Migration + Convergence
System Dossier

1. Executive summary

Source claim, not independently verified by this demo

The central finding. The migration did not reveal a small application waiting to be deployed. It exposed a partially built operating organism whose major problem was continuity: multiple databases, duplicated logic, stale runtime paths, unproven timestamps, uncustodied local work, and automation that could stop at exactly the seam where the next action needed to happen. The work since migration began has therefore been less like a database move and more like systems archaeology followed by reconstitution.

The strongest asset is not any one DeedSonar screen, property score, recipe, or agent. It is the emerging control architecture that makes data, decisions, software changes, runtime state, and external actions traceable to evidence and bounded by authority. For every material state transition it increasingly answers five questions: what is known, how do we know it, how fresh is it, what can legally or operationally happen next, and what exact proof closes the step?

What exists now (source claims). Canonical production authority sits on the efxo target while the older lvdi spine is legacy/forensic and explicitly refused by key writers. DeedSonar has a property identity lattice, multi-recipe membership, signal fusion, next-best-evidence, qualification, demand matching, contact-action authority, receipts, courts, freshness semantics, and production-proof harnesses. LandOS ingestion and engine converged around source custody, facts-as-of discipline, cheap-source unlocks, bounded/resumable CRM projection, and explicit legacy-spine refusal. The remaining gap is live activation and proof of the final event-driven loop across runtime, receipt watcher, Sentinel verdict ingest, Git custody, deployment, and commercial outcome recording.

2. What this dossier is actually measuring

A conventional migration report asks whether rows moved. That standard is insufficient here. The system has multiple layers of truth: source custody, normalized facts, canonical property identity, recipe membership, scoring, qualification, contactability, action authority, software custody, runtime deployment, and commercial outcome. A migrated row can be present yet semantically wrong, fresh-looking yet stale, callable-looking yet unauthorized, or deployed yet pointed at the wrong database.

Four simultaneous objectives: data convergence (one canonical production authority, legacy state reconciled without destroying forensic lineage); semantic convergence (duplicated functions and vocabularies agree across DeedSonar, LandOS, ingestion, recipes, CRM, and contact layers); execution convergence (workers, schedulers, receipts, courts, Git custody, runtime verification, and outcome handling reconnected); commercial convergence (machinery turned into customer-visible intelligence products).

Evidence standard used in the dossier

3. The convergence timeline

4. What the migration exposed

Eight representative forensic discoveries, each a failure of authority, identity, freshness, custody, or continuity that would have caused silent corruption or false confidence in a conventional stack.

Two database realities. Several tools and environment variables still pointed toward lvdi, creating the possibility of a second CRM writer or split-brain execution. LandOS-engine PR #37 converged the engine onto efxo; PR #39 hardened that decision by refusing the legacy spine at process start. The contact graph received an even stricter resolver in PR #30: contact tools require the explicit canonical variable and reject lvdi with no override.

Data-only restoration erased structural guarantees. LandOS-engine PR #40 found the efxo landos schema had been copied as data without the legacy post-data structural layer: 0 of 31 tables had primary keys, unique/foreign keys, or indexes. On-conflict writers consequently failed. Two tables were triple-loaded. Use the calculator in section 9 to verify the ratios.

Identity logic was more dangerous than ranking logic. The curative import found real cases where an APN without FIPS, a placeholder situs, raw owner strings, or loose substring matching could collapse different properties or classify people as entities. LandOS-ingest PR #11 refuses unresolved county identity and placeholder situs; PR #28 corrected owner-kind classification; PR #29 corrected entity fallback and address normalization. DeedSonar PR #63 formalized one canonical opportunity per property and proved TS/SQL property-key parity on all 20,734 NC opportunities.

Seemingly reasonable facts were semantically unsafe. A tax balance cannot be treated as mortgage-cleared equity. The normalized fact became an upper bound. The same work distinguished a disjunctive vacancy-or-code tag from independent vacancy and code facts. Signal fusion in PR #64 adopted KNOWN / UNKNOWN / CONFLICTED semantics so disagreement is preserved instead of flattened.

