Basalt Sea Press — Economic analysis
Whether a $4.0M county contribution is recovered through the revenue the facility generates.
The numbers, what produced them, over what period, and which industries they land in.
Read this part if you read nothing else. It is written for someone deciding whether to support the project, not for an economist.
Five of the words in this report look like plain English and mean something narrower. A reader who takes them at face value will believe they have understood and will be wrong in a predictable direction — usually upward. This page is here so that does not happen.
Take a bakery in town that wins a large new contract.
It hires four people to fill the order. Those four jobs are the direct effect — the ones at the business itself.
To bake more, it buys more flour from a mill up the road, more packaging, more delivery runs. The work that creates at the mill, the box supplier and the haulier is the indirect effect — jobs at the businesses that supply the bakery.
The four new bakers, and the people at the mill, spend their wages: groceries, rent, a dentist, a haircut. The work that supports is the induced effect — jobs paid for out of wages earned elsewhere in the chain.
Add the three together and you have the total. The whole of it depends on the first four jobs actually happening, and on the flour, the boxes and the haircuts being bought here rather than somewhere else. Both of those are assumptions, and the rest of this report is largely about how far they can be trusted.
| Word | What it sounds like | What it actually counts |
|---|---|---|
| Output | sounds like what the business produces and sells | the total value of all sales set off across the region, counted at every step. A dollar of groceries bought by a worker is counted again when the shop restocks. It is a measure of activity, not of wealth created, and it is always larger than the money anyone ends up with. |
| Value added | sounds like the benefit to the region | wages, profits and taxes on production — output minus what was bought from someone else to make it. This is the figure comparable to a county's GDP. It is the honest size number, and it is smaller than output. |
| Jobs | sounds like the number of people who get hired | an average level of employment supported while the activity continues, including people at suppliers and at businesses where the wages get spent. One job lasting a year and twelve people working a month each count the same. Construction figures are job-YEARS and are not ongoing positions at all. |
| Multiplier | sounds like a return on investment | how much total activity follows each dollar of direct activity. Every dollar spent anywhere has one, including the dollar spent on something else instead, so a large multiplier is not by itself an argument for anything. |
| Displacement | sounds like a technicality | the share of the money that was already being spent inside the region and has simply moved. It is the difference between a project that grows an economy and one that rearranges it. Of the twenty published studies we read while building this method, the five advocacy studies checked for it netted nothing at all, and the one that measured it — an arts study surveying its own attendees — found 49.4% of local attendees would have spent locally anyway. Those twenty were selected rather than sampled, so that is a witness and not a rate. |
These figures are placeholders showing what the analysis produces. They are NOT findings about this economy. Do not quote them, in any setting, for any purpose.
The pre-publication review reached: Provisional only — must be stamped throughout.
A conventional economic impact study answers the first of these and stops. The other four are what decide whether a project happens, and they are where projects fail.
| The question | What answers it | In this study | Note |
|---|---|---|---|
| Will it produce what is claimed? | The impact model, and how much of it is new money | Examined | See the section named below. |
| Can it actually be built? | Zoning, overlays, land and the approvals still outstanding | Not examined | This study did not examine it. An answer here would need work that was not commissioned, and its absence is not evidence that the answer is favorable. |
| Can it be paid for? | What the capital stack can secure, and what must be equity | Not examined | This study did not examine it. An answer here would need work that was not commissioned, and its absence is not evidence that the answer is favorable. |
| Can it be staffed? | Whether the workforce can live within reach of the work | Not examined | This study did not examine it. An answer here would need work that was not commissioned, and its absence is not evidence that the answer is favorable. |
| Will it survive? | Coverage, volatility and what the competition does to it | Not examined | This study did not examine it. An answer here would need work that was not commissioned, and its absence is not evidence that the answer is favorable. |
What is withheld. Any answer, in either direction. Nothing in this study reports what an examination would have found, and no figure elsewhere in the document is offered in place of one.
Why. The analysis that would settle it was not run. Most often the engagement did not commission it; sometimes an input it needs was never supplied. Which of those it was is recorded for each item rather than left to the reader, because a question nobody paid to ask and a question that could not be answered are different facts arriving as the same blank. In this case: can it actually be built; can it be paid for; can it be staffed; will it survive.
What can still be used. Everything here that was examined, at the strength printed beside it, and the list of what was not. What the study looked at, it answers completely; what it did not, written down, is the scope of the engagement made visible — which is what a board needs in order to decide whether to buy the rest.
What would clear it. Commissioning the named analysis. Each unexamined item is named rather than summarized, and what running it would take is stated with it, so what is left is a finite list somebody can price rather than an open-ended request for more study.
What this does not mean. That what was not examined would have come back favorably, or unfavorably. An absence of analysis reads exactly like an analysis that came back clean, and this document keeps the two apart everywhere precisely because nothing about the blank itself distinguishes them.
Annual Operations is already net of displacement. 28% of its demand is spending the region would have seen anyway, so $2.5M of gross output was removed before any figure above was computed, and $6.4M is what remains. Basis: Intercept surveys at three comparable facilities.
