend as optimized.
Analyze a 5-10% contestable pool from flexible legacy homelessness funding; clarify it is not a blind across-the-board cut.
Define contestable pool mechanics plainly: identify eligible flexible dollars; exclude restricted, legally committed, rent, and protected life-safety funds; flag 5-10% for renewal review; score with independent acuity-adjusted metrics; renew, condition, transition/rebid, or reallocate; provide transition and appeals; reinvest in homelessness/housing outcomes.
Contestable-pool analysis compares blunt cut, provider performance challenge, holdback / contestable pool, reallocation-first, and targeted new money after bottleneck validation.
Contestable-pool safeguards protect high-performing and life-safety services, include transition time, appeals/review, reinvestment into better outcomes, and independent review rather than provider self-grading.
Option table includes spend less, stay current, reallocate current dollars, 5-10% contestable pool, and spend more targeted.
Every option row quantifies what money is freed, moved, added, bought, or risked losing, using source-backed figures or clearly labeled planning assumptions.
Section 5 options table columns are Option, Fiscal move, What money buys / frees / moves, Expected durable impact, Public-harm effect, Main risk, and Evidence confidence, matching docs/xprize/prompts/decision_memo_synthesis_prompt.md.
Appendix A uses an evidence ledger with confidence limited to VERIFIED, SECONDARY, INFERRED, and NEEDS-DATA; Appendix B includes benefit-stack logic, funding-source logic, service-unit vs durable-impact distinction, and Diagnostic tests (live).
Include a full-spectrum housing/intervention cost table and a $10M service-unit translation table; warn that service units are not verified impact units.
The 1-page decision card uses columns: post-reset budget path, annual fiscal move, what $10M equals, when this path makes sense, and key risk if chosen without reset; do not include a 'decision' column.
The 1-page decision card includes decision requested, recommended action, quantified options, fiscal impact, what money buys/frees/moves/risks, top risks, studies needed, and an approve / modify / reject decision line.
Separate homelessness response spending, public external cost, and total social/public cost; do not invent a decision-grade external-cost number when the packet lacks one.
Frame risks as risks of deciding without the reset: blind cuts, blind new spending, wrong bottleneck, provider punishment, public-cost blind spot, and governance failure.
Run a reallocation-first reset; targeted +$50M-$100M/year equals $250M-$500M over five years and is allowed only after validated bottlenecks remain that optimized current dollars cannot solve.
Treat any +$50M-$100M/year scenario as illustrative target-sizing only, not a funded proposal, not a real-estate acquisition fund, and not a KCRHA bailout.
Any added money identifies a funding source or offset: new revenue, current homelessness reallocation, other city/county reallocation, state/federal match, philanthropy/institutional partnership, or proven avoided crisis costs.
State it is not a blind cut, provider self-evaluation exercise, or spend-more-because-need-is-large memo.
Lower-cost / higher-impact alternatives include reallocation, targeted prevention, diversion, retention, CES/navigation throughput, vacancy/lease-up reduction, outcome-based contracts, Medicaid/state financing, fund braiding, procurement/audit/contract management, and high-cost crisis-user focus.
Include a required 'Studies needed to better inform strategy and future action' section.
Studies needed cover spend map, cost-per-durable-outcome, public-cost / external-harm, contract performance, benchmarking, system-flow, prevention/diversion targeting, and fund-source flexibility.
Include a required 'What would change the recommendation' section.
Convert recommendation-changing risks into a plain-language risk table with trigger, owner, mitigation, and decision implication.
Use bracketed placeholders such as [insert legal authority], [insert approving body], [insert reset staffing estimate], [insert baseline metric], [insert target threshold], [insert data owner], and [insert verification source] instead of invented certainty.
Do not imply exact savings, exact benchmark targets, or named-provider cuts unless source-backed and reviewed.

Source Packet Quality Gate

Packet-readiness check only. This does not validate source truth, confirm final claim support, or run live research.

