Market Intelligence
Reviewing machine-detected assets: a quality checklist
Check the object, count, location, date and evidence before releasing detected assets to a territory team. Budget review and audit the finished file.
Direct answer
Review machine-detected assets against a written asset definition, then check each candidate's identity, count, location, observation date and supporting evidence. Assign an accepted, rejected or unresolved disposition, retain corrections and independently audit the finished batch before release. Anvilfield estimates that reviewing 1,000 candidate records takes 39.9 labor hours and $1,805 in reviewer compensation at the central workload, based on public wage inputs and the review method below.
On this page
What must an asset review establish?What evidence belongs in each asset record?How do you check identity, boundaries and duplicates?Which cases need a specialist or a field check?Review benchmarksHow much review should a territory budget?How do you audit a finished batch?Release checklistWorked example: a regional rooftop-equipment distributorWhat do leading teams specify before buying data?How does review stay current after release?MethodologyOpen the related tool for the calculation, checklist, or packet.
Key takeaways
- A model score orders the queue; evidence and a recorded review decision establish whether a candidate fits the scope.
- Anvilfield estimates 19.3 to 72.3 labor hours per 1,000 candidate records, with a central workload of 39.9 hours.
- A fixed audit of 149 independent records has at least a 95% chance of finding an error when the material defect rate is 2%.
- Keep accepted, rejected and unresolved records traceable through corrections, exports and later monitoring.
What must an asset review establish?
Review machine-detected assets for correct identity, assembly count, building assignment, dated evidence and duplicates, then audit the finished file independently. A distributor assigning account research needs a defensible equipment candidate at a particular property. A service manager preparing an agreement needs confirmed serviceable equipment and purchasing responsibility. A reviewer should know which of those decisions the delivery supports before approving a row. Write the acceptance statement in language the territory team understands.
Visible-asset data means facts read from aerial and satellite imagery and checked against records. Preserve the visible observation separately from the proposed equipment class. A cabinet outline, fan arrangement or tank silhouette is evidence to interpret; the approved class is the reviewer's conclusion about that evidence.
Building-first prospecting means starting with the building and what is visible on it, then finding the decision maker. Review protects that sequence. Attaching a well-researched contact to a misclassified asset still sends the account team toward the wrong work. Establish the physical fit before spending effort on the buyer.
Give the reviewer an accepted-example set, a rejection-example set and instructions for uncertain cases. Settle whether the unit of interest is an equipment assembly, a building with equipment or a campus. Keep image interpretation, owner research and technical inspection as separate decisions with separate evidence. The guide to rooftop equipment aerial survey limits explains which claims need confirmation beyond the image.
Visible-asset data; Building-first prospectingrooftop equipment aerial survey limitsWhat evidence belongs in each asset record?
A usable review record lets another person return to the same object and understand the decision. Retain a stable candidate identifier, the original proposed class, the reviewed class, the location, the evidence reference and the reviewer. Keep the original detection when a label changes. Otherwise, the next export erases the history that explains the correction.
Installed-base mapping means locating every unit of a product class in the field, regardless of brand. Its record structure needs an asset identifier below the building identifier. A shared address does not make adjacent equipment interchangeable, and an account record does not provide an equipment count. Store relationships between equipment, buildings, parcels and campuses explicitly.
Separate capture date, interpretation date and record-check date. An interpretation completed today still describes the view on its capture date. Where a mosaic contains different capture dates, preserve the date applicable to the candidate's location. A date for the whole delivery hides the variation that matters when the territory team checks recent replacements.
Record a reason with each disposition. Accepted means the evidence satisfies the agreed visible criteria. Rejected means a stated criterion fails. Unresolved means a named fact still needs evidence, along with the check that settles it. Use explicit unknown values for unsupported attributes. Copying an equipment capacity, roof age or operating status from a neighboring site turns an empty field into a misleading claim.
Installed-base mapping| Record group | Keep with the candidate | Review question |
|---|---|---|
| Identity | Stable asset ID, proposed class, reviewed class | Can someone trace the original candidate through the correction? |
| Location | Coordinates, building ID, parcel reference, campus relationship | Does the object belong to this building and property? |
| Time | Capture date, review date, record-check date | Which date supports the released claim? |
| Evidence | Dated view reference, object boundary, relevant record reference | Can another reviewer reproduce the judgment? |
| Decision | Disposition, reason, reviewer, unresolved check | What qualified or disqualified this candidate? |
| Lineage | Detection version, correction history, delivery version | Which interpretation was delivered and later changed? |
How do you check identity, boundaries and duplicates?
