Research for this quiz

Start automatic intake research as an administrator, or download the manual research prompt below. Research saves drafts for review.
Anne Stewart

Multnomah County, OR OR · anne-stewart-neighborhood-match
Saved counties: Multnomah County, Washington, Clackamas, Yamhill
Saved cities / areas: Portland, Lake Oswego, West Linn, Tualatin, Tigard, Beaverton, Aloha, Hillsboro, Forest Grove, Wilsonville, Aurora, Oregon city, Beavercreek, Stafford, Milwaukie, Gladstone, clackamas, Happy Valley, Gresham, Newberg, McMinnville, Dundee, Cornelius, Carlton, Gaston, Sherwood
Priority communities: Not supplied
Research guidance: 2026-09-11-v3.1 · usually 100–300 useful areas; never a quota
Choose the output your next step needs

Research output
Includes proposed branch weights, exact JSON templates and local testing instructions. New quizzes must adopt the exported shared engine for scores to match. These are not recovered live quiz weights.
Confirm state / region for this assignmentResearch only these areas (optional)Additional instructions for this download (optional)
Preview the complete prompt
# ONE LEAD OS — Anne Stewart

Quiz: anne-stewart-neighborhood-match
Quiz record ID: 6aa1bf7be0f48dddc2933aac
Guidance version: 2026-09-11-v3.1
Output: Local Data Engine project for a NEW quiz

This is a manual research assignment. Downloading it does not start research, send email, import data or update a quiz. Research happens in your separate research assistant and browser. Run the returned data through the appropriate local engine/adapter before integration.

Saved intake information is context to verify, not factual evidence or a scoring instruction. Never turn protected-class descriptions into housing recommendations.

# ONE LEAD OS research assignment — Anne Stewart — Multnomah County, OR

Agent: Anne Stewart

State / region: OR

Market ID: anne-stewart-neighborhood-match

Approved county IDs: multnomah, washington, clackamas, yamhill

Home types: single_family, condo, townhome, high_rise

Pilot / due date: Not supplied — resolve before work that depends on it.

Scope: Discover supported areas across the approved market.

Selected areas: Whole market / see scope above

Service boundaries and exclusions: Saved counties: Multnomah County, Washington, Clackamas, Yamhill
Saved cities: Portland, Lake Oswego, West Linn, Tualatin, Tigard, Beaverton, Aloha, Hillsboro, Forest Grove, Wilsonville, Aurora, Oregon city, Beavercreek, Stafford, Milwaukie, Gladstone, clackamas, Happy Valley, Gresham, Newberg, McMinnville, Dundee, Cornelius, Carlton, Gaston, Sherwood
Verify the approved service boundary before dependent collection.

Priority starting points: Not supplied — resolve before work that depends on it.

Requirements / specific gaps: Relocation buyers, downsizers, families looking for good schools
We serve 4 counties in and around the Portland Metro area: Clackamas county, Multnomah county, washington county and yamhill county
These are proposed new-quiz branch settings, not extracted live weights. Review which home types and questions this market supports. The saved survey does not contain an exact engine adapter. Unsupported dimensions require implementation in the shared runtime before use.

Approved starting sources / IDX links: Use approved original sources; no authorized IDX credentials are included in this download.

## How this assignment fits the engine
Conduct research in your browser/research assistant. The local Data Engine does not browse sources. Return evidence-backed JSON for local quiz simulation and gauntlet tests. Copying this prompt, preparing a handoff, and running deterministic tests make no model API calls. The optional API key in Data Engine powers later explanation/grading only.

Use the approved territory and home-type configuration below. Typical whole-market coverage is 100–300 useful distinct areas with no quota or hard cap; never pad. A selected-area batch has no minimum count. Never expand a focused assignment to manufacture 100 records.

This is the shared NEW-quiz contract. The local simulator and exported component use identical scoring code/settings. Do not claim parity with an existing custom quiz. These starting weights need source-backed cases and owner review.

The full working rules and structural examples are embedded below. Named local files are not evidence that you accessed them. If a required current project, source or attachment is unavailable, identify it explicitly. Never fabricate a source visit or test result.

## Research instructions
# Copy-paste assignment for the researcher's assistant

Read START-HERE.md, RESEARCH-PLAYBOOK.md, SCORING-AND-GAUNTLET.md and the intended market/question contract before working. Build an evidence-backed neighborhood dataset for the agent's approved service geography using browser computer use, official public exports/APIs, document extraction and local validation as appropriate.

Normally target 100-300 useful distinct areas. There is no hard cap or minimum. Never pad to reach a number. If the supported result is about 90, document the discovery coverage and exclusions and prepare that explanation for agent review. Count a geographic area once even when it has several home-type segments. Preserve scope and provenance; do not copy the training examples into real research.

Start with a candidate/source roster and five varied pilot records. Research in batches of 15-25. Open original sources rather than treating search snippets as proof. For every scored/displayed claim, record URL, publisher, checked/as-of date, scope, source excerpt/page, units, home type and conflict status. Prefer county/municipal sources for geography and records, association documents for dues/access, builder sources for product details and authorized typed IDX/listing evidence for market guidance. Use scripts for bulk normalization and validation; use computer use to inspect interactive source details and verify copied search links.

Separate detached, condo, townhome and high-rise classifications under the intended contract. Keep price, fees, ownership, resident amenities and tradeoffs specific to each product. Unknown is not false or zero. Selected price examples are not market endpoints. Builder base prices are not finished-home prices. Public, paid and planned amenities are not on-site resident access. Do not infer protected traits or use demographics to recommend housing.

Run the local gauntlet after each batch. Write independent buyer expectations before tuning and preserve holdout cases. Classify issues as evidence gaps, factual conflicts, mapping errors, engine bugs, unsupported questions or real market scarcity. Research the affected fields, rerun regression cases and preserve run hashes. Do not adjust expected answers merely to obtain a pass. Do not claim a source, browser, API or test was checked unless it actually was.