Automation had silent skips. Routes that could not apply to a row produced no explicit outcome. PR #19 made those routes produce receipts or named halts; PR #21 made inapplicable routes reasoned NO_HIT states. "Could not run" is now data instead of absence.

5. The DeedSonar / LandOS intelligence stack

Source claims, not independently verified by this demo

The 20,734-property intelligence core. The NC estate is the primary proving ground. DeedSonar PR #63 established 20,734 canonical opportunities with 0 false merges, false splits, unknown identity collisions, or cross-county conflicts in the curative import. PR #64 fused signals across ten evidence families and reproduced the receiver partition across all 20,734 properties. Seven faulty recipe clauses and evaluator behaviors were exposed in the process.

Multi-recipe intelligence. PR #57 created the membership layer needed for secondary opportunities. The rehearsal reached 21,527 memberships across the 20,734 properties, with 641 secondary eligible memberships beyond the seed set.

Next-best-evidence as an intelligence primitive. PR #62 identifies the next evidence action, estimates information value, cost, and latency, and records a research receipt. On a read-only production pass, every property received a continue decision with a lawful next visit or an explicit operator-only path.

Cheap-source unlock philosophy. Exhaust existing lawful, free, already-custodied signals before buying a new source. The source-unlock program exposed valuation, parcel, code/vacancy, tax-year thresholds, parcel graph, tax-sale calendar, deed tenure, and federal hazard joins without creating unnecessary database seams or new paid acquisitions. The system also records when an official source is unavailable or incomplete rather than pretending the gap does not exist.

Source factory and custody. LandOS-ingest PR #14 bound each run receipt to the code commit that produced the rows. PR #15 made ArcGIS pulls prove completeness against server-side counts. PR #16 and #17 added Union and Mecklenburg coverage with explicit handling for taxpayer-level staging where parcel identity does not exist in the source. PR #13 corrected a placeholder ZIP problem that had fabricated roughly 2,200 positive absentee-owner readings.

6. The autonomous organism

What makes the autonomy model different. Mission state is designed to survive model/process death. Receipts are transition objects recording what happened, why, inputs, outputs, authority, as-of state, uncertainty, and proof. Independent courts separate production work from the decision to accept that work. Git custody is a control surface: an independently passed feature branch can be pushed and turned into a PR, while direct main writes and production deployment remain separately authorized. Action authority is persisted: contact permission is a state that can expire, be revoked, or be invalidated. Queues and routes are evidence-bearing: a source or route that cannot apply becomes an explicit NO_HIT, HALTED, or operator step instead of disappearing.

The customer tape is the proof instrument. DeedSonar PR #65 ran one real production opportunity through a 15-step PROPERTY to RECEIPT path. The run passed the identity, evidence, qualification, and disqualifier gates, then stopped at CONTACT_LADDER because no living party had been established. The system discovered the next missing capability, repaired the missing operator-research writer, and then re-ran to find the next missing capability: the contact projection had not yet been run against production. A conventional demo hides the missing capability. A governed execution tape discovers it, records it, routes the defect, repairs the seam, and resumes.

7. Governance, courts, and the refusal model

The system is not optimized to maximize autonomous action. It is optimized to maximize correct, explainable progress. The result is a system that can refuse to act when evidence, authority, freshness, or provider state is inadequate.

Action authority. DeedSonar PR #71 adds an append-only contact_action_authority decision per candidate x number x channel, with explicit states NOT_GRANTED, AUTHORIZED, DENIED, REVOKED, EXPIRED, and INVALIDATED. The authorization model requires an entitled administrator, a basis, an attestation, and an expiry bounded at 180 days. This is stronger than embedding a boolean "callable" field in a CRM table. Use the authority expiry calculator in section 9.

The final rung remains intentionally gated. The read-side authority model is present, but live dial/SMS action has not been treated as automatically enabled. The customer tape blocked at CONTACT_LADDER because the candidate had no established living party and 0 seller-side numbers. This is precisely the behavior a serious system should exhibit. The product should monetize higher certainty, not higher recklessness.