No displacement adjustment was applied to Construction, so that figure is gross: spending the region would have seen anyway is counted there as new. That share is unknown rather than zero.
The adjusted and the unadjusted figures above are therefore on different bases, and are not comparable with each other without that in mind.
Over the 2026–2037 horizon the project generates $71.0M in total regional output undiscounted, or $57.1M in present value at 4.0%. Once construction ends and operations reach full capacity, the recurring level is 64.0 jobs — the figure a community experiences year after year, and the one that should anchor any long-run expectation.
Every figure in this report is meaningless on its own. A number of dollars is large or small only against something, and the something has to be named and printed or the reader cannot check it. This page puts each headline figure over a denominator from Illustrative Region’s own accounts — the same accounts that produced the figure.
| Figure | Size | Against | How big that is | What it says |
|---|---|---|---|---|
| Annual Operations: jobs supported, per year | 64 jobs | all jobs in Illustrative Region | 22,628 jobs | The project accounts for 0.28% of all jobs in Illustrative Region. |
| Annual Operations: value added, per year | $4.9M | the whole economy of Illustrative Region (value added, which is what a county's GDP measures) | $1.82B | The project accounts for 0.27% of everything Illustrative Region produces in a year. |
Nothing in Illustrative Region sits within a factor of two of the Annual Operations in either direction, so there is no industry to say it is “about the size of”. Naming the nearest anyway would put a number many times larger or smaller beside this one under a heading claiming they are comparable. The project is smaller than every industry in this region: the smallest, Utilities, supports 1,028 jobs against the project’s 64. That is a finding about scale, and it is the honest counterweight to a share expressed as a percentage: a fraction of a percent of a county is a real number and a small one.
| Nearest industry by size | Jobs it supports here |
|---|---|
| Utilities | 1,028 |
Each of these could have been divided by something and printed as a share. Each would have been wrong in a predictable direction — upward. A refusal here is a finding about the figure, not a gap in the analysis.
What is withheld. This figure as a share of the economy of Illustrative Region. Output is not divided by regional GDP anywhere in this report, and no sentence describes the project as increasing the size of this economy by the output figure.
Why. Output is not comparable to a region's GDP and is not divided by it here. Output counts the same dollar again at every step it passes through; GDP counts it once. Their ratio would overstate the project's share of this economy — an error we found once in the twenty published impact studies read while this method was built, twenty we selected rather than a sample of the field, so it stands as one witness and not as a rate. That report headlined an $11.2bn increase in the size of a state's economy where its own value added was $6.96bn, a 61% overstatement produced by swapping one word, in a table whose gloss on the output row read that GDP is a measure of the overall size of the economy.
What can still be used. Value added is the comparable figure, and it is anchored above at $4.9M.
What would clear it. Nothing supplied later closes this one, and it is not waiting on an input. Output and value added measure different things, so no denominator makes their ratio mean what a share of an economy means. The refusal is permanent, and it is the finding rather than a stage on the way to one.
What this does not mean. That the output figure is wrong, or that it is inflated. Output is a real quantity, correctly computed here, and it is the right number for the question output answers — how much business the project moves through this region's suppliers. It is the wrong numerator for a share of an economy, which is a different question that value added answers.
What is withheld. This figure as a share of the workforce in Illustrative Region. It is not divided by regional employment here, and no percentage of the region's jobs is stated for it anywhere in this report.
Why. A one-time phase produces job-YEARS, which is a quantity of work, not a number of positions that exist at any moment. Dividing it by the region's employment would produce something that reads as a share of the workforce and is not one. This figure is not anchored that way here, and it is not added to the annual jobs figure anywhere in this report.
What can still be used. The figure itself, read as the work it is: a crew of a given size for a given number of months. The construction phase's own duration, stated in the timeline, is what converts between the two, and the conversion is arithmetic a reader can do on the page.
What would clear it. Nothing supplied later turns this ratio into a share, because a job-year and a job are different units and no additional data reconciles them. What a denominator in the same unit would allow is a different and legitimate comparison — this region's own annual construction employment, against which a construction phase can honestly be sized.
What this does not mean. That the construction work is not real, or that it is smaller than the annual jobs figure it cannot be added to. It is a genuine quantity of employment and it is often the largest number in a study of this kind. What it is not is a standing level of jobs, and the two are kept apart here because a report that adds them reports a decade of temporary work as though it were permanent.
What is withheld. Every per-resident figure. No figure in this report is divided by a resident count, and none is stated in a unit a household can hold.
Why. No population was declared for the study place, so there is no resident count in this study to divide by.
What can still be used. The share-of-region anchors above, which use employment and value added and need no population at all. Nothing on this page depends on the missing count except the per-resident line itself.
What would clear it. One ACS table supplies it, at no cost and without re-running anything here. It is the cheapest of the missing inputs on this page to close.