The local simulator and exported new-quiz component use oneleados-new-quiz-v1.0.0 and identical code. Its initial weights are proposals to review. It does not reproduce legacy custom quizzes or implement school/commute scoring. The Matt Tilley reference is a separate pinned target, not a universal adapter. Identify the exact target adapter/importer work and browser checks still needed; record them as not run until verified.

Optional AI review happens only when the researcher invokes AI review lab with their own locally connected provider key, after the readiness gate. They can use OpenAI SOL and a separate strong grader or another supported provider/model. Normal collection/gauntlet commands never call it. Send only the fixed shortlist and supplied facts; enforce SOL-RERANK-PROMPT.md, unchanged scores, fixed IDs, original bands and source-grounded explanations. Never put a key in a dataset or shared artifact. Do not publish, contact agents or modify a live quiz as part of research.

Deliver frozen dataset/config/cases, coverage roster, claim ledger, reports and research tickets, source-review status, agent discussion brief and release checklist. Distinguish automated passes from human review and not-run checks. End with exact counts of unique areas and product segments, supported coverage, remaining gaps and the next action owner.

## Output a single importable project

Prefer one UTF-8 JSON file with top-level keys `name`, `records`, `config` and `scenarios`. `records` is the canonical segment array. `config.model_version` must be `oneleados-new-quiz-v1.0.0`. All three parts must have the same market_id. Copy the exact record shape from templates/research-segment.json, config shape and values from the supplied market configuration, and scenario syntax from training/scenarios.json. The examples are structural references, never factual evidence.

If you are in ChatGPT without direct filesystem access, have the researcher attach those template/reference files and the current project JSON. Do not claim to have read local files merely because this prompt names them. Return the revised JSON as a downloadable file and a separate source/research summary. The researcher drops that file into Data Engine and supplies the actual resulting report; never invent a simulated runner result.

Write at least ten source-reviewed buyer cases and five reserved holdout cases before tuning, with distinct buyer profiles and meaningful included/excluded/ranking expectations. Cover every offered home type and boundary/unknown-data conditions. Do not turn duplicated synthetic personas into an alleged independent evaluation. Keep manual source, target-browser, importer and IDX checks separately named.

Use the app's research-next-pass.json to target further collection. Preserve canonical IDs and source history. Never fill an unknown just to clear a warning. If weight or question changes seem justified, propose them with evidence and expected behavior for explicit review; never change the expected outcomes merely to pass. Normally target 100-300 distinct areas, with justified smaller/larger coverage and no padding.


## Source research playbook
# Dataset research and handoff playbook

ONE Lead OS | Data Engine 3.0.0 | September 11, 2026

## 1. The assignment

Build a useful, evidence-backed set of places that the agent's neighborhood quiz can actually compare. Research comes first; scores reveal where the research or question design needs work. We normally aim for 100-300 distinct high-quality areas. That range is a planning guide, not a hard minimum or ceiling. Do not stop at 100 if the market supports more useful coverage. Do not pad a market that supports fewer.

The owner describes the demo as having about 90 supportable areas. Treat that as the owner's example, not a fresh audit of the live demo. A final set of 90 can be correct. Explain the searched geography, sources, excluded duplicates, weak candidates and remaining gaps to the agent. Do not manufacture subdivisions, divide one place into arbitrary pieces or include out-of-scope places to reach a count.

Use the approved geography and priorities in the current assignment. Confirm the intended quiz branches, supported home products, IDX destination and release rules from that client's intake. A sample configuration is not evidence for the assigned market.

Deliver researched source files, a structured dataset, an evidence ledger, a coverage roster, repeatable buyer scenarios, local reports, unresolved research tickets and an agent review brief. The company engineer owns the new-route integration, lossless data handling and final publication. The researcher owns traceable facts and reproducible research runs.

## 2. Define the quiz-to-data contract before collecting hundreds of rows

List every question, its displayed options, when it appears, whether it is a hard requirement or preference, the exact data fields it uses, its score rule and the facts that can appear in results. Use templates/question-data-contract.csv. A question is supported only when its answer is collected correctly, mapped correctly, scored or routed intentionally, and reflected honestly in results.

Start with home type. A detached home, a condo, a townhome and a high-rise unit can have different price evidence, ownership, dues, maintenance, amenities and tradeoffs in the same geographic area. Keep a stable area_id and a separate segment_id for each supported product. Count the area once. Record legal ownership and physical form separately; a building's form does not establish its legal ownership. Define mutually exclusive quiz classifications so a high-rise condo is not duplicated merely under two labels.

Choose the branch before asking branch-specific questions. Attached-home branches may emphasize total required dues, included maintenance and resident amenities more heavily. Detached-home branches may emphasize property setting and location preferences. These are design choices to document, not universal facts. Mixed selections mean the buyer accepts either selected home type; score each segment under its own branch and consolidate results by parent area.

Clear hidden answers when the buyer changes branches. Do not let an old condo pool or fee answer silently influence a detached-only result. Write cases for switching forward and back. A question such as timing may affect follow-up instead of rankings; document that purpose rather than inventing a score effect. If the current data cannot support a proposed question, research the missing evidence or explicitly change the question contract with the owner.

## 3. Build the source register and candidate roster

Work county by county, municipality by municipality, then subdivision or recognized area. Record each authoritative portal's URL, coverage, field strengths, export method, access limits and last successful visit. Prefer existing official exports or documented APIs for bulk records when available. Use computer use for interactive maps, filters, association pages and difficult documents. Search results are discovery aids, not evidence for final claims.

Search systematically with patterns such as:

- `[county] Tennessee GIS subdivision map`, `[county] assessor parcel search`, `[city] planning subdivision plats`.
- `"[area name]" "homeowners association"`, `"[area name]" covenants`, `"[area name]" dues filetype:pdf`.
- `"[area name]" condos`, `"[area name]" townhomes`, `"[builder]" "[community]" quick move in`.
- `site:[official domain] "[area name]"`, followed by the exact document title, address or phase when needed.