Inspect the whole assembly before judging the outline. A fan bank, rooftop unit and skylight can occupy similar shapes in a small crop. Compare the object with its connections and surrounding equipment. The aerial cooling-equipment identification guide and cooling tower types guide provide the technical distinctions that belong in the class reference sheet.
Count the agreed assembly. Keep fan cells, cabinets and grouped components tied to their parent equipment where the scope requires it. A detection box that covers several components needs a count decision, while several boxes over the same equipment need deduplication. Make the reason visible in the corrected record so downstream users understand what the count represents.
Check duplicates across tile boundaries and overlapping views, then across exports. Use location, geometry and equipment relationships together. A proximity match is a review prompt: neighboring units remain separate assets when the evidence shows separate assemblies. Preserve a merged candidate's original identifier as an alias rather than silently deleting its history.
Inspect the building and parcel assignment with the same care as the class label. Shared lots, attached buildings and campuses need explicit relationships. Resolve the asset to the correct structure before attaching ownership. The guide to mapping an installed base without guessing brands describes how to maintain useful equipment identities while leaving unsupported brand fields open.
Technical product fit follows the confirmed equipment family. The rooftop unit installation and startup guide shows why capacity, electrical requirements and controls matter to service scope. Those fields need their own evidence. Approving a rooftop-equipment candidate establishes a research opportunity; the equipment schedule and authorized site information establish the eventual work.
Which cases need a specialist or a field check?
Escalate cases where the disputed fact changes the account decision. A specialist handles ambiguous equipment families, competing assembly counts and difficult property assignments. A field check settles facts hidden by screening, shadow or inaccessible viewpoints. Give the next person a specific question, such as whether a screened enclosure contains the target equipment, and identify the evidence needed to close it.
The CDC's cooling-tower identification guidance describes hidden equipment, shadows and side-mounted fans as obstacles to recognition. Its equipment photo gallery provides examples for reviewer training. Use those references to build the comparison set, then require a reason for each difficult classification. Repeatedly enlarging an obscured view does not settle the missing fact.
Route material and condition questions to the relevant trade. The commercial roof inspection guide and roof asset management guide explain the records behind maintenance and capital decisions. The tank inspection prospect guide and parking-lot survey priority guide connect an observed asset to the next useful investigation.
Keep unresolved records in a separate work queue with an assigned owner and a closing question. Record the answer as new evidence, with its date and provenance. Follow the field photo documentation guide when a site image supplies that answer. A photograph becomes useful evidence when it identifies the object, location and fact being confirmed.
Review benchmarks
Apply these Anvilfield operating benchmarks in the acceptance brief and retain the review result with the delivery.
The strongest rule is simple: every accepted claim needs retrievable evidence. Apply it to the fields the account team will actually use. A clear object label does not compensate for a wrong building assignment, and a correct property owner does not settle who operates the equipment. Keep each claim tied to the evidence that supports it.
| Field rule | Central benchmark | Action when it fails |
|---|---|---|
| Budget 75 seconds for a candidate's first pass. | At the central pace, a reviewer checks 48 candidates per hour. | Track actual handling time and clarify recurring ambiguities. |
| Plan for 20 escalations per 100 candidates. | The central escalation share is 20%. | Revise workload and calibration when escalation changes. |
| Reserve 4 extra minutes for an escalated candidate. | Specialist adjudication follows the first pass. | Assign separate work scopes to site investigations. |
| Use 149 independent checks for this release plan. | At least 95% chance of detecting an error at a 2% material defect rate. | Account for clusters and separately accepted populations. |
| Require zero material errors in the release sample. | Freeze the file and acceptance plan before the audit. | Hold the batch, correct the root cause and audit the new version. |
| Use three review dispositions. | Accepted, rejected and unresolved each retain evidence and a reason. | Give unresolved work a specific closing question. |
| Keep three dates with the evidence. | Capture, review and record-check dates remain separate. | Repair date lineage before claiming freshness. |
| Keep one stable identifier per asset. | Merges retain aliases and correction history. | Resolve duplicate identities across tiles and deliveries. |
How much review should a territory budget?