MISSINGTotal source count detected

Detected 0 source-labeled lines or citation signals.

MISSINGOfficial/public-agency source count detected

No source-labeled lines with public-agency signals were detected.

MISSINGRecent source count detected

No 2024-2026 source-labeled lines were detected.

MISSINGFiscal dollar facts detected

No dollar lines with fiscal/budget language were detected.

MISSINGUnit-cost facts detected

No dollar lines with unit-cost language were detected.

MISSINGDurable-impact unit evidence detected

No durable-impact unit evidence was detected. The packet can support interim performance-budget controls, but not final intervention-by-intervention allocation without cost per durable impact unit.

MISSINGClaim audit entries detected

No key claim, fact, or cost rows were detected for source-to-claim audit preview.

MISSINGWeak/unsupported/estimated claim flags detected

No weak, unsupported, estimated, or needs-verification signals were detected by the deterministic rules.

Coverage flags

MISSINGSeattleNo deterministic Seattle coverage signal was detected.

MISSINGKing CountyNo deterministic King County coverage signal was detected.

MISSINGKCRHANo deterministic KCRHA coverage signal was detected.

MISSINGaffordable housingNo deterministic affordable housing coverage signal was detected.

MISSINGhomelessness preventionNo deterministic homelessness prevention coverage signal was detected.

MISSINGshelterNo deterministic shelter coverage signal was detected.

MISSINGpermanent supportive housingNo deterministic permanent supportive housing coverage signal was detected.

MISSINGbehavioral healthNo deterministic behavioral health coverage signal was detected.

MISSINGgovernance/accountabilityNo deterministic governance/accountability coverage signal was detected.

Total packets: 1
Claims: 0
Facts: 0
Cost figures: 0
Assumptions: 0
Gaps/conflicts: 0
Weak/unsupported: 0

Packet 1: Research/source packet

Other / 0 characters

Run the deterministic preview for this packet to surface claims, facts, costs, assumptions, gaps, conflicts, and source signals.

Advanced: synthetic strategy demo

Synthetic strategy demo

Mock planning workflow. Use this only when you want the synthetic strategy packet instead of the source-packet-grounded Quality Review workflow above.

Civic problem required

Optional: improve the recommendation

Planning budget

Geography

Current spend

Time horizon

Goals

Constraints

Target outcomes

local demo

Ready for mock run

Synthetic strategy demo: the mock provider produces the full packet without API keys.

Selected source-packet execution: mockSynthetic strategy demo provider: mockQuality Review source mode: not generatedCalculation: deterministic-app-codeSource-packet workflow: local PDF import, manual prompt support, selected-output generation.

Advanced: workflow modes

Workflow modes

Demo mode is not the ceiling of the product.

The current demo runs in safe, repeatable mock mode for reviewers. The intended pro-grade path uses stronger models and real source packets for judgment-heavy work, while fiscal math, score totals, allocation percentages, source labels, and audit logs stay deterministic or source-backed.

Demo mode

Mock/synthetic, repeatable, and safe for judges. It demonstrates workflow shape without live research.

Efficient real mode

Lower-cost models handle extraction and formatting, while a stronger model handles synthesis.

Pro-grade output mode

Stronger models support judgment-heavy research, diagnosis, options, red team, and final recommendation, with deterministic app code for math, labels, allocations, and audit records.

Max-Max / flagship mode

Reserved for the most critical synthesis, red-team, and final recommendation stages where explicit reasoning-effort control is required.