8. The migration was a reconstruction, not a copy

Production authority. The canonical DeedSonar target is Supabase project efxowqudmsfqjmbeidjc. The legacy lvdi project lvdimujkfheeyqchjzra remains valuable as forensic/parity evidence but is not treated as a write target. PR #72 established byte-level migration parity and reconstruction courts. PR #74 codified the operating rule: efxo is production; lvdi is legacy; merge is deploy; database writes and migrations are founder-gated.

Custody and restore. The production backup work fixed a subtle but serious failure in the existing verification path. A valid 3,629-line production TOC was being interpreted as 0/3 markers because grep exited early, causing pipefail to report failure and the script to delete the dump. PR #36 fixed the verifier and then produced a real 166 MB efxo dump to cold storage on the external NEURUH-HD volume. A disposable PostgreSQL 17 restore matched production across the major DeedSonar counts, including 70,834 lead_import_rows, 33,079 acquisition leads, 21,527 memberships, 20,734 canonical opportunities, 11,903 contact evidence rows, 6,439 parties, 923 party phones, and a 176-row migration ledger.

Estate custody. PR #66 repaired another custody gap: a sealed estate could have no receipt head, no county import receipts, and only 1,609 of 22,343 payload digests. The fix made the estate receipt content-bound and removed the unsafe bulk-approve behavior. "Sealed" without proof of what was sealed is an attractive nuisance.

Freshness. PR #76 demonstrated the danger in a partial facts estate. Only 4 of 22,343 rows had a real facts date of 2026-08-19, while the majority remained unknown. The product face was therefore changed to say FACTS UNKNOWN with explicit coverage instead of presenting Aug 19 as if it described the entire estate. Use the freshness calculator in section 9 to reproduce the coverage figure.

Customer tape and the final production reality. PR #82 ran the customer tape read-only and stopped at CONTACT_LADDER with HUMAN_STEP_REQUIRED. Candidate 7421053a had a deceased owner, no public estate case, 0 seller-side numbers, and 0 authorized contact actions. The run performed no dial, SMS, or webhook. Autonomy reaches a boundary, proves the boundary, and requests the smallest missing human step.

9. Current state as of September 14, 2026

This table deliberately separates source custody, merged code, provider state, and live behavioral proof. A system can be excellent while still having a short list of live activation gates.

10. What is actually innovative here

Many individual ingredients exist elsewhere: vector search, rules, ETL, queues, LLM agents, property data, CRMs, workflow engines, and CI/CD. The innovation is in the composition and the refusal boundaries.

10.1 Evidence is an execution primitive, not just a citation

A source artifact is not simply attached to a record. It can generate a custodied run, a hash, a normalized signal, a fact date, an uncertainty state, a next-best-evidence decision, and eventually an action-authority transition. Evidence becomes something the machine can route and verify.

10.2 Uncertainty is preserved all the way down

The stack repeatedly refuses a binary worldview. UNKNOWN is not FALSE. CONFLICTED is not LATEST. A partial estate is not CURRENT. A phone is not CALLABLE. A tax balance is not mortgage-cleared equity. A taxpayer-level list is not a parcel. A buyer match is not qualification.

10.3 Autonomous agents operate inside courts

Workers create changes, but independent courts decide whether those changes qualify for promotion. That is a different control model from "let the agent edit production."

10.4 The system can monetize research itself

Next-best-evidence turns uncertainty into a productizable work queue. A customer can buy an opportunity dossier, a research sprint, a forensic audit, a lead rescue, or a managed opportunity desk because the system can quantify what is known and what evidence would materially change the decision.

10.5 Source reuse creates an economic moat

The source-unlock discipline repeatedly extracts more value from data already in custody rather than purchasing more feeds. One artifact can support multiple recipes and evidence families. This is data economics rather than data accumulation.

10.6 The system is self-describing at the transition level

Receipts, source hashes, Git SHAs, migration ledgers, run checkpoints, authority states, and court results collectively form a machine-readable history of why the organism is in its current state. This is closer to an operational memory substrate than a conventional application log.