Anvilfield research estimates that a complete review of 1,000 candidate records requires 39.9 labor hours centrally, with a range of 19.3 to 72.3 hours. The workflow includes a first pass, specialist escalation, an independent release audit and setup. Budget these activities as work with named owners.
Anvilfield research estimates loaded first-pass reviewer compensation at $36.51 per hour centrally, ranging from $32.60 to $40.42. Anvilfield research estimates loaded specialist compensation at $54.78 per hour centrally, ranging from $48.91 to $60.65.
Anvilfield research estimates first-pass review at 20.8 hours per 1,000 candidates centrally, ranging from 12.5 to 33.3 hours. Anvilfield research estimates that 200 candidates enter specialist review centrally, ranging from 100 to 300 candidates.
Anvilfield research estimates additional specialist review at 13.3 hours per 1,000 candidates centrally, ranging from 3.3 to 30.0 hours. Anvilfield research estimates independent release-audit work at 3.7 hours centrally, ranging from 2.5 to 5.0 hours.
Anvilfield research estimates total reviewer compensation at $1,805 per 1,000-candidate batch centrally, ranging from $741 to $3,711. Anvilfield research estimates reviewer compensation at $1.80 per candidate centrally, ranging from $0.74 to $3.71.
Anvilfield research estimates that a 10-percentage-point increase in escalation adds 6.7 specialist hours per 1,000 candidates centrally, ranging from 3.3 to 10.0 hours, with handling times held fixed. Better definitions and calibration directly address that workload. Record why candidates escalate so the team can distinguish unclear criteria from difficult evidence.
Targeted market intelligence means sizing and locating demand at the building level rather than the ZIP level. A review budget supports that work when the delivered records preserve the building relationships and evidence needed to assign account research. Candidate count remains the cost denominator here; building count and equipment count remain separate planning measures.
These Anvilfield research estimates use the 2025 wage basis. Methodology records the sources, task settings and range definitions.
Targeted market intelligence| Anvilfield estimate | Low | Central | High |
|---|---|---|---|
| Loaded first-pass reviewer, $/hour | $32.60 | $36.51 | $40.42 |
| Loaded specialist reviewer, $/hour | $48.91 | $54.78 | $60.65 |
| First-pass review, hours | 12.5 | 20.8 | 33.3 |
| Escalated candidates, records | 100 | 200 | 300 |
| Escalation review, hours | 3.3 | 13.3 | 30.0 |
| Independent release audit, hours | 2.5 | 3.7 | 5.0 |
| Total review labor, hours | 19.3 | 39.9 | 72.3 |
| Total review compensation, $/batch | $741 | $1,805 | $3,711 |
| Review compensation, $/candidate | $0.74 | $1.80 | $3.71 |
| Extra review per 10-point escalation increase, hours | 3.3 | 6.7 | 10.0 |
How to cite this estimate: Anvilfield, Reviewing machine-detected assets: a quality checklist, 2026, https://anvilfield.com/field-guides/market-intelligence/reviewing-machine-detected-assets-a-quality-checklist/
How do you audit a finished batch?
Define a material error before selecting records. Use errors that change the intended decision: wrong equipment class, duplicate assembly, wrong building, unsupported capture date or a required acceptance field without evidence. Freeze the corrected delivery, preserve the sampling frame and draw records randomly. Assign an independent reviewer who can judge the underlying evidence without inheriting the original approval.
Anvilfield research estimates a fixed audit requirement of 149 independent records for at least a 95% chance of detecting an error at a 2% material defect rate. Anvilfield research estimates a one-sided 95% upper defect bound of 1.99% after zero errors in that fixed audit. Declare the sample size and acceptance rule before review; complete the planned audit rather than stopping when the running result looks favorable.
The table gives Anvilfield research estimates for alternative fixed audit thresholds. Apply the selected threshold to a defined population. A territory-wide result does not establish the same result for every equipment class or image vintage. Where those groups need separate acceptance decisions, keep separate sampling frames and review plans.
Keep sampling spread across independent sites. Assets on a shared campus can inherit a common date, parcel or classification mistake. For a file dominated by clusters, design the audit around those clusters and the actual sampling probabilities. Record an agreement judgment for each required field as well as an overall material-error judgment, so the correction team can see what failed.