What this does not mean. That the method cannot produce a per-resident figure. Its absence here is a gap in what was commissioned, not a limit of the method, and the same run produces the figure the moment the count is supplied.
| Study area | Illustrative Region |
| Analysis type | Impact analysis - prospective change |
| Client | Illustrative County Board of Commissioners |
| Dollar year | 2024 |
| What is counted | the project, its suppliers, and the wages workers spend locally (Type II (closed on households at the accounts-implied rate)) |
| Regional model | supplied to us already built for this region — see Part V |
| Discount rate | 4.0% |
| Evidence tier | illustrative |
| Prepared by | Project Red Team |
These figures answer for the region: what the project adds to output, value added and employment inside the study area. They are not a finding about whether the project pays. Whether the operator covers its costs and services its debt, whether a jurisdiction that contributes money gets it back, and what the land is worth if an entitlement never arrives are three further questions, each asked of a different party, and a figure in this section is not evidence about any of them. A region can gain from a facility that never pays its lender back.
| Measure | Direct | Indirect | Induced | Total |
|---|---|---|---|---|
| Output | $5.9M | $2.3M | $2.9M | $11.1M |
| Labor income | $2.0M | $756,622 | $1.1M | $3.8M |
| Value added | $4.2M | $1.7M | $2.2M | $8.1M |
| Employment | 43.7 | 21.1 | 28.1 | 92.9 |
The arithmetic, in words: the project itself accounts for 44 of these. That activity supports a further 49 at suppliers and at the businesses where the wages are spent, for a total of 93 — counted as job-years of temporary work.
One-time total over the construction period. Employment is job-years of temporary work, not permanent positions. Output multiplier 1.870. Type II (closed on households at the accounts-implied rate).
| Measure | Direct | Indirect | Induced | Total |
|---|---|---|---|---|
| Output | $3.5M | $1.1M | $1.8M | $6.4M |
| Labor income | $1.3M | $366,675 | $651,475 | $2.3M |
| Value added | $2.7M | $837,864 | $1.4M | $4.9M |
| Employment | 36.9 | 10.1 | 17.1 | 64.0 |
The arithmetic, in words: the project itself accounts for 37 of these. That activity supports a further 27 at suppliers and at the businesses where the wages are spent, for a total of 64 — counted as jobs, as an average level while it operates.
Recurring annually at full operation. Output multiplier 1.815. Type II (closed on households at the accounts-implied rate).
The one-time and annual figures above are in different units and must not be added together. A construction impact happens once; an operations impact happens every year. The timeline section below places both on a common footing, which is the only defensible way to combine them.
The effect concentrates in accommodation, food service, and recreation, which takes 35% of the total output effect. The next is retail trade at 19%.
Reading it: the direct column is the project's own purchases landing in that industry, the indirect column is what those industries buy locally in turn, and the induced column is what the wages from both buy in shops and clinics. An industry with a large indirect number and a small direct one is a supplier the project never contracts with.
| Industry | Direct | Indirect | Induced | Total | Share |
|---|---|---|---|---|---|
| ACC — Accommodation, food service, and recreation | $2.1M | $30,548 | $147,698 | $2.3M | 35.5% |
| RET — Retail trade | $738,720 | $103,075 | $339,338 | $1.2M | 18.5% |
| HLT — Health, education, and social services | $191,520 | $68,522 | $282,438 | $542,480 | 8.5% |
| TRN — Transportation and warehousing | $257,184 | $141,410 | $97,661 | $496,255 | 7.8% |
| PRO — Professional and business services | $218,880 | $86,392 | $110,436 | $415,708 | 6.5% |
| FIN — Finance, insurance, and real estate | $0 | $69,906 | $296,169 | $366,075 | 5.7% |
| MFG — Manufacturing | $6,566 | $115,557 | $127,048 | $249,172 | 3.9% |
| CON — Construction | $0 | $119,000 | $78,087 | $197,088 | 3.1% |
Timeline covers 2026-2037. Annual activities are held at their final ramp level through the horizon; one-time activities occur only in their scheduled years.
| Year | Regional output | Jobs in that year | Recurring or not |
|---|---|---|---|
| 2026 | $5.5M | 46.5 | not yet at the recurring level |
| 2027 | $5.5M | 46.5 | not yet at the recurring level |
| 2028 | $3.8M | 38.4 | not yet at the recurring level |
| 2029 | $5.1M | 51.2 | not yet at the recurring level |
| 2030 | $6.4M | 64.0 | at the recurring level |
| 2031 | $6.4M | 64.0 | at the recurring level |
| 2032 | $6.4M | 64.0 | at the recurring level |
| 2033 | $6.4M | 64.0 | at the recurring level |
| 2034 | $6.4M | 64.0 | at the recurring level |
| 2035 | $6.4M | 64.0 | at the recurring level |
| 2036 | $6.4M | 64.0 | at the recurring level |
| 2037 | $6.4M | 64.0 | at the recurring level |
Each figure in the jobs column is the employment supported within that single year. Adding the column down produces job-years of work, which is a different unit from an ongoing job and is not added to the annual jobs figure anywhere in this report. The recurring level is 64.0 jobs, and the 4 years before it are lower because construction and ramp-up have not yet given way to full operation; work counted in those years lasts as long as the year it sits in rather than continuing after it.
The undiscounted total and the present value describe the same stream of activity. The undiscounted figure is larger and is the one usually quoted; the present value is the one that can be compared against a cost incurred today. Both appear here so that neither has to be taken on trust.
