OGM Scout — Tax Roll & Pipeline

A tax record is a lead and supporting evidence — never title. Lead is an ordering aid (who to look at first), not a value. Classification comes from wells, permits, lease status and the roll itself; open a row to see what the record proves, suggests and does not establish.

Groups are the system's claim that these rows are the same person or entity (same name key, same ZIP). Zip-less rows fold into a single matching group. Reachability repeats what the appraisal district recorded — a mailing address it wrote as UNKNOWN, a name carrying ESTATE or HEIRS — not a verified finding. A tax roll is a lead, never title evidence. Nothing on this screen establishes ownership; "Find them" opens a public search in a new tab and sends nothing anywhere until you click it.

Five triggers put a tract here: a rig drilling now, a recent spud, a recently filed permit, a tract you flagged as an area of interest, or an owner the tax roll flags. Priority is the sum of the triggers that fired — a tract with a rig and a recent spud outranks either alone. A trigger is a reason to research the tract, never a finding about it. Nothing on this screen establishes ownership. The report itself is produced by the Preliminary Report Bot under its landman ruleset and imported through prelim_api.php; mark a tract reported with its batch UID to link the two.

Title confidence is measured, not assumed: it is built from the starting point, whether the chain walks and balances, whether every grantor appears as a prior grantee, the open defects, and how current the search is. Any item that turns on a question of law caps it at 0.75 and is marked for an attorney. The mineral chain itself is walked by the agent layer (ogm_agents); where that is not deployed the walk reports unavailable and ownership stays unknown rather than being guessed at here.

A lease status is never a bare "leased". Past the primary term the question is what holds it — production in paying quantities, a shut-in payment, or a continuous-development programme — and a Pugh or retained-acreage clause can leave it alive on part of the acreage and dead on the rest. What has come open is listed as reverted, and what would settle the question is listed under needs.

The decline fit starts at the peak month, not the first, and months with no reported production are excluded as downtime rather than read as decline. The b‑factor is capped (SCOUT_B_CAP) because an uncapped hyperbolic tail is how an EUR runs away; where the cap binds, the well says so. These are labels from the evidence loaded, not a certified reserve report — a booked estimate needs a qualified petroleum engineer.

Two verdicts are shown, never merged: Queue is the platform's BUY/NEGOTIATE/WATCH/PASS band (and an analyst's override wins there), Decision is the ten-class call from the spec rubric. Covered is how much of the scoring frame could actually be measured — a category with no data leaves the numerator and the denominator rather than scoring zero, so a high score on thin coverage is visible as thin. An acquire_* decision needs three independent positive signals from different evidence families; short of that it is demoted, and the demotion is shown.

A decimal that disagrees is a question, not an accusation: it is usually a different unit factor, a participation the operator computed on other acreage, or an interest sold after the examination date. Each finding carries what would settle it. Revenue expectations use the platform price deck rather than the operator's realised prices, so a variance is a prompt to compare realisations and deductions — never a computed amount owed.

Six deterministic critics run on every target, cost nothing and catch most of what a bad row can do. Two model critics run only on what survives them still holding BUY or NEGOTIATE, and are shown a fact sheet with every conclusion stripped out — shown the answer, a critic anchors on it and rubber-stamps, and you have paid for a second opinion and received a second copy of the first one. A critic can only demote, cap or flag: there is no path in the chain that improves a recommendation. The ones it kills are usually the more interesting read.

An alert is not "something changed" — the change feeds already record every change. It is a change that alters what to do about a tract, and each one carries what changed, why it matters and the next move. Every alert is keyed to the fact behind it, so a nightly run refreshes a standing alert rather than raising it again; a monitor that fires every night gets muted, and a muted monitor is worse than none. An alert on a tract already in the pipeline opens a task on that opportunity, so it lands where the work is.

The loop: what the engine said when the opportunity opened (frozen in a prediction snapshot) against what actually happened. Fewer than five outcomes is anecdote. This table reports; it does not move any knob by itself.