11. Defensibility and moat

The real moat is not secrecy alone. Some components are reproducible. The moat comes from the accumulation of contracts, edge cases, evidence semantics, source adapters, court patterns, runtime controls, and real-world outcomes. The longer the organism runs, the more expensive it becomes to reproduce the exact history that teaches it where real workflows fail.

12. Turning the infrastructure into revenue

The commercial mistake would be selling "the platform" before customers can feel the economic result. The correct sequence is to expose narrow, high-value outputs while using the same substrate underneath.

Why the dossier is more valuable than a lead list. A lead list transfers research cost to the buyer. A dossier absorbs research cost and returns a decision. That is the monetizable edge. A $499-$2,500 evidence-bound report (source range) can be a front door to recurring pipeline, a $5,000+ forensic engagement, or a managed desk. The architecture underneath supports all three without changing the core data model.

13. Neuruh.com proof strategy

The strongest visual demo. Take one property and show the actual state progression: source artifact, canonical property, fused evidence, recipe memberships, qualification gates, next-best-evidence, demand match, contact ladder, action authority, receipt. The stop condition should be visible when action is not authorized.

14. Remaining gates and risk register

15. Investor and sales narrative

The problem. The market has an abundance of AI that can summarize, search, score, draft, and recommend. Businesses still pay humans to reconcile conflicting records, determine what is current, decide whether an action is allowed, chase missing evidence, move work between systems, verify that changes really deployed, and figure out what happened afterward.

The insight. The bottleneck is not model intelligence alone. The bottleneck is trusted execution. Real-world automation fails at boundaries: stale data, ambiguous identity, conflicting sources, permission gaps, provider drift, hidden side effects, and work that lives only inside a chat or terminal.

The product. Neuruh supplies a governed execution layer that makes those boundaries explicit. It turns evidence into state, state into bounded decisions, decisions into authorized actions, and actions into receipts. DeedSonar is the first high-value proving ground because real estate contains unusually rich combinations of public data, time sensitivity, identity ambiguity, legal authority, geographic context, and economic consequence.

The proof. During convergence, the system did not merely add features. It found and repaired real failure modes: a migration ledger that could hide schema changes, a data-only restore missing constraints, false dates covering tens of thousands of rows, property identity collision paths, unsafe equity inference, silent source-route skips, a wrong-database writer risk, and a backup verifier that would delete valid production dumps. It also demonstrated a receipted customer tape that stopped rather than guessing when authority was missing.

The expansion. Once the execution substrate is stable, the same architecture can operate across any domain with structured evidence and bounded action: insurance claims, equipment resale, procurement, title/escrow preparation, due diligence, collections, compliance research, logistics exceptions, and other decision-heavy workflows.

Appendix A. GitHub evidence ledger explorer

High-value evidence inventory reviewed for the dossier, summarized rather than dumping PR bodies. Filter by repository.

Appendix B. Notion canonical evidence map

Appendix C. State vocabulary

Appendix D. What to say vs what not to say

Appendix E. Bottom line

The asset. The enduring asset is the architecture that connects evidence, state, authority, software custody, runtime verification, and commercial outcomes. DeedSonar demonstrates the value of the substrate; Neuruh is the substrate.

The migration has materially changed the quality of the system. It converted hidden divergence into explicit state, turned several dangerous assumptions into hard refusals, and connected a real-world intelligence graph to an emerging governed execution loop. The remaining work is now sharply bounded: activate and prove the live event chain, finish the founder-gated production repairs, close the contact/action lane safely, and push the first commercial outcome loop hard enough to create real calibration data.

Source note: Human-readable source identifiers correspond to GitHub repository/PR records and Notion canonical pages reviewed during preparation. Current claims are bounded by the evidence available as of September 14, 2026; architecture pages are not treated as substitutes for live provider state.

16. Convergence calculators

Working calculators where the source defines the math. Every figure below reproduces a number stated in the dossier; change an input and the verdict recomputes.