No tribal rate appears in the Illustrative County tax structure, so there is no tribal line above. That is an absent rate, not zero revenue: this study was never given a rate to apply, and printing a zero would report a finding where there is only a gap in the inputs. Supply the tribal rate and this section reports the line.
A further $1,211,173 in state and local revenue arrives once, during the build. It is shown separately because it is not repeatable, and adding it to the figures above would present the construction year as though it were every year.
| Revenue source | Each year at full operation |
|---|---|
| Sales and Excise | $439,529 |
| Property | $269,389 |
| Personal Income | $267,874 |
| Social Insurance | $290,329 |
| Corporate Profits | $246,096 |
| Total | $1,513,216 |
Tax revenue is derived from average effective rates applied to modeled tax bases. It is not a jurisdiction-specific tax calculation and should not be used as a revenue forecast without reconciliation against actual collections. Tax structure: Illustrative County (Project Red Team illustrative rates — replace with jurisdiction actuals)
These are revenue figures only. Serving the households a project brings costs money — schools, roads, public safety — and this study has not been given the jurisdiction's service costs, so it cannot say whether the revenue above exceeds them. That is the first question a commissioner will ask. The costs come from the jurisdiction's own budget divided by its own households; supply them and this section reports a net position instead.
Final-demand multipliers: the total regional effect generated per $1.00 of new final demand delivered to each industry. Employment is expressed as jobs per $1 million. When a study supplies a stated direct job count, the employment headlines preserve that count and use the multipliers only for indirect and induced jobs; this jobs-per-dollar column will not reproduce the published employment total.
| Industry | Output | Labor income | Jobs per $1M |
|---|---|---|---|
| AGR — Agriculture, forestry, and fishing | 1.657 | 0.452 | 15.8 |
| MIN — Mining and resource extraction | 1.659 | 0.454 | 15.7 |
| UTL — Utilities | 1.889 | 0.654 | 15.4 |
| CON — Construction | 1.887 | 0.655 | 15.5 |
| MFG — Manufacturing | 1.859 | 0.564 | 12.7 |
| WHL — Wholesale trade | 1.721 | 0.588 | 17.8 |
| RET — Retail trade | 1.720 | 0.588 | 17.8 |
| TRN — Transportation and warehousing | 1.721 | 0.589 | 17.8 |
| FIN — Finance, insurance, and real estate | 1.827 | 0.745 | 16.8 |
| PRO — Professional and business services | 1.825 | 0.746 | 16.9 |
| HLT — Health, education, and social services | 1.826 | 0.744 | 16.8 |
| ACC — Accommodation, food service, and recreation | 1.859 | 0.685 | 25.2 |
Whether the economy has the labor, the suppliers and the structure the projection assumes, and who the gains reach.
For the people who would have to make it work: a county manager, a workforce board, a planning department.
The modeled job mix for 64 jobs averages 10% below the regional median wage ($41,199 job-weighted average wage against a $46,000 median). 61% of new labor income reaches the two lowest household brackets, against 49% of households in the region — more broadly distributed than the status quo.
| Occupation | Jobs | Share | Average wage |
|---|---|---|---|
| Food service workers | 9.5 | 14.9% | $24,500 |
| Retail salespersons | 5.7 | 8.9% | $31,000 |
| Housekeeping and cleaning | 4.3 | 6.6% | $26,000 |
| Supervisors and managers | 4.2 | 6.6% | $48,583 |
| Recreation and fitness workers | 3.8 | 5.9% | $29,000 |
| Front-desk and clerks | 3.5 | 5.5% | $30,500 |
| Cashiers | 2.9 | 4.5% | $26,500 |
| Stock and material movers | 2.1 | 3.2% | $33,000 |
| Maintenance | 1.5 | 2.3% | $38,000 |
| Nursing assistants and aides | 1.4 | 2.1% | $33,000 |
| Registered nurses | 1.0 | 1.5% | $78,000 |
| Office and administrative | 0.9 | 1.4% | $39,620 |
Staffing patterns were available for 69% of the jobs modeled. The other 31% are real jobs whose occupational mix this study cannot break down, and they are left out rather than distributed by assumption.
A further 6% was broken down and simply did not reach the twelve largest occupations. That share is missing from the table above but not from the analysis, which is why the Share column sums to less than the coverage figure.
Everything above asks what a change would do. This asks a different question, and it is the one a workforce plan, a tax-base argument or a grant narrative actually needs: how much of Illustrative Region is each industry today, once its suppliers are counted. The two questions have different arithmetic and are not interchangeable.
| Industry | Output attributable to it | Share of regional output | Share of regional jobs | What an impact run would have added |
|---|---|---|---|---|
| Manufacturing | $360.7M | 15.0% | 9.5% | 3.4% |
| Health, education, and social services | $325.2M | 13.6% | 12.9% | 1.1% |
| Construction | $294.4M | 12.3% | 10.0% | 2.7% |
| Retail trade | $293.6M | 12.2% | 13.7% | 2.2% |
| Finance, insurance, and real estate | $279.9M | 11.7% | 11.1% | 1.1% |
| Accommodation, food service, and recreation | $268.2M | 11.2% | 17.7% | 0.9% |
| Agriculture, forestry, and fishing | $241.5M | 10.1% | 10.1% | 1.8% |
| Mining and resource extraction | $239.1M | 10.0% | 9.9% | 1.4% |
| Transportation and warehousing | $207.9M | 8.7% | 9.7% | 2.1% |
| Wholesale trade | $204.6M | 8.5% | 9.6% | 2.2% |
| Utilities | $200.4M | 8.4% | 6.7% | 2.3% |
| Professional and business services | $194.7M | 8.1% | 7.8% | 1.3% |
Run as an impact analysis instead, Manufacturing would have come out 3.4% higher — $372.8M against $360.7M.
