# Evidence-first recovery pilot
Created: 2026-09-18

## Objective and decision

Determine whether James's existing unwanted contacts contain a small number of well-supported, economically sensible cases, then automate the repeatable clerical work. Research supports a pilot, not an income promise or unattended litigation.

Start with one receiving number and up to 20 candidate conversations, initially from the most recent 12 months. Preserve older relevant material too; this sample window is not a limitations rule. Prepare at most three packets. Stop pilot preparation at five hours of James's time or 30 calendar days after evidence becomes available, whichever comes first. These are proposed operating caps, not permission to spend money. No paid tools, legal fees, filing fees, or external communications are authorized by this plan.

## Work checklist and acceptance

- [x] Research federal routes, current developments, reuse options, and evidence requirements.
- [x] Create preservation, intake, review, and outcome templates.
- [x] Check available Messages access; document failure to obtain an evidence export.
- [x] Receive state/county, personal-use status, and approximate number tenure; preserve as user-reported private intake.
- [ ] Confirm residence/contact dates, exact number ownership dates, registry proof, existing consent, and prior disputes.
- [ ] Screen recent contacts immediately for the Texas §304.257 notice/verified-complaint prerequisites; the preparation cap must not delay a relevant deadline.
- [ ] Preserve an approved sample and verify source fidelity.
- [ ] Review up to 20 threads, with written reasons for candidate, exclude, or unresolved.
- [ ] Prepare up to three packets with exhibits and strongest counterarguments.
- [ ] Obtain jurisdiction-specific review of the best packet and an actual cost proposal.
- [ ] James decides whether to authorize a specific demand, report, or filing based on the finished packet.
- [ ] Track results; decide whether automation is worth building.

The preparation pilot succeeds if it produces at least one fully traceable packet suitable for legal review within the cap, or a documented no-go explaining why none qualifies. This is a process milestone, not proof of a viable income stream. Recovery validation requires actual received funds and measured costs/time; unresolved cases remain pending, not wins.

## Sequence

| Phase | Actions and owner | Exit evidence |
|---|---|---|
| 1. Intake | James supplies jurisdiction, use/ownership and consent facts; agent records unknowns without guessing | Completed private intake and registry evidence |
| 2. Preserve | James/authorized local tool captures complete threads and call material; agent manifests and checks copies | Source-matched sample, hashes, exceptions log |
| 3. Triage | Agent proposes categories and entity leads; James confirms recollections; all candidates reviewed | Candidate ledger, defense notes, excluded/unresolved cases retained |
| 4. Attribute | Agent researches public entity records and independent business sources; reviewer separates sender from advertised seller | Identity evidence and service-address candidates with confidence and contradictions |
| 5. Prepare | Agent assembles exhibits, timeline, claim theories, and questions; lawyer/reviewer validates local law | At most three complete packets, no invented facts or damages totals |
| 6. Escalate | After separate authorization, prepare exact recipient/content/action for review; James or counsel handles approved legal acts | Sent-document copy, authorization record, delivery/filing evidence if applicable |
| 7. Measure | Agent logs labor/cost/receipts; James evaluates whether to continue | Outcome ledger with pending cases and actual cash distinguished |

Before Phase 6, resolve jurisdiction/standing, consent and exceptions, defendant attribution, limitations, damages counting, proper service, any arbitration/release, and review cost. Counsel involvement is recommended for validation. Do not wait for new violations. Notice and pre-suit requirements are theory-specific; apply the [Texas supplement](../research/texas-supplement.md), including §304.257, before adopting a timetable.

## Triage policy

- Prioritize repeated apparent marketing from an identifiable operating business, with intact context and a plausible path to service and collection.
- Place unclear consent, mixed use, uncertain attribution, ambiguous transport, and incomplete history in `needs_review`.
- Keep transactional/political/charitable/scam classifications separate. Exclusion from the pilot is not a legal determination that a contact was lawful.
- A single message may be relevant to a different legal route even if it cannot satisfy the repeat-contact DNC theory alone.
- Retain rejected cases and reasons so the project can measure false positives rather than presenting only favorable examples.

## Economics: measure money, not claimed damages

Primary outcome: cash actually received minus attributable out-of-pocket costs. Also report that amount divided by all James hours, including intake, review, correspondence, court, and collection. Keep build effort visible separately and show an all-in figure as well. If hours are zero/unknown, the rate is undefined, not infinite.