For each candidate, record canonical name, aliases, county, municipality, boundary or geographic definition, area type, parent relationship and why it belongs in the quiz. Name variations and marketing phases may be one area. An independently named building or phase may deserve its own record when geography, access, product or market evidence distinguishes it. Explain the decision. Broader districts and subdivisions can both be useful, but document their overlap so they do not masquerade as independent recommendations.

Use the coverage roster to show discovered, researched, held, excluded and ready candidates. Preserve exclusions with reasons. Compare coverage by county, home type, price segment and useful property setting; use those dimensions to find research gaps, not to impose numeric quotas. Stop discovery when further passes mainly return duplicates, out-of-scope places or unsupported candidates and remaining omissions are documented.

## 4. Pilot five varied records, then collect in batches

Before scaling, research approximately five deliberately different cases: a mixed-product area, a condo/building, a detached area, a fee-uncertain area and a new-construction or upper-price case. A pilot should expose schema and source problems. Review these with the company engineer and agent where their local judgment is needed. Do not collect 250 rows using an untested interpretation of fees or boundaries.

Then work in batches of about 15-25 areas. Maintain a queue: identity, source evidence, product segmentation, prices/fees, amenities/lifestyle, geography/IDX, copy, validation and reviewer checks. Checkpoint files after each batch. An interrupted browser session should resume from the ledger, not repeat the whole market.

## 5. Collect each fact from the appropriate source

Identity and boundaries: use county GIS, plats, municipal planning and the association's own documents. Distinguish official boundaries, a developer's plat and informal market usage. An assessor parcel record establishes parcel facts; it does not by itself establish every amenity or the market definition of a neighborhood. Record coordinate scope: area center, entrance or building, with its evidence.

Home type and ownership: inspect typed listings, plats, association documents and builder product pages. Do not copy a condo price or fee into detached housing just because both share an area name. Verify whether a fee belongs to the master association, a sub-association, a building or a separate membership.

Prices: prefer authorized typed listing/sales evidence and dated public transaction records, with a stated lookback and sample method. Keep sold prices, current asking prices and builder base prices separate. Record home type, date, sample size, inclusions/exclusions and any skew from a luxury outlier or unusual sale. A few selected examples are examples, not a proven market floor and ceiling. Assessed value is not automatically a market price. Upper markets can remain open-ended. Do not invent a maximum just to fill an integer field.

Fees: use current association schedules, budgets, builder disclosures or sufficiently scoped listing disclosures. Preserve the source's amount, billing period, mandatory/optional status and what it covers, then normalize required recurring dues to a monthly total. Annual dues divide by 12; quarterly dues divide by 3. Keep one-time transfer fees, special assessments, taxes, insurance and optional club costs separately identified in evidence and narrative. If the recurring total is not established, mark it unknown. No result found is not evidence of zero dues.

Amenities: verify both existence and access for the particular home type. A public park nearby, an optional paid golf club, a private backyard pool and a shared resident pool are different facts. A planned clubhouse is not operating. Check the source page and any membership qualifications rather than treating a marketing icon as proof. If a master community has a pool, verify whether the specific phase/building has access. Record known absence only when evidence actually supports absence; otherwise use unknown.

Lifestyle: translate labels into observable property or place characteristics, such as a documented trail connection, historic built form or public water access. Keep subjective language qualified. A straight-line distance does not prove a walking route. Do not score or describe an area by resident demographics, inferred protected traits or a preferred kind of household. Stay with the buyer's expressed housing preferences and verified physical features.

Schools, when the actual target asks: use official district address tools, boundary maps and state reporting for the named school, measure and year. Do not convert a state measure into an invented A-F neighborhood grade. Do not promise every address has the same assignment. The shared new-quiz scorer does not implement school scoring; it needs a documented target adapter and specific tests before that question is supported.

Drive times, when the target asks: pin destination, origin scope, route, mode and measurement conditions. An entrance may be more appropriate than a centroid. Separate estimates from observed route results and traffic-sensitive measurements. The shared new-quiz scorer does not implement travel-time scoring. Never claim its gauntlet validated a commute it never calculated.

Current inventory and IDX: use the agent's authorized search destination. Apply the real area, home type and relevant filters in the browser; copy the resulting URL; reload it and confirm the filters survived. A generic city search is not a verified neighborhood search. Record exact scope and checked time. A functioning link is not proof that matching homes currently exist. An IDX display's public visibility does not automatically grant a bulk feed; use the authorized feed/export or permitted browser workflow.

## 6. Use computer use and extraction together

Open the actual source. Inspect its visible text or DOM, map legend, active filters and document links. Use normal UI controls to select county layers, search a subdivision, open its record and read the relevant fields. Save the exact URL plus the record/document identifier when the URL alone does not preserve state. Take a screenshot for map scope or a visually important claim; note which claim it supports.

For PDFs, download through the source's normal link when available, extract native text and tables, then inspect the relevant rendered page. Use OCR only when needed and check numbers, decimal points, footnotes, units and multi-column reading order visually. A table can place annual dues beside quarterly dues; extraction without headers can silently corrupt the amount.

Use scripts to normalize units, join stable IDs, detect duplicates, validate schema and compare revisions. Use documented official API/export data to avoid hundreds of repeated UI actions. Keep API paging, filters, timestamps and field mappings with the extract. Do not reverse-engineer hidden access or bypass login/CAPTCHA restrictions. If blocked, use another appropriate public source, the authorized account, or record the source limitation.

A script or AI can draft a structured claim from a page; the researcher must ensure the cited source actually supports that exact claim and home type. Save a compact excerpt or factual note with the source, not entire scraped sites. Cache reusable permitted research locally and revisit changed fields instead of rerunning every source for every buyer scenario.