That difference is not a range and the larger figure is not a more optimistic reading. It is an error for this question. An impact run lets the industry buy from itself through the multiplier rounds, so the answer includes output the region did not produce — and Illustrative Region cannot have produced more of a thing than it produced. The figure in the table is the one that can be defended.
Impact answers: what would change if this activity were added or removed. It is the right question for a proposal.
Contribution answers: how much of the economy already is this. It is the right question for something operating.
Of the twenty published impact studies read while this method was built — a set we chose and read by hand, so the count describes those twenty and not the field — nineteen reported one of these numbers in the other's language. The one that published both, and disclosed the gap, was the most checkable document in the set.
These shares overlap and must not be summed: every industry draws on the same suppliers, so adding them counts the shared supply chain once per industry.
Before asking what a project does to a region, it is worth establishing what the region already is. The same matrix that produces the impact tables answers a different and often more useful set of questions: which industries pull the rest of the economy along, which the rest depends on, and what would be lost if one of them went away.
| Key sector | Buys locally | Depended upon | Share of output |
|---|---|---|---|
| HLT — Health, education, and social services | 1.02 | 1.21 | 10.8% |
| MFG — Manufacturing | 1.04 | 1.03 | 10.8% |
| FIN — Finance, insurance, and real estate | 1.02 | 1.24 | 9.3% |
Extraction assumes nothing replaces the sector. Over any real horizon something partly does — imports substitute, other employers absorb some workers — so these are upper bounds on the loss, not forecasts. They are most useful as a ranking of exposure.
| Sector | Initial | First round | Supply chain | Induced | Total | Induced share |
|---|---|---|---|---|---|---|
| MFG | 1.00 | 0.33 | 0.10 | 0.42 | 1.859 | 23% |
| HLT | 1.00 | 0.20 | 0.07 | 0.56 | 1.826 | 31% |
| RET | 1.00 | 0.21 | 0.07 | 0.44 | 1.720 | 26% |
| FIN | 1.00 | 0.20 | 0.07 | 0.56 | 1.827 | 31% |
| CON | 1.00 | 0.30 | 0.09 | 0.49 | 1.887 | 26% |
How much of the answer is the project and how much is the method, what the biggest assumption rests on, and what else the same evidence would support.
For a sceptical reader, or the analyst hired to argue the other side. Parts of it are technical.
19 observations were governed for this study — checked for source, vintage and consequence before any of them was allowed to affect a number, and 13 things a study of this kind would normally rest on are absent. Each entry says what it is, what it cost or bought, and where it came from.
A study of this kind would normally rest on these, and this one cannot. Each is an absence the analysis noticed rather than a question nobody asked, and each one bounds what the rest of this document can claim.
| What it is | Value | What it changed | Source |
|---|---|---|---|
| The BLS Quarterly Census of Employment and Wages (QCEW) employment coverage produced no reading for Illustrative Region 2024, and this sensor cannot separate a source that answered badly from a precondition that stopped it before it was asked, so which of the two happened here is not established. Neither one is a finding that the source had nothing to show. It is one of this study's default regional sensors. | — | The regional employment coverage and suppression check did not run, so sector coverage is not established by QCEW. | BLS QCEW annual CSV (2024) |
| The USAspending.gov federal award flow was not read for Illustrative Region 2024: this study did not supply the geography or identifier the source needs, so it was never asked. It is one of this study's default regional sensors. | — | The federal-award sensor cannot test the funding environment without county geography. Grant, match, and federal-support claims still need award-level evidence. | USAspending.gov API (2024) |
| The SAM.gov contract-opportunity pipeline was not read for Illustrative Region 2024: this study did not supply the geography or identifier the source needs, so it was never asked. It is one of this study's default regional sensors. | — | The SAM.gov opportunity sensor cannot test the federal procurement pipeline without a state or ZIP. Contract, solicitation, set-aside, and procurement-notice claims still need notice-level evidence. | SAM.gov Opportunities API (2024) |
| The EPA Toxics Release Inventory (TRI) facility context was not read for Illustrative Region 2024: this study did not supply the geography or identifier the source needs, so it was never asked. It is one of this study's default regional sensors. | — | The EPA TRI facility sensor cannot check regional environmental-regulatory context without county geography. Site, emissions, cleanup, and permitting claims still need project-specific records. | EPA Envirofacts DMAP REST service (2024) |
| The EIA electricity price series was not read for Illustrative Region 2024: this study did not supply the geography or identifier the source needs, so it was never asked. It is one of this study's default regional sensors. | — | The EIA electricity sensor is state-level, so it cannot test power-cost exposure without a resolved state. Energy-sensitive operating claims still need a tariff or utility bill. | EIA Open Data API (2024) |
| The NOAA Climate Data Online weather record was not read for Illustrative Region 2024: this study did not supply the geography or identifier the source needs, so it was never asked. It is one of this study's default regional sensors. | — | The NOAA climate sensor could not form a location query. Weather, heat, precipitation, resilience, insurance, and seasonality claims still need dated climate or engineering support. | NOAA Climate Data Online API (2024) |
| The FRED macro financing series was asked for Illustrative Region 2024 and did not come back with a checkable answer, so what it would have shown is unknown rather than nothing. It is one of this study's default regional sensors. | — | The FRED macro financing sensor did not produce a checkable result. Interest-rate, inflation, escalation, and capital-stack claims still need dated source support. | FRED API (2024) |