Before recovery, use scenario analysis only:

`scenario net value = assumed probability of cash recovery × assumed net receipt if recovered − expected unreimbursed costs − James hours × chosen hourly value`

All inputs start `unknown`; do not invent success probabilities. Hypothetical arithmetic: 20% × $1,000 minus $100 costs minus 3 hours × $50/hour = −$50. This is not a forecast. Do not subtract contingency fees twice if the receipt input is already net of fees.

Track: candidates reviewed, defensible packets, human reversals, minutes per packet, expenses, elapsed days, written settlements, cash received, pending cases, and optional unwanted-contact frequency. Never optimize for number of demands, maximum alleged damages, or more incoming spam.

## Agentic Version

| Tier | Allowed autonomy | Explicit boundary |
|---|---|---|
| Simple automation | Authorized local exports, manifest/hash checks, duplicate detection, chronology, blank packet assembly | No legal conclusions or outbound actions |
| Semi-agentic — target after pilot | Local classification, evidence-linked entity leads, missing-fact questions, draft packets, outcome tracking | Human verifies facts and law; action authorization is separate and tied to exact content/recipient |
| Fully agentic internal preparation — later | Incremental approved imports, changed-evidence detection, refreshed internal drafts, meaningful-change alerts | Fully autonomous legal demands/filings/settlements are out of scope; the agent stops at review-ready |

Five preconditions for recurring operation:

1. **Objective metric:** realized net recovery per James hour, with evidence correctness as a hard constraint. Pre-recovery proxy: reviewed packet-preparation minutes; it cannot establish profitability.
2. **Measurement:** timestamped work/outcome ledger, receipts, expense records, source-linked review decisions. Legal-quality judgments need a qualified reviewer; a model's self-score is not validation.
3. **Action space:** import approved evidence, flag gaps, organize exhibits, rank for review, draft, and update internal status. No silent expansion of collection scope.
4. **Stopping condition:** pilot caps above. For later scheduled operation, propose one daily local pass capped at 100 new messages/10 minutes and pause on backlog; do not schedule until manually validated and authorized. No paid API calls by default.
5. **Rollback/guardrail:** originals immutable by policy with backup; version derived records and rules. Any source mismatch, invented assertion, privacy leak, or unauthorized outbound attempt stops the run. Preserve logs, revert derived output to last reviewed version, and require manual restart after correction.

**Missing now:** preserved corpus, verified legal facts/jurisdiction, recovery observations, validated importer, reviewer acceptance, and a configured private runtime. Unattended operation is not ready.

## Automation roadmap after the pilot

Use existing export tooling plus local structured storage before adding a UI or a long-running agent. The named reason for custom code is integration of James's own evidence, review workflow, and realized outcomes.

1. **Read-only importer:** normalize approved exports into events; source IDs/hashes, completeness diagnostics, stable deduplication, original/local/UTC timestamps. Test malformed input, duplicate reruns, outgoing opt-outs, blank bodies, edits/reactions, attachments, multiple destination numbers, and DST. Pin the exporter and fixture contract.
2. **Review queue:** facts separate from findings; every extracted assertion points to source evidence. Local model suggestions only. Identity and consent remain unresolved unless supported. Treat all message/website text as untrusted data. No tool execution, external fetching, or outbound permissions in the classification process.
3. **Packet builder:** static local reports with exhibit index, timeline, adverse evidence, missing facts, cited rule versions, and reviewer sign-off. No default dollar demand. Disable live links/assets; redact shared copies, retain private originals.
4. **Outcome ledger:** distinguish alleged exposure, negotiated amount, judgment, collection, costs, and received funds. Store approvals by document hash + recipient + purpose + date; changed content invalidates approval.
5. **Only then recurrence:** run manually first, verify idempotency and privacy, then consider a quiet monitor that reports meaningful changes or failures. No monitor has been created.

Future implementation needs its own software implementation plan and tests. This document is a research/pilot roadmap, not a ready-to-run software specification.

## Validation and decision gates

- Every material statement in a packet can be traced to an exhibit; unsupported statements remain questions.
- A sample import is reconciled against source UI and any discrepancies stop packet generation.
- Deliberate synthetic counterexamples (consented marketing, duplicate export, forged seller identity, personal message, prompt injection) do not become automatically approved claims.
- No model output can transition a case to send/file/settle authority. Verify this at the action boundary, not just in prompts.
- State-specific legal review is complete before an actionable legal theory or final deadline is adopted.
- At the preparation cap, proceed only if a credible packet exists; otherwise document no-go or targeted missing evidence. Do not build an app to compensate for lack of viable cases.
- Expand after a reviewed recovery attempt only if measured burden and likely economics justify it; repeat the calculation with actual receipts when available.