Workflow stage	Demo/mock mode	Efficient real mode	Pro-grade output mode
Problem intake	User prompt plus synthetic defaults	Structured intake with lightweight normalization	GPT-5.5 high/xhigh intake review with stakeholder and constraint framing
Decision framing	Repeatable synthetic decision frame	Stronger model shapes the decision question	GPT-5.5 high/xhigh frames decision rights, tradeoffs, and verification burden
Research/diligence	Synthetic diligence patterns, no live research claim	Lower-cost extraction plus stronger synthesis	GPT-5.5 high/xhigh research over real source packets, web findings, and uploaded files
Source extraction	Mock/source-label examples	Lower-cost model extracts facts and citations	Source-backed extraction with file/web research review and traceability
Variables/assumptions	Synthetic assumptions with labels	Extracted values plus stronger synthesis for gaps	GPT-5.5 high/xhigh assumption discipline with deterministic source labels
Issue tree/hypotheses	Repeatable issue tree	Stronger model for issue framing	GPT-5.5 high/xhigh issue decomposition and hypothesis testing plan
Diagnosis/root causes	Synthetic diagnosis	Stronger synthesis over extracted facts	GPT-5.5 high/xhigh diagnosis grounded in real evidence packets
Option generation	Synthetic portfolio options	Stronger model generates options from extracted context	GPT-5.5 high/xhigh option design with feasibility, equity, and implementation constraints
Comparative scorecard	Deterministic scoring totals	Model suggests rationale; app computes totals	GPT-5.5 high/xhigh score rationale; deterministic app code computes totals
Fiscal/allocation model	Deterministic allocation math	Extracted fiscal inputs; app computes allocation math	Source-backed fiscal inputs; deterministic app code computes allocations
Scenario comparison	Synthetic scenario comparison	Stronger synthesis with deterministic scenario math	GPT-5.5 high/xhigh scenario judgment with deterministic comparison math
Current-vs-proposed spend	Synthetic/current-input comparison only	Extracted current-spend inputs where provided	Source-backed baseline review with deterministic comparison labels
Red team	Repeatable red-team critique	Stronger model critiques recommendation and gaps	GPT-5.5 high/xhigh red team across evidence, fiscal, implementation, and public-risk claims
Recommendation synthesis	Synthetic recommendation packet	Stronger model synthesizes extracted work	GPT-5.5 high/xhigh final recommendation with explicit uncertainty and verification gates
Executive brief/final memo	Repeatable memo output	Stronger model drafts concise memo	GPT-5.5 high/xhigh executive memo with source-backed caveats and decision framing
Public transparency summary	Synthetic transparency language	Stronger model formats public summary	GPT-5.5 high/xhigh public transparency summary tied to evidence and unresolved risks
Audit log	Deterministic mock audit trace	Provider/model/prompt trace plus deterministic records	Deterministic audit log for sources, prompts, provider calls, math, labels, and errors

Advanced: synthetic demo reference

Example demo prompt

Reduce homelessness and increase affordable housing in King County with a constrained public capital budget. Recommend a portfolio that balances impact, equity, implementation feasibility, cost-effectiveness, leverage, and risk.

Synthetic demo warning

This XPRIZE demo uses King County-style synthetic planning inputs. It is not an official King County engagement, does not provide final funding authority, and should be treated as planning guidance until local source documents, budget constraints, legal requirements, and accountable decision-makers verify it.

Civic Staffer AI

Source packets in, decision package out

Inputs

Outputs

Next action

From executive recommendation to audit trail

Civic-Staffer-AI source-packet workflow

Generate from research/source packets

Decision question / problem statement

Upload research/source packet

Research/source packet

Source Library

Import stored source text

Add metadata-only record

Choose execution mode

Select outputs

Planned per-agent routing policy

Output package quality check

Generate selected output package

Source size / run risk

Actual API call for this run

Create source packets

ChatGPT Deep Research source-packet prompt

Gemini Deep Research source-packet prompt

Deterministic extraction preview

Decision Memo Quality Gate

Source Packet Quality Gate

Coverage flags

Packet 1: Research/source packet

Synthetic strategy demo

Optional: improve the recommendation

Ready for mock run

Demo mode is not the ceiling of the product.

Demo mode

Efficient real mode

Pro-grade output mode

Max-Max / flagship mode

Synthetic demo warning