| The USDA NASS Quick Stats agricultural land-value series was not read for Illustrative Region 2024: this study did not supply the geography or identifier the source needs, so it was never asked. It is one of this study's default regional sensors. | — | The NASS land-value sensor cannot form a county query. Parcel value, acquisition downside, alternative land use, and highest-and-best-use claims still need appraisal or project-specific land evidence. | USDA NASS Quick Stats API (2024) |
| The BLS Local Area Unemployment Statistics (LAUS) series was not read for Illustrative Region 2024: this study did not supply the geography or identifier the source needs, so it was never asked. It is one of this study's default regional sensors. | — | The labor-market sensor cannot test whether the jobs claim lands in a tightening or seasonal labor market. | BLS LAUS API (2024) |
| The BEA transfer-receipt share of personal income was not read for Illustrative Region 2024: this study did not supply the geography or identifier the source needs, so it was never asked. It is one of this study's default regional sensors. | — | The regional-income sensor cannot test transfer dependence or commuting leakage without county geography. | BEA Regional API (2024) |
| The American Community Survey county poverty rate was not read for Illustrative Region 2024: this study did not supply the geography or identifier the source needs, so it was never asked. It is one of this study's default regional sensors. | — | The ACS profile sensor could not form one county query from the supplied geography; a multi-state regional profile needs an explicit aggregation policy. | Census ACS API (2024) |
| The American Community Survey share of housing units available was not read for Illustrative Region 2024: this study did not supply the geography or identifier the source needs, so it was never asked. It is one of this study's default regional sensors. | — | The housing sensor could not form one county query from the supplied geography; multi-state housing capacity needs an explicit aggregation policy. | Census ACS API (2024) |
| The HUD USER housing-market data was not read for Illustrative Region 2024: this study did not supply the geography or identifier the source needs, so it was never asked. It is one of this study's default regional sensors. | — | The HUD User sensor could not form a FMR, Income Limits, HUD-USPS crosswalk, or USPS NCWM query. Rent, income-limit, affordability, ZIP-geography, and address-vacancy claims still need dated public support. | HUD User dataset APIs (2024) |
This table is placed before the findings so that a reader meets it before forming a view.
What is withheld. Every claim that would have rested on the absent evidence. Where a figure could only be established by something on that list, this study reports the limit instead of the figure.
Why. The engine adjudicates each expected item before any of it reaches a number, and it records the ones it looked for and did not find. Each is an absence the analysis noticed rather than a question nobody asked, and what each one costs the conclusion is printed in its own row.
What can still be used. Every observation that was found, checked and used, with its source, vintage and consequence — the larger part of the ledger, and the part the figures in this report actually rest on. An absence bounds what may be concluded from those figures. It does not withdraw them.
What would clear it. Supplying the named items. The list is enumerated rather than described and each row says what it would decide, so this is a collection task with a scope rather than an open research question.
What this does not mean. That the study is unsound, or that the absent evidence would have argued against the project. Nothing was measured in either direction. A study that lists what it could not see is showing its own bound, which is the opposite of a study that had the same gaps and printed none of them.
Published figures the arithmetic rests on directly. If one of these is wrong, the numbers downstream of it are wrong by the same proportion.
| What it is | Value | What it changed | Source |
|---|---|---|---|
| regional accounts rests on ILLUSTRATIVE — not a finding. | — | Sets the headline evidence tier floor; the gate cannot publish the affected numbers as findings. | Project Red Team synthetic region generator — structural placeholder, not an observed economy (2024) |
| Leontief total requirements rests on ILLUSTRATIVE — not a finding. | — | Sets the headline evidence tier floor; the gate cannot publish the affected numbers as findings. | Leontief inverse of the regional direct-requirements matrix |
| induced effects rests on ILLUSTRATIVE — not a finding. | — | Sets the headline evidence tier floor; the gate cannot publish the affected numbers as findings. | household closure at the accounts-implied consumption rate |
| 16 model inputs for purchaser-to-producer price conversion rest on ILLUSTRATIVE — not a finding. | — | Sets the headline evidence tier floor; the gate cannot publish the affected numbers as findings. | Project Red Team illustrative margin rates — replace with BEA supply-table margins before publication |
| 2 model inputs for tax revenue rest on ILLUSTRATIVE — not a finding. | — | Sets the headline evidence tier floor; the gate cannot publish the affected numbers as findings. | Project Red Team illustrative rates — replace with jurisdiction actuals |
| occupational composition rests on ILLUSTRATIVE — not a finding. | — | Sets the headline evidence tier floor; the gate cannot publish the affected numbers as findings. | Project Red Team illustrative staffing patterns — replace with BLS OES industry-occupation matrix for the study region |
Every figure above rests on choices an analyst made: which closure to use, how to regionalize a national table, what to assume about coefficients nobody measured for this region. Each choice is defensible and each moves the answer. This section reports how far, so the precision of the headline is not overstated.