## 7. Keep a claim ledger and resolve conflicts explicitly

Every scored or displayed assertion needs traceable source support. Record claim ID, area and segment, field, value, original units, normalized units, URL, publisher/document, checked date, effective/as-of date, excerpt or page number, scope, confidence and conflict state. The canonical JSON has evidence references; the CSV ledger can carry the fuller audit trail.

Use present, absent and unknown distinctly. Evidence references and a valid URL establish traceability, not truth by themselves. Resolve conflicts by authority for that field, matching product/scope and recency. A current association budget can clarify recurring dues while a county deed is stronger for a recorded transaction. Do not average contradictory fee values or choose the more favorable claim. Record both, investigate and keep the unresolved field qualified or unknown.

Write summaries and tradeoffs from reviewed claims. Do not let generated copy create facts outside the ledger. Recheck volatile prices, dues, construction and inventory close to release under a documented freshness plan. Set refresh owners and intervals by field and market; the kit does not automatically declare an old source fresh.

## 8. Run the gauntlet and use the failures to direct research

Validate each batch, then run frozen synthetic controls and independently specified buyer cases. The local HTML report shows shortlisted segments, scores, components, missing evidence and failed assertions. The JSON output preserves all evaluated segments, exclusions and hashes. Research tickets identify gaps and count affected cases at the score-dimension level; this is a triage aid, not proof of the exact cause of every rank.

Classify each weak result before changing anything: missing source evidence, contradictory fact, incorrect mapping, engine rule defect, unsupported question, or real market scarcity. Add the next concrete source action and owner. A true no-match case should remain no-match until the buyer explicitly changes a constraint. A research gap should trigger a targeted pass; a scoring bug belongs to the engineer. Do not alter expected outcomes just to produce a green report.

After corrections, rerun the same cases plus new regression cases for the discovered failure. Reserve holdout cases chosen before tuning. Have the agent review representative shortlists and explain misses using actual facts, rather than requiring their favorite place to win every scenario. SOL review is optional after deterministic results stabilize; it cannot repair missing evidence or certify the dataset.

## 9. Freeze, compare with the target and hand off

Freeze dataset, config, scenario and prompt versions. Compare the intended live quiz's exact code, home-type branches, importer, defaults, hard filters, scoring, tie handling, shortlist limits and AI bands to the local result. The new-quiz export uses the exact engine and renderer mounted by the local simulator. A legacy quiz needs its own exact adapter; it must not be silently migrated. The included Matt Tilley reference is target-specific and pinned separately. Mount the new-quiz export on the intended new-quiz route and verify the frozen cases there before publication.

Round-trip the data through the actual importer in an approved review environment. Read it back and compare IDs, distinct counts, product records, nulls, price qualifications, fee semantics, enum values, URLs and evidence. Watch especially for import defaults that turn unknown into zero or discard qualifiers. Run browser checks on desktop/mobile for branch changes, hidden answers, result edits, short/empty results, AI fallback and IDX filters.

Deliver the coverage explanation, completed sources and ledgers, frozen files, pass/fail/not-run results, unresolved items, agent comments and refresh plan. Public display subsets must reconcile with the master area's count and segment count. The release checklist records the exact reviewed version; an import success or synthetic test pass is not release approval. Do not publish or send agent messages as a side effect of this local tool.


## Scoring and test contract
# Scoring and gauntlet contract

## One shared engine for new quizzes

`tools/research-engine.mjs` is oneleados-new-quiz-v1.0.0, the shared opt-in runtime for new quizzes. The local simulator and exported renderer import it directly; there is no second copy of the math to drift. The earlier Matt Tilley bundle under reference/tilley remains a separate, target-specific reference. Existing working quizzes are not changed.

The included pilot template starts with proposed branch weights and vocabularies that need agent review and independent cases. The new-quiz export preserves those approved settings with its data and code. It is not a claim that these weights reproduce a pre-existing target route. See PRODUCTION-ENGINE-AUDIT.md for the production-source findings and chosen architecture.

## Actual implemented rules

Only declared home types and counties are accepted. Home types selected together use OR. Each record has one parent area and one home-type classification; duplicate area/type records fail validation. The primary shortlist shows one best segment per parent area, with deterministic segment-ID ties. There is no 100- or 300-record input limit. top_count limits displayed results only (1-20), independently of dataset size.

The config chooses visible questions and weights by home type. The buyer's visible-question set is the union of selected branches; hidden grouped score answers are pruned. Each segment's budget, fees, amenity, lifestyle and construction components use that segment's branch visibility and weights. Home type is required routing input. County routing is conditional on question visibility; hidden county answers are cleared. Required amenity and age-product filters also respect per-segment visibility. This engine reports visibility; the actual frontend must implement the same branching and clearing behavior.

Proposed pilot weights, in order budget / fees / amenities / lifestyle / construction:

- Detached: 40 / 15 / 15 / 25 / 5.
- Condo: 35 / 25 / 25 / 10 / 5.
- Townhome: 35 / 20 / 20 / 20 / 5.
- High-rise: 30 / 25 / 30 / 10 / 5.

These are inspectable starting choices, not recovered live target rules. Edit market.json under a new config_version when the intended contract differs. Lifestyle and amenity keys are configurable. School, commute, bedroom, lot size and maintenance scoring are not implemented dimensions in this generic model; unknown buyer keys fail instead of silently pretending to score them. Extend the code and tests or use an exact target adapter before offering those dimensions.

For active components, score = 100 * sum(weight * fit) / sum(active weights), rounded to two decimals. Fit is between zero and one. Missing evidence earns no unsupported positive points and remains marked unknown or limited. Evidence coverage is a separate weighted diagnostic (supported=1, limited=0.5, unknown=0), not an estimated probability of truth. With no active scored preferences the score is zero with a no_active_preferences flag; do not display it as a meaningful buyer-fit percentage.