A. Facts-freshness coverage

PR #76: only 4 of 22,343 rows had a real facts date of 2026-08-19. Coverage = known / total x 100. Unknown rows stay UNKNOWN.

B. Triple-load detector

PR #40: audit_log held 27,636 rows against an expected 9,212; comps held 1,371 against an expected 457. Ratio = actual / expected. A ratio of 3 means the table was triple-loaded.

C. Recipe membership density

PR #57 rehearsal: 21,527 memberships across 20,734 properties, with 641 secondary eligible memberships beyond the seed set. Average memberships per property = memberships / properties. Extra memberships = memberships minus properties.

D. Contact authority expiry

PR #71: authorization requires an entitled administrator, a basis, an attestation, and an expiry bounded at 180 days. Expiry = grant date + days (maximum 180).

Sample loaders

Load the dossier's own recorded cases into all four calculators at once.

17. Dossier completeness meter

Tick the fields this dossier record actually carries. The meter counts completed fields against the eight required fields for a convergence dossier record. Anything unticked is listed as missing, and the record is not treated as complete until all eight are present.

18. Share this dossier state

The current calculator inputs and completeness ticks are encoded into the page URL as a base64 JSON hash (#d=). Opening a shared link restores the same state. Nothing is sent to a server; the state lives only in the link.

19. Try-it walkthrough

Five steps through one convergence case, using the dossier's own numbers. Click each step to advance.

20. How it works / What this is not

How it works

This page is a faithful standalone demo of the 19-page source dossier. The section structure, tables, figures, and numbers follow the source. The four calculators reproduce the arithmetic the source defines: freshness coverage, load ratio, membership density, and the 180-day authority expiry. The two explorers filter the PR ledger by repository and the state table by capability state. The completeness meter scores a dossier record against eight required fields. The share link encodes the current calculator and completeness state in the URL hash and restores it on load.

What this is not

• Not a live system. Nothing here reads GitHub, Notion, Supabase, or any production database. All figures are source claims from 14 Sep 2026.
• Not proof of current state. The dossier is bounded by the evidence available as of 14 Sep 2026. Architecture pages are not substitutes for live provider state.
• Not an autonomous execution system. The event-driven receipt to Sentinel to Git loop was an activation gate at the time of writing, and end-to-end autonomous commercial execution was not yet fully proven.
• Not a lead list. The 20,734 canonical NC opportunities are a source claim about property identity parity, not 20,734 perfect leads.
• Not a contact tool. No dial, SMS, or webhook is performed here, and the customer tape example deliberately stops at CONTACT_LADDER.

21. Honest gaps

• Stale evidence. Every claim is bounded by the evidence available as of 14 Sep 2026. Nothing on this page verifies the current state of DeedSonar, LandOS, or the efxo target.
• Activation gate open. The event-driven receipt to Sentinel to verdict ingest to Git custody to deployment loop had no live binding proof at review time. End-to-end autonomous commercial execution was not yet fully proven.
• Founder-gated DDL. Structural parity repair for the 31 landos tables (PR #40) remained founder-apply where DDL is required. Merged code is not the same as an applied database change.
• Contact lane held. The authority read model is live but live dial/SMS action remains separately gated. Candidate 7421053a had 0 seller-side numbers and 0 authorized contact actions.
• Estate freshness mostly unknown. Only 4 of 22,343 rows carried a real facts date. The majority of the estate remained facts-UNKNOWN at review.
• Demand matching limited. Schema and mutations were proven, but the production run was limited by permissions.
• No live verification in this demo. PR statuses, counts, and dates are transcribed from the source PDF. No GitHub or Notion call was made to re-verify them.
• Layout not eyeballed in a real browser. Controls were tested in jsdom only; visual layout was not reviewed in a live browser.

Source: NEURUH_MIGRATION_CONVERGENCE_DOSSIER_2026-09-14.pdf (19 pages, evidence window 5 Aug - 14 Sep 2026). Related: Neuruh Hub.