Across defensible methodological choices the output-weighted average multiplier ranges 1.321 to 1.809 — a 27% spread that owes to the analyst's choice of method rather than to anything about the project. Of that spread, the closure alone accounts for 26 of the 27 percentage points.
| Alternative | Result | Change | Why it is defensible |
|---|---|---|---|
| Type I (open model — direct and indirect only) | 1.321 | -26.3% | Open model. No induced effects claimed at all — the conservative floor and the hardest number to argue with. |
| Alternative | Result | Change | Why it is defensible |
|---|---|---|---|
| Coefficients low (5th percentile, 15% CV) | 1.772 | -1.1% | The coefficient matrix is itself an estimate; this is the low end of its plausible range. |
| Coefficients high (95th percentile, 15% CV) | 1.809 | 1.0% | The high end of the coefficient matrix's plausible range. |
Spread attributable to method is not the same kind of uncertainty as spread attributable to the project. A project risk can be managed — negotiate the local-content share, verify the attendance forecast. Method spread cannot be reduced from the client's side; it is a property of the state of the art, and the only honest response is to disclose it and to report the conservative end of this method spread alongside the headline. That end is a floor over the choices an analyst could have made. It is a different construction from a percentile of a combined uncertainty range, and the two are not interchangeable.
The 90% range for output-weighted average multiplier runs 1.773 to 1.81, centred on 1.81 — a width of 2% of that centre, of which 0% is the project's own parameters the client can act on and the rest is method and measurement.
| Source of uncertainty | Share of the variance | What it means for you |
|---|---|---|
| measurement | 100% | how well the underlying data measures this region. Reducible only by buying better data, and sometimes not then |
Read that column as variance, which is not the same thing as width. Variance shares are the ones that add up; the widths they stand for do not, because a width goes as the square root of a variance. A source carrying 50% of the variance carries about 71% of the width, and one carrying 25% carries about 50% of it. The gap runs one way: every figure in that column understates how far its source moves the answer, and it understates the largest source by the most.
The range above assumes the sources vary independently, which is the conventional combination and the narrow end. If they move together instead — and they tend to, because the same analyst chooses the method and the assumptions under the same pressure — the range is about the same width — 2% of the central estimate against 2%, with the optimistic end at 1.808 rather than 1.810.
Nobody has measured the dependence for this study. Both bounds are shown because picking one would be taking a position on a question the evidence does not settle, and the wider one is the relevant bound for a downside decision.
A conventional impact study picks one explanation for what a project does and quantifies it. Every serious objection a reviewer raises is really a claim that a different explanation applies. Those explanations are named here and weighed against the evidence rather than assumed away.
No structured evidence was registered for this study, so the competing explanations remain unweighed. The favorable explanation is assumed rather than established.
This practice has not yet closed the loop on a forecast — no projection has been checked against what actually happened. Results here carry the uncertainty of an unvalidated method, and this sentence will remain in our reports until that changes.
The supplied table method has no completed forecast-versus-outcome record yet. Results carry the uncertainty of an unvalidated method.
Three readings of the same study, each out of 100. Analysis support: 100 is an analysis that argues for going ahead on every count and is settled about it; 0 is an analysis that argues for nothing, or cannot hold a direction. Evidence support: 100 is an evidence base that is uncontested, uncontaminated, and settled; 0 is an evidence base that supports no recommendation in either direction. Track-record support: 100 is a practice with a validated forecasting record on this method; 0 is a practice with no record on this method at all. The gap between them is the finding rather than any one level: the three measure different things and are not required to agree. All three are reviewer diagnostics rather than board figures, and none of them is comparable with the readiness or case figures in Part IV.
Narrative quarantined — evidence too contested. The evidence base is dominated by interested or placeholder sources. The recommendation is withheld until it is replaced — not because the answer would be unfavorable, but because no answer derived from this evidence would mean anything.
What is withheld. Any statement that further measurement would or would not pay for itself here. The section that normally prices each candidate measurement against the decision value at stake did not run, so this study makes no such comparison at all.
Why. A collection plan reached this report with no measurements in it. That is a missing input rather than a computed result: the engine prices what it is given, and it was given nothing. A study with an empty plan and a study whose plan priced everything and found nothing worth buying reach this section by different routes and mean opposite things, and the difference is invisible unless it is stated.
