skills/used-car-investigator/references/research-budget.md

# Research Budget and Stop Rules

Use this router before browsing. The goal is decision completeness, not source accumulation.

## Research stages

### Stage A — Identity and listing

In one batch where possible, resolve the listing, exact powertrain, manufacturer specifications, fuel, original warranty, and seller terms. Stop when the exact architecture is resolved or the missing VIN/build-sheet fact is identified.

### Stage B — Expensive risks and maintenance

Search only for failure modes capable of changing the score, maximum price, PPI, or walk-away decision. Prefer manufacturer campaigns, warranty extensions, recalls, technical bulletins, and technically specific specialist sources.

Do not run separate searches for minor nuisances. Consolidate engine, transmission, AWD/battery, and maintenance queries by exact architecture.

### Stage C — Price and transaction

Batch local comparable searches with dealership/regulator/review searches. For Standard mode, stop after:

- three close local comps, or a clearly disclosed lack of them
- the written advertised fee/warranty/PPI terms
- one regulator/licensing check where available
- one broad review search; corroborate only a material negative theme

Use one additional query only when the first dealer search reveals a material concern requiring confirmation.

## Budgets

| Mode | Search queries | Opened pages | Dealer research | Comparable target |
|---|---:|---:|---:|---:|
| Rapid | 5 | 4 | Listing terms only; one reputation query if material | 2–3 |
| Standard | 10 | 8 | One regulator + one reputation search; corroborate concerns only | 3–5 |
| Deep | 18 | 12 | Regulator, review themes, complaints/news if signalled | 5–8 |

Search queries should be batched into as few tool calls as practical. Open only pages needed to confirm decision-changing claims or details not present in snippets.

## Evidence packet

Keep at most 14 rows in the working evidence packet. Rank candidates by:

1. safety/title/odometer gate
2. catastrophic mechanical exposure
3. deterministic near-term cost
4. price/value impact
5. transaction recourse
6. driving/practical fit

Discard low-impact trivia unless the buyer asked about it.

## Stop conditions

Stop researching when additional sources are unlikely to change:

- the primary verdict
- confidence level
- fair-value range or maximum price
- PPI priorities
- dealer concern rating

If evidence conflicts, report the conflict and lower confidence instead of searching indefinitely.

## Token discipline

- Extract one concise note per source; do not copy source prose.
- Reuse the evidence packet across every section.
- Group citations supporting the same claim.
- Prefer tables for repeated fields and one-sentence interpretations.
- Do not duplicate risks in “what matters,” subsystem tables, value, and bottom line.
- Omit sections that contain no decision-changing information.