Budget: a supported finite typical range fully within the buyer's range fits 1. Partial overlap fits 0.5 + 0.5 * intersection_width / market_range_width. A boundary touch receives 0.5. No overlap fits 0. An open upper market overlapping the budget receives limited fit 0.5. A selected observed example in budget receives limited fit 0.6; examples outside the budget do not establish that the market is unaffordable. Builder base price is unknown for finished-home affordability. A hard budget requires positive supported or qualified overlap, not a promise of a current available home.

Fees: a fully supported recurring monthly total within the maximum fits 1; entirely above fits 0; a range crossing the limit gets proportional partial fit. A hard fee limit requires the entire supported range to be within the maximum. Unknown fails a hard fee requirement, even if the limit is zero. This field covers recurring required dues; separate assessments and optional costs still need explicit copy and review.

Amenities and lifestyle: fit is the fraction of requested keys with supported presence. Known absence and unknown earn no positive credit, but retain different states. Only on-site access counts as a resident amenity. Nearby public, separate membership and planned entries do not earn that credit. A required resident amenity excludes records with absence or unknown. Construction matches an explicit preference for current new construction or documented absence; hard unknown/mismatch excludes. Age-restricted products use explicit documented restriction and explicit buyer choice, with no demographic inference.

## Prewritten cases and interpretation

Start with training/scenarios.json for syntax. Each automated case needs id, kind, buyer and expect. kind distinguishes synthetic controls, reviewer cases and holdouts. A manual case has mode:manual and a reason and remains not_run. Independent assertions can include top_contains, top_excludes, eligible_contains, eligible_excludes, top_count, score_above, visible, same_scores_as and component checks. See tests/engine.test.mjs for boundary cases.

Write expected outcomes from source-backed reasoning before observing results. Cover every offered type, mixed selections, each county, boundary budgets, unknown/zero fees, hard/soft requirements, partial amenities, public/paid/planned access, hidden-answer changes, no-match cases and upper-price cases. A generated sweep checks implementation invariants; it does not replace expert expectations. Hold out cases from tuning. Browser routing, actual source truth and live inventory need separate checks.

All automated cases also test repeated deterministic results, input immutability, bounded eligible primary scores and unique parent areas. Reports distinguish engine_status (scored/no_supported_match) from test status (pass/fail/not_run). An expected empty result is a passing test; an unexpected empty result is a failure. The provided training run has five automated cases, 40 assertions and one separate manual check.

Research tickets flag missing prices, fees, offered amenities/lifestyle, new-construction state, age restrictions, inventory, IDX verification and draft review. Affected-case counts are diagnostic impact counts for unknown/limited components, not causal attribution. The audit does not fetch sources, prove source truth, enforce freshness, validate every tradeoff claim or know when a geography is fully researched. Use the ledger and release checklist for those judgments.

## SOL is a bounded second stage

The deterministic engine chooses eligible candidates, scores and primary shortlist. SOL can order close calls within precomputed contiguous bands and write explanations from supplied facts. It cannot add/drop places, change scores or override hard requirements. Bands are anchored to the highest score in each new band, preventing chained ties from drifting far from the original score. The new-quiz template uses five points; actual targets can differ. The reviewed demo used six-point bands, so pin target behavior rather than generalizing one width.

Packets contain fixed facts, per-candidate evidence, buyer property preferences, model/config versions, input hashes and a packet hash. Changing notes or data requires a new packet; old responses fail packet matching. The local validator requires every original ID exactly once, stable band positions, complete explanations and candidate-owned evidence references. A small wording check catches some unsafe or unsupported phrases but is not a full factual or policy validator.

A real evidence ID can still be attached to a false claim. Human factual review must compare each sentence to that evidence, including price qualifiers, resident access, unknown costs and inventory limitations. Invalid AI output falls back to deterministic results and separately reviewed template explanations; this kit does not generate a production fallback UI. A successful SOL response saves a review artifact and receipt, not an automatic quiz update.

## Production parity checklist

For the supplied new-quiz contract, mount the exported component and preserve its files and bundle. For a legacy target, a separate pinned adapter must include its actual question normalizer, engine and market-specific hooks. Compare the same frozen inputs with the live/review target: visible questions, pruned answers, eligibility, component values, scores within a stated numeric tolerance, tie rules, distinct shortlist and AI bands. Record the commit and dataset/config hashes. Run importer/readback checks and browser behavior as separate results. Keep unknown/not-run checks visible and stop short of claiming target parity while they remain unverified.

The branded application adds the >90% independent-case gate, question evidence adviser, calibration proposals, version invalidation, provider connections and AI grading described in DATA-ENGINE-WORKFLOW.md. The CLI report by itself does not apply every application release gate.