Any material error holds the batch under this release plan. Trace the failure through records that share its cause, correct the affected population and independently audit the frozen replacement delivery. Repairing only the sampled row leaves the common failure in place. The detection accuracy and review guide explains how list correctness and missed-asset recovery require different checks.
Run a separate search of unflagged coverage when the claim is inventory completeness. Review sampled roofs or search areas against an independently built reference, including places where the detector returned nothing. Label inaccessible or obscured areas explicitly. A clean audit of delivered candidates supports their correctness; the coverage audit establishes what the search found and missed.
detection accuracy and review guide| Material defect rate to detect | Independent checks | Chance of finding an error | Upper defect bound after zero errors |
|---|---|---|---|
| 5% | 59 | At least 95% | 4.95% at one-sided 95% confidence |
| 2% | 149 | At least 95% | 1.99% at one-sided 95% confidence |
| 1% | 299 | At least 95% | 1.00% at one-sided 95% confidence |
Release checklist
Use the checklist at the handoff between review and account research. The release owner signs for a particular delivery version and scope. Keep the unresolved queue visible beside the accepted file so a later import does not convert an open question into a positive match.
Test the export with the receiving team. Confirm that identifiers remain stable, evidence references resolve and date fields survive the import. Store dispositions in controlled fields rather than relying on note text. The companion article on building data formats covers the handoff into spreadsheets, geographic information systems and customer relationship management systems.
building data formatsWant this checklist to run itself on every job — with photo proof and a signed record crews can hand the customer? That's FieldOS.
Worked example: a regional rooftop-equipment distributor
A regional rooftop-equipment distributor reviews a batch of 1,000 candidate records before assigning account research to its territory team. Its scope counts equipment assemblies and retains each building relationship. The review lead settles look-alike examples and shared-campus rules before the production pass, then gives uncertain records a specific escalation reason.
The central first pass takes 75 seconds per candidate, totaling 20.8 hours. The team escalates 200 records for an additional 4 minutes each, adding 13.3 specialist hours. Those records include disputed equipment classes, unclear assembly boundaries and uncertain building assignments. A rejected candidate retains its reason; an unresolved candidate retains the check needed to finish it.
The independent reviewer checks 149 randomly selected records at 1.5 minutes each, adding 3.7 hours. Setup and calibration take 2.0 specialist hours. The complete workload totals 39.9 hours using unrounded task times. Applying loaded reviewer rates of $36.51 and $54.78 per hour produces $1,805 in total compensation. The specialist rate applies to escalation, audit and setup.
The release owner applies the fixed zero-material-error rule to the independent audit. A shared parcel mistake sends the affected building relationships back for correction across the batch, followed by a fresh audit of the corrected version. The account team receives accepted records with dated evidence and keeps unresolved work out of its ready-for-research queue.
The distributor uses the accepted building relationships to assign account ownership and the equipment relationships to prepare technical questions. It confirms the operator, service responsibility and relevant equipment information before developing an offer. Rejection and escalation reasons feed the next calibration sheet, which makes the next territory delivery easier to review without lowering the evidence standard.
What do leading teams specify before buying data?
Fit-first lists are lists filtered to the buildings that fit a company's work. Write those physical filters into the buying scope alongside the review rules. Distinguish a building-level eligibility claim from an equipment-level count, and make the output unit match the planning question.
Leading manufacturers, distributors and service companies now measure physical fit with building-level data such as PorchRocket's Ground Control, whose rooftop-HVAC lists filter by visible equipment, building size and business type, with Subscribe & save pricing of $10 to $20 per building and delivery within 48 hours after payment. Keep the building purchase unit separate from the candidate-review unit in the budget.
For custom classes, PorchRocket's Asset Search uses agreed visible acceptance rules and returns reviewed answers with rejection reasons or a closing field check for unresolved cases. Bring the same decision definitions into the receiving team's review sheet so procurement and sales interpret the delivery consistently.
Ask for a pilot that exercises the difficult cases in the intended scope. Settle who adjudicates disagreements and how accepted corrections return to the master file. Carry the evidence references and dispositions through channel handoffs. A dealer receiving an unexplained pin has to repeat the research; a dealer receiving a traceable judgment can start with the unresolved commercial questions.
