
On this page
Every field in the Terrabillion queue carries two attachments: a confidence score and a pointer to its source. Low-confidence fields queue for human review; high-confidence fields ship with spot checks. Nothing in between gets silently rounded up. This post describes the labeling system — what the scores mean, how they are set, and why the gaps stay visible instead of being paved over.
The three bands
Confidence resolves into three bands, and the bands drive behavior. High confidence (0.85 and above) means the value came from agreement between independent sources or from a pattern-matched field with a long clean history — a due date in the standard ODNR header block, a well count confirmed by both scope table and plat. These ship immediately and enter the spot-check sample. Medium (0.6–0.85) means a single credible source or agreement with a caveat, like a sign-flipped coordinate we auto-corrected. These ship flagged, with the caveat attached. Low (below 0.6) means the pipeline could not defend the value, and these never ship unreviewed — they wait for an analyst or appear explicitly marked as unverified.
The thresholds are arbitrary; the discipline is not. What matters is that every value lands in a band with consequences, and that the consequences are the same whether the packet came from a clean digital PDF or a third-generation photocopy.
Where scores come from
Scores are computed, not guessed, and each extraction path has its own scoring logic. Rule-based extraction — dates, amounts, contact blocks — scores from pattern specificity and historical precision: a date matching the county's standard header format in the expected position scores higher than the same date found in body text. Model-assisted extraction — scope narratives, location descriptions — scores from a combination of the model's own assessment and agreement checks against the structured fields: a scope summary mentioning six wells when the table lists six scores higher than one mentioning an unconfirmed count.
Geocoding has the most elaborate ladder, because locations arrive in three mutually disagreeing dialects and graduate only on agreement. The principle is identical everywhere else: independent agreement raises confidence, single sources cap it, and disagreement routes to humans with both sources attached.
Why gaps stay visible
The tempting alternative is gap-filling: when the pipeline cannot extract a value, guess quietly and move on. We do the opposite — the field appears with its low score and its source pointer, visibly marked. Three reasons. First, bidders verify what they can see is weak and trust what they can see is strong; hidden guesses get trusted uniformly, which is how estimators end up pricing phantom wells. Second, visible gaps generate corrections, and corrections are the training signal that improves the pipeline — our launch-week experience showed agency staff correcting fastest precisely because they could see what needed fixing. Third, honesty about uncertainty is the only policy that survives being wrong in public, which happens weekly at our volume.
Calibrating against reality
Scores mean nothing without calibration, so we calibrate weekly. The spot-check sample — fixed rate, stratified across counties and packet types — produces a precision number per band per week, and drift triggers investigation: if high-band precision slips below its floor, we tighten thresholds or fix the extraction path before the next batch ships. Quarterly, we publish the methodology internally and review whether the bands still match reviewer experience. They have moved twice: once when plat reconciliation improved enough to promote a class of locations, and once when a new county's packet dialect humbled a pattern we thought was universal. Calibration answers how accurate we are; our coverage methodology answers how complete.
What we think at Terrabillion
Data quality is a user interface problem. The choice is never between perfect data and flawed data — it is between flawed data with visible seams and flawed data with hidden ones. We choose visible seams everywhere: confidence bands on fields, uncertainty radii on map dots, under-review flags on disputed profile claims. Our users make five- and six-figure decisions from this screen; the least we owe them is showing our work. Any data product that will not show you its uncertainty is asking for faith, and faith is a poor foundation for a bid.
More blog posts to read

Geocoding 40,000 well records without losing our minds
County plats, section-township-range grids, and hand-typed latitudes all disagree. Here is the pipeline that reconciles them into one map.

Making the bid queue fast with virtualized lists
Five thousand rows brought the queue to its knees. Rendering only the visible ones took us from 4fps back to 60.
Orphan-well intel, monthly
Contact sales