## Exact market configuration for this job
```json
{
  "model_version": "oneleados-new-quiz-v1.0.0",
  "market_id": "anne-stewart-neighborhood-match",
  "config_version": "anne-stewart-neighborhood-match-proposed-v1",
  "status": "research_only_not_production_approved",
  "county_ids": [
    "multnomah",
    "washington",
    "clackamas",
    "yamhill"
  ],
  "top_count": 5,
  "rerank_band_width": 5,
  "lifestyle_keys": [
    "trails",
    "walkable",
    "water_access",
    "golf",
    "historic"
  ],
  "amenity_keys": [
    "pool",
    "fitness",
    "clubhouse",
    "spa"
  ],
  "branches": {
    "single_family": {
      "visible_questions": [
        "home_types",
        "county_ids",
        "budget_min",
        "max_monthly_fees",
        "amenities",
        "lifestyle",
        "new_construction",
        "age_choice",
        "notes"
      ],
      "weights": {
        "budget": 40,
        "fees": 15,
        "amenities": 15,
        "lifestyle": 25,
        "construction": 5
      }
    },
    "condo": {
      "visible_questions": [
        "home_types",
        "county_ids",
        "budget_min",
        "max_monthly_fees",
        "amenities",
        "lifestyle",
        "new_construction",
        "age_choice",
        "notes"
      ],
      "weights": {
        "budget": 35,
        "fees": 25,
        "amenities": 25,
        "lifestyle": 10,
        "construction": 5
      }
    },
    "townhome": {
      "visible_questions": [
        "home_types",
        "county_ids",
        "budget_min",
        "max_monthly_fees",
        "amenities",
        "lifestyle",
        "new_construction",
        "age_choice",
        "notes"
      ],
      "weights": {
        "budget": 35,
        "fees": 20,
        "amenities": 20,
        "lifestyle": 20,
        "construction": 5
      }
    },
    "high_rise": {
      "visible_questions": [
        "home_types",
        "county_ids",
        "budget_min",
        "max_monthly_fees",
        "amenities",
        "lifestyle",
        "new_construction",
        "age_choice",
        "notes"
      ],
      "weights": {
        "budget": 30,
        "fees": 25,
        "amenities": 30,
        "lifestyle": 10,
        "construction": 5
      }
    }
  }
}
```

## Record structure — fictional template, never factual evidence
```json
[
  {
    "market_id": "anne-stewart-neighborhood-match",
    "area_id": "REPLACE-AREA",
    "segment_id": "REPLACE-AREA--single_family",
    "name": "REPLACE verified name",
    "home_type": "single_family",
    "county_id": "multnomah",
    "status": "draft",
    "summary": "REPLACE with sourced buyer copy",
    "ownership_type": "REPLACE with sourced ownership",
    "physical_form": "REPLACE with sourced physical form",
    "tradeoffs": [
      "Fictional training only."
    ],
    "evidence": [],
    "price": {
      "kind": "unknown"
    },
    "fees": {
      "status": "unknown"
    },
    "lifestyle": [],
    "amenities": [],
    "construction": {
      "state": "unknown"
    },
    "age_restriction": {
      "status": "unknown"
    },
    "geography": {
      "lat": null,
      "lng": null,
      "scope": "REPLACE",
      "evidence_ids": []
    },
    "idx": {
      "status": "missing"
    },
    "inventory": {
      "status": "unknown"
    }
  }
]
```

## Scenario structure — replace placeholders with independent buyer expectations
```json
{
  "market_id": "anne-stewart-neighborhood-match",
  "cases": [
    {
      "id": "condo-pool-case",
      "kind": "reviewer",
      "buyer": {
        "home_types": [
          "condo"
        ],
        "budget_min": 250000,
        "budget_max": 500000,
        "budget_hard": true,
        "amenities": [
          "pool"
        ],
        "lifestyle": [
          "trails"
        ]
      },
      "expect": {
        "eligible_contains": [
          "REPLACE_WITH_REAL_CONDO_SEGMENT_ID"
        ],
        "eligible_excludes": [
          "REPLACE_WITH_REAL_DETACHED_SEGMENT_ID"
        ]
      }
    },
    {
      "id": "target-quiz-parity",
      "mode": "manual",
      "reason": "Run the same dataset and answers in the pinned destination adapter and preview; compare scores, eligibility, explanations and home-type branches."
    }
  ]
}
```