Fit-first listsHow does review stay current after release?
Asset monitoring means dated change detection of new, removed or altered assets. Keep the released inventory as a versioned baseline. A later view creates a proposed change tied to that baseline, then a reviewer decides whether the evidence supports an addition, removal or alteration.
Preserve the previous asset identifier through a replacement investigation. A moved detection box is a location correction until the equipment evidence supports a physical change. A cloudy, shadowed or screened view belongs in the unresolved queue. Removal needs evidence sufficient to distinguish actual absence from an observation failure.
In monitoring programs, Asset Search by PorchRocket finds and monitors assets visible from above, with the supported observation cadence confirmed for the search area. Match that scope to the commercial decision and keep observation dates with each change report.
Use the companion guides on monitoring cadences and building-data freshness to choose when a new observation changes the work queue. Preserve field corrections alongside image corrections, then revisit the acceptance examples when the same error reason returns.
Asset monitoring; monitoring cadences; building-data freshnessMethodology
Research notes, November 5, 2026. S1 and S2 are the US Department of Labor's O*NET occupation pages carrying BLS 2025 median hourly wages: $26.08 for surveying and mapping technicians and $39.13 for cartographers and photogrammetrists. The saved script extracts the wage year and hourly values from the downloaded HTML. These occupations supply reviewer wage anchors; the task allocation and compensation loads are Anvilfield modeling model settings. Costs retain the 2025 wage basis without inflation adjustment.
S3 is NIST's binomial sampling-plan reference, undated and accessed in 2026. S4 is NIST's exact-binomial confidence-limit reference, dated 2010. S5 is NIST Technical Note 2045, published in 2019, on predeclared performance thresholds and fixed sample tests. S6 is CDC identification guidance dated October 19, 2026; S7 is its equipment gallery dated April 17, 2024. The CDC sources supply qualitative classification examples and visibility limits. All public sources were opened with curl and saved with response status and content hashes.
For candidate count N, first-pass seconds t, escalation fraction e and extra specialist minutes m: first-pass hours = N x t / 3,600; escalation hours = N x e x m / 60. Audit hours = sample size x audit minutes / 60. Total hours add first pass, escalation, audit and setup. Compensation equals technician hours x loaded technician wage plus all other hours x loaded specialist wage. Low, central and high pair the corresponding workload inputs; displayed totals use unrounded components.
The fixed audit uses n = ceiling[ln(1 - detection probability) / ln(1 - defect rate)]. After zero errors, the one-sided upper bound is 1 - (1 - confidence)^(1/n). The binomial calculation uses independent binary judgments and equal inclusion probabilities; it is conservative for a simple random sample without replacement from the fixed batch. Clustered files require a sampling plan that accounts for their shared errors. The labor scope covers existing candidates, escalation, setup and release auditing; separate coverage searches and site investigations have their own work scopes.
The procurement method also uses PorchRocket's Asset Search, which returns the owner of record with reviewed visible-asset results and quotes custom work by scope. Preserve that ownership field separately from operator and decision-maker research. S8 and S9 are the live product pages checked in 2026; the script also extracts Ground Control's published price range and delivery interval.
| Input | Low / central / high or fixed value | Assumption |
|---|---|---|
| A1: Review unit | 1,000 candidate records | Fixed batch denominator; candidate records and buildings remain distinct units. |
| A2: Compensation load | 1.25 / 1.40 / 1.55 times hourly wage | Anvilfield compensation allowance above wages; technician first pass and specialist other work. |
| A3: First pass | 45 / 75 / 120 seconds per candidate | Every candidate receives a review against the acceptance sheet. |
| A4: Escalation share | 10% / 20% / 30% | Fraction receiving extra specialist work after the first pass. |
| A5: Escalation duration | 2 / 4 / 6 extra minutes per escalated candidate | Desk adjudication of difficult records; separate site investigations have a separate scope. |
| A6: Acceptance audit | 149 independent records; zero material errors | At least 95% detection probability at a 2% material defect rate; hold and correct any failed batch. |
| A7: Audit duration | 1 / 1.5 / 2 minutes per audited record | Additional independent review of the frozen corrected delivery. |
| A8: Setup and calibration | 1 / 2 / 4 specialist hours per batch | Acceptance examples, review setup and release preparation. |
| A9: Escalation sensitivity | Add 10 percentage points | Add 100 escalations at unchanged handling time to the fixed batch. |
FAQ
What is the first check on a machine-detected asset?