## Complete fictional examples of field shapes — never copy as market evidence
These records demonstrate supported prices, fees, evidence references and scoped geography. Their IDs, locations and .invalid sources are fictional and belong to training only. Use this job’s config and researched facts for actual output.
```json
[
  {
    "market_id": "training",
    "area_id": "example-a",
    "segment_id": "example-a--single_family",
    "name": "Fictional example-a",
    "home_type": "single_family",
    "county_id": "training-county",
    "status": "draft",
    "summary": "Fictional training record. Not a real neighborhood.",
    "ownership_type": "fee_simple",
    "physical_form": "detached",
    "tradeoffs": [
      "Fictional training only."
    ],
    "evidence": [
      {
        "id": "example-a--single_family-e1",
        "url": "https://example.invalid/training/example-a--single_family",
        "checked_at": "2026-09-11",
        "as_of": "2026-09-11",
        "scope": "example-a--single_family",
        "claim": "FICTIONAL TEST FIXTURE. All facts are invented to test the software; never import or publish."
      }
    ],
    "price": {
      "kind": "typical_range",
      "min": 500000,
      "max": 900000,
      "open_upper": false,
      "method": "synthetic_test_fixture",
      "sample_size": 20,
      "as_of": "2026-09-11",
      "evidence_ids": [
        "example-a--single_family-e1"
      ]
    },
    "fees": {
      "status": "confirmed_zero",
      "min": 0,
      "max": 0,
      "unit": "USD_per_month",
      "total_required_dues": true,
      "evidence_ids": [
        "example-a--single_family-e1"
      ]
    },
    "lifestyle": [
      {
        "key": "trails",
        "state": "present",
        "evidence_ids": [
          "example-a--single_family-e1"
        ]
      }
    ],
    "amenities": [
      {
        "key": "pool",
        "state": "absent",
        "access": "on_site",
        "evidence_ids": [
          "example-a--single_family-e1"
        ]
      }
    ],
    "construction": {
      "state": "absent",
      "evidence_ids": [
        "example-a--single_family-e1"
      ]
    },
    "age_restriction": {
      "status": "unrestricted",
      "evidence_ids": [
        "example-a--single_family-e1"
      ]
    },
    "geography": {
      "lat": 35.9,
      "lng": -84.1,
      "scope": "fictional_test_anchor",
      "evidence_ids": [
        "example-a--single_family-e1"
      ]
    },
    "idx": {
      "status": "missing"
    },
    "inventory": {
      "status": "unknown"
    }
  },
  {
    "market_id": "training",
    "area_id": "example-a",
    "segment_id": "example-a--condo",
    "name": "Fictional example-a",
    "home_type": "condo",
    "county_id": "training-county",
    "status": "draft",
    "summary": "Fictional training record. Not a real neighborhood.",
    "ownership_type": "condominium",
    "physical_form": "attached",
    "tradeoffs": [
      "Fictional training only."
    ],
    "evidence": [
      {
        "id": "example-a--condo-e1",
        "url": "https://example.invalid/training/example-a--condo",
        "checked_at": "2026-09-11",
        "as_of": "2026-09-11",
        "scope": "example-a--condo",
        "claim": "FICTIONAL TEST FIXTURE. All facts are invented to test the software; never import or publish."
      }
    ],
    "price": {
      "kind": "typical_range",
      "min": 200000,
      "max": 350000,
      "open_upper": false,
      "method": "synthetic_test_fixture",
      "sample_size": 20,
      "as_of": "2026-09-11",
      "evidence_ids": [
        "example-a--condo-e1"
      ]
    },
    "fees": {
      "status": "range",
      "min": 500,
      "max": 500,
      "unit": "USD_per_month",
      "total_required_dues": true,
      "evidence_ids": [
        "example-a--condo-e1"
      ]
    },
    "lifestyle": [
      {
        "key": "trails",
        "state": "absent",
        "evidence_ids": [
          "example-a--condo-e1"
        ]
      }
    ],
    "amenities": [
      {
        "key": "pool",
        "state": "present",
        "access": "on_site",
        "evidence_ids": [
          "example-a--condo-e1"
        ]
      }
    ],
    "construction": {
      "state": "absent",
      "evidence_ids": [
        "example-a--condo-e1"
      ]
    },
    "age_restriction": {
      "status": "unrestricted",
      "evidence_ids": [
        "example-a--condo-e1"
      ]
    },
    "geography": {
      "lat": 35.9,
      "lng": -84.1,
      "scope": "fictional_test_anchor",
      "evidence_ids": [
        "example-a--condo-e1"
      ]
    },
    "idx": {
      "status": "missing"
    },
    "inventory": {
      "status": "unknown"
    }
  },
  {
    "market_id": "training",
    "area_id": "example-b",
    "segment_id": "example-b--single_family",
    "name": "Fictional example-b",
    "home_type": "single_family",
    "county_id": "training-county",
    "status": "draft",
    "summary": "Fictional training record. Not a real neighborhood.",
    "ownership_type": "fee_simple",
    "physical_form": "detached",
    "tradeoffs": [
      "Fictional training only."
    ],
    "evidence": [
      {
        "id": "example-b--single_family-e1",
        "url": "https://example.invalid/training/example-b--single_family",
        "checked_at": "2026-09-11",
        "as_of": "2026-09-11",
        "scope": "example-b--single_family",
        "claim": "FICTIONAL TEST FIXTURE. All facts are invented to test the software; never import or publish."
      }
    ],
    "price": {
      "kind": "typical_range",
      "min": 450000,
      "max": 700000,
      "open_upper": false,
      "method": "synthetic_test_fixture",
      "sample_size": 20,
      "as_of": "2026-09-11",
      "evidence_ids": [
        "example-b--single_family-e1"
      ]
    },
    "fees": {
      "status": "range",
      "min": 200,
      "max": 200,
      "unit": "USD_per_month",
      "total_required_dues": true,
      "evidence_ids": [
        "example-b--single_family-e1"
      ]
    },
    "lifestyle": [
      {
        "key": "trails",
        "state": "present",
        "evidence_ids": [
          "example-b--single_family-e1"
        ]
      }
    ],
    "amenities": [
      {
        "key": "pool",
        "state": "present",
        "access": "on_site",
        "evidence_ids": [
          "example-b--single_family-e1"
        ]
      }
    ],
    "construction": {
      "state": "absent",
      "evidence_ids": [
        "example-b--single_family-e1"
      ]
    },
    "age_restriction": {
      "status": "unrestricted",
      "evidence_ids": [
        "example-b--single_family-e1"
      ]
    },
    "geography": {
      "lat": 35.9,
      "lng": -84.1,
      "scope": "fictional_test_anchor",
      "evidence_ids": [
        "example-b--single_family-e1"
      ]
    },
    "idx": {
      "status": "missing"
    },
    "inventory": {
      "status": "unknown"
    }
  },
  {
    "market_id": "training",
    "area_id": "example-c",
    "segment_id": "example-c--condo",
    "name": "Fictional example-c",
    "home_type": "condo",
    "county_id": "training-county",
    "status": "draft",
    "summary": "Fictional training record. Not a real neighborhood.",
    "ownership_type": "condominium",
    "physical_form": "attached",
    "tradeoffs": [
      "Fictional training only."
    ],
    "evidence": [
      {
        "id": "example-c--condo-e1",
        "url": "https://example.invalid/training/example-c--condo",
        "checked_at": "2026-09-11",
        "as_of": "2026-09-11",
        "scope": "example-c--condo",
        "claim": "FICTIONAL TEST FIXTURE. All facts are invented to test the software; never import or publish."
      }
    ],
    "price": {
      "kind": "typical_range",
      "min": 250000,
      "max": 400000,
      "open_upper": false,
      "method": "synthetic_test_fixture",
      "sample_size": 20,
      "as_of": "2026-09-11",
      "evidence_ids": [
        "example-c--condo-e1"
      ]
    },
    "fees": {
      "status": "unknown"
    },
    "lifestyle": [
      {
        "key": "trails",
        "state": "absent",
        "evidence_ids": [
          "example-c--condo-e1"
        ]
      }
    ],
    "amenities": [
      {
        "key": "pool",
        "state": "present",
        "access": "nearby_public",
        "evidence_ids": [
          "example-c--condo-e1"
        ]
      }
    ],
    "construction": {
      "state": "absent",
      "evidence_ids": [
        "example-c--condo-e1"
      ]
    },
    "age_restriction": {
      "status": "unrestricted",
      "evidence_ids": [
        "example-c--condo-e1"
      ]
    },
    "geography": {
      "lat": 35.9,
      "lng": -84.1,
      "scope": "fictional_test_anchor",
      "evidence_ids": [
        "example-c--condo-e1"
      ]
    },
    "idx": {
      "status": "missing"
    },
    "inventory": {
      "status": "unknown"
    }
  },
  {
    "market_id": "training",
    "area_id": "example-d",
    "segment_id": "example-d--townhome",
    "name": "Fictional example-d",
    "home_type": "townhome",
    "county_id": "training-county",
    "status": "draft",
    "summary": "Fictional training record. Not a real neighborhood.",
    "ownership_type": "fee_simple",
    "physical_form": "attached",
    "tradeoffs": [
      "Fictional training only."
    ],
    "evidence": [
      {
        "id": "example-d--townhome-e1",
        "url": "https://example.invalid/training/example-d--townhome",
        "checked_at": "2026-09-11",
        "as_of": "2026-09-11",
        "scope": "example-d--townhome",
        "claim": "FICTIONAL TEST FIXTURE. All facts are invented to test the software; never import or publish."
      }
    ],
    "price": {
      "kind": "typical_range",
      "min": 350000,
      "max": 500000,
      "open_upper": false,
      "method": "synthetic_test_fixture",
      "sample_size": 20,
      "as_of": "2026-09-11",
      "evidence_ids": [
        "example-d--townhome-e1"
      ]
    },
    "fees": {
      "status": "range",
      "min": 250,
      "max": 250,
      "unit": "USD_per_month",
      "total_required_dues": true,
      "evidence_ids": [
        "example-d--townhome-e1"
      ]
    },
    "lifestyle": [
      {
        "key": "trails",
        "state": "present",
        "evidence_ids": [
          "example-d--townhome-e1"
        ]
      }
    ],
    "amenities": [
      {
        "key": "pool",
        "state": "unknown",
        "access": "on_site",
        "evidence_ids": [
          "example-d--townhome-e1"
        ]
      }
    ],
    "construction": {
      "state": "absent",
      "evidence_ids": [
        "example-d--townhome-e1"
      ]
    },
    "age_restriction": {
      "status": "unrestricted",
      "evidence_ids": [
        "example-d--townhome-e1"
      ]
    },
    "geography": {
      "lat": 35.9,
      "lng": -84.1,
      "scope": "fictional_test_anchor",
      "evidence_ids": [
        "example-d--townhome-e1"
      ]
    },
    "idx": {
      "status": "missing"
    },
    "inventory": {
      "status": "unknown"
    }
  },
  {
    "market_id": "training",
    "area_id": "example-luxury",
    "segment_id": "example-luxury--single_family",
    "name": "Fictional example-luxury",
    "home_type": "single_family",
    "county_id": "training-county",
    "status": "draft",
    "summary": "Fictional training record. Not a real neighborhood.",
    "ownership_type": "fee_simple",
    "physical_form": "detached",
    "tradeoffs": [
      "Fictional training only."
    ],
    "evidence": [
      {
        "id": "example-luxury--single_family-e1",
        "url": "https://example.invalid/training/example-luxury--single_family",
        "checked_at": "2026-09-11",
        "as_of": "2026-09-11",
        "scope": "example-luxury--single_family",
        "claim": "FICTIONAL TEST FIXTURE. All facts are invented to test the software; never import or publish."
      }
    ],
    "price": {
      "kind": "typical_range",
      "min": 2000000,
      "max": null,
      "open_upper": true,
      "method": "synthetic_test_fixture",
      "sample_size": 20,
      "as_of": "2026-09-11",
      "evidence_ids": [
        "example-luxury--single_family-e1"
      ]
    },
    "fees": {
      "status": "range",
      "min": 600,
      "max": 600,
      "unit": "USD_per_month",
      "total_required_dues": true,
      "evidence_ids": [
        "example-luxury--single_family-e1"
      ]
    },
    "lifestyle": [
      {
        "key": "trails",
        "state": "present",
        "evidence_ids": [
          "example-luxury--single_family-e1"
        ]
      }
    ],
    "amenities": [
      {
        "key": "pool",
        "state": "absent",
        "access": "on_site",
        "evidence_ids": [
          "example-luxury--single_family-e1"
        ]
      }
    ],
    "construction": {
      "state": "absent",
      "evidence_ids": [
        "example-luxury--single_family-e1"
      ]
    },
    "age_restriction": {
      "status": "unrestricted",
      "evidence_ids": [
        "example-luxury--single_family-e1"
      ]
    },
    "geography": {
      "lat": 35.9,
      "lng": -84.1,
      "scope": "fictional_test_anchor",
      "evidence_ids": [
        "example-luxury--single_family-e1"
      ]
    },
    "idx": {
      "status": "missing"
    },
    "inventory": {
      "status": "unknown"
    }
  }
]
```

## Return format
Return one downloadable UTF-8 project JSON with name, records, config and scenarios, plus a separate coverage/source/change summary. Keep all market_id values aligned with the supplied config. For a selected-area batch, return a clearly labeled partial project; it is not a completed market. Preserve current project data when it is supplied. The operator will import it into Data Engine, run the actual gauntlet and give you the actual resulting research tickets. Do not claim to have run the engine unless you actually did.
Where the prompt goes

Run the downloaded prompt in your separate research assistant with browser/source access.
For a new quiz, import its project JSON into Data Engine. The dataset supplies facts; the configuration supplies questions and home-type weights.
Use Quiz simulator and Gauntlet locally. They require no API key. Research gaps and review weight changes before retesting.
Connect your own key in AI review lab for optional explanations and grading after readiness. The engineer then verifies the new route uses the same exported code and settings.