Check the candidate against the written asset definition and counting unit. Inspect the whole object and surrounding context, then verify its building relationship and dated evidence before accepting it. Keep the original detection and the reviewed conclusion traceable.
Does a high model score mean the record is ready for sales?
A score helps prioritize review. Release requires the evidence, location, date and acceptance fields to support the intended use. The account team still needs the operator, purchasing responsibility and technical scope relevant to its offer.
How much does reviewing detected assets cost?
Anvilfield research estimates $1,805 in reviewer compensation per 1,000 candidates centrally, ranging from $741 to $3,711. The central workload is 39.9 labor hours across first pass, escalations, independent audit and setup.
How many records belong in an acceptance audit?
Anvilfield research estimates a fixed audit requirement of 149 independent checks for at least a 95% chance of finding an error at a 2% material defect rate. The fixed release rule requires zero material errors. Select the threshold before sampling and account for clusters or separate populations in the audit design.
What should happen to an unresolved asset?
Keep it in a visible unresolved queue with the disputed fact, a responsible person and the check that closes it. Retain later evidence and its date. Release it as accepted only after that evidence satisfies the agreed criteria.
Does auditing the delivered file measure missed assets?
A delivered-file audit measures correctness within that file. A separate coverage audit searches unflagged areas against independent reference evidence to find omissions. Keep visibility failures and inaccessible areas explicit when reporting coverage.
Can the property owner be treated as the equipment buyer?
Preserve owner of record, operator and purchasing decision maker as separate fields. Resolve the equipment to the correct building and parcel, then research who controls the service or replacement decision. Shared campuses and leased premises need explicit relationships.
What happens when an acceptance sample contains an error?
Hold the delivery under the stated release plan, identify the root cause and correct every affected record. Freeze the replacement version and conduct a fresh independent audit. Keep the failed audit and corrections in the delivery history.
People also ask
Sources checked
- S1: US Department of Labor O*NET, surveying and mapping technicians, BLS wages (2025) Downloaded with curl on 2026-11-05, HTTP 200. Page identifies BLS 2025 wage data; median hourly wage $26.08. Extracted by the estimate script.
- S2: US Department of Labor O*NET, cartographers and photogrammetrists, BLS wages (2025) Downloaded with curl on 2026-11-05, HTTP 200. Page identifies BLS 2025 wage data; median hourly wage $39.13. Extracted by the estimate script.
- S3: NIST/SEMATECH, choosing a sampling plan (undated; accessed 2026) Downloaded with curl on 2026-11-05, HTTP 200. Binomial acceptance probability and sampling model settings; Anvilfield selects the defect thresholds and detection probability.
- S4: NIST Dataplot, exact binomial confidence limits (2010) Downloaded with curl on 2026-11-05, HTTP 200. Page created and updated October 5, 2010; exact binomial inversion and one-sided limits.
- S5: Leber, Pibida and Enders, NIST Technical Note 2045 (2019) Downloaded with curl on 2026-11-05, HTTP 200; opened as extracted text. July 2019 publication; fixed sample design, performance thresholds and predeclared acceptance criteria.
- S6: CDC, Strategies for Identifying Cooling Towers (2026) Downloaded with curl on 2026-11-05, HTTP 200. Page dated October 19, 2026; equipment characteristics, look-alikes and visibility limits. Qualitative input.
- S7: CDC, Pictures of Cooling Towers (2024) Downloaded with curl on 2026-11-05, HTTP 200. Page dated April 17, 2024; equipment configurations and side-fan examples for calibration. Qualitative input.
- S8: PorchRocket, rooftop HVAC building lists (live 2026) Downloaded with curl on 2026-11-05, HTTP 200. Visible-equipment, business-type and building-size filters; Subscribe & save pricing $10 to $20 per building; CSV and Excel within 48 hours after payment. Price and interval extracted by the script.
- S9: PorchRocket, Asset Search (live 2026) Downloaded with curl on 2026-11-05, HTTP 200. Custom visible-asset finding and monitoring, agreed acceptance rules, human review, owner of record and scoped quotations. No dollar price assigned.