What can still be used. Every figure above it. The pricing question is separable from the analysis -- it asks what would be worth spending to reduce the uncertainty the study already reports, and the study reports that uncertainty either way, in the range, the variance shares and the sensitivity tables.
What would clear it. A collection plan naming the measurements under consideration, each with a method and an estimated cost. The decision value is computed here; the price is not something this analysis can supply for itself.
What this does not mean. That the study has been checked for what more evidence would be worth and found it not worth buying. Nothing was compared. An unpriced list is not a list of things ruled out.
What the analysis supports, what it does not, and what would have to change before a recommendation could be made at all.
The recommendation, what it rests on, and what would have to change. Written to be read alongside Part I.
No recommendation is stated anywhere in this part. Any figure below that reads as one is not one, and the absence is a decision rather than an omission.
Release status is this document's ruling on itself and the first thing in this part to read. It is not a score. At best it is an artifact that is release-safe for the scope stated on it; at worst, an artifact that is not release-safe, with the blocking reason printed beside it. This study is Provisional only — must be stamped throughout.
PROVISIONAL — STRUCTURAL DEMONSTRATION ONLY. The figures in this document rest on illustrative or incomplete inputs. They show what the analysis produces, not what the economy does. Not for distribution, citation, or decision-making.
What put it there:
1 categorical stop was recorded as well. A categorical stop is not a low score: it withholds the recommendation whatever the arithmetic underneath it says, and each one is named where the cascade is set out below.
This document prints several figures out of 100, in this part and in Part III. They are not on one scale, they measure different things, and they are not required to agree — a study can be ready to decide on evidence a reviewer would argue with, and a case can survive the evidence and still be stopped by something about the practice that produced it. The release status above is the verdict, and it is not a number.
| Figure | What it measures | Top of its scale | Written for |
|---|---|---|---|
| Release status | Whether the artifact may leave the building, and on what scope. | It is a ruling, not a score | It is written for a board. |
| Decision readiness | How much of the plausible range clears the declared criterion, how far the expected case sits from the threshold, and what evidence tier the base tables carry. | 100 is a criterion that holds across the whole plausible range, with the expected case sitting a clear two standard deviations from the threshold, on published regional data | It is written for a board to read beside its criterion and the top risks, and it is not the verdict. |
| Epistemic quality | The quality of the evidence base, not whether the project is good. | 100 is every assumption resting on observed published evidence, impact tables built from published regional data, and a decision metric that barely moves across the plausible range | It is a reviewer's diagnostic rather than a board figure. |
| Case strength | How much recommendation authority survives the governance cascade. | 100 is a case with every governance layer assessed and none finding against it | It is a technical reviewer's diagnostic rather than a board figure. |
| Analysis support | How strongly the analysis itself argues for action. | 100 is an analysis that argues for going ahead on every count and is settled about it | It is a reviewer's diagnostic rather than a board figure. |
| Evidence support | Whether the evidence base supports the analysis. | 100 is an evidence base that is uncontested, uncontaminated, and settled | It is a reviewer's diagnostic rather than a board figure. |
| Track-record support | Whether the practice has a validated record on this method. | 100 is a practice with a validated forecasting record on this method | It is a reviewer's diagnostic rather than a board figure. |
| Divergence | How far the analysis argues past the evidence and how far the trust signals sit apart. | 100 is an analysis arguing far past what its evidence supports while the two trust signals contradict each other | It is a reviewer's diagnostic rather than a board figure. |
A conventional study presents a conclusion. This one presents the arithmetic behind it. The analysis opens at the strength the evidence supports, and every governance layer in turn takes what it is entitled to. No layer can add strength back.
Two kinds of layer, kept deliberately apart. Layers about the project — the evidence, the structure of the case, what the region can absorb — move the score. Layers about the analyst — this firm's forecast accuracy, whether its method is proven, how much other work it is carrying — are measured and reported but never scored, because none of them change what will happen at the site. They govern whether the firm should be speaking at all, and they can stop a recommendation outright.
WITHHOLD — the case scores 30 out of 100, where 100 is a case with every governance layer assessed and none finding against it, and 0 is a case with no admissible evidence behind it. At this level the recommendation is not supportable in either direction: the analysis can be published, the recommendation cannot. The evidence opened the case at 33 and the governance layers took 3 off it. It is a technical reviewer's diagnostic rather than a board figure. It measures how much recommendation authority survives the governance cascade. Weaker than 90% of a synthetic reference population (provisional) — a reference set, not a list of real comparable projects.
The layer that found most against this case was evidence quality (finding 0.45). 5 layers were not assessed (forecast record, regional absorption, lens disagreement, practice capacity, claim strength); an unassessed layer is not a layer that passed. A recommendation is withheld regardless of strength: arbitration says the analysis may be published but no recommendation may be drawn from it. Publish the analysis and say plainly what would have to change.
Finding is what a layer concluded on its own, and it is a multiplier, so it runs the other way from its name: 1.00 is a layer that looked and found nothing against the case, and 0.00 is a layer that found everything it could find against it. Lower is worse. A finding of 0.60 means the layer saw enough to cut the case by forty percent. Cost is what that took off the recommendation, which is always less, because these are not independent lines of evidence but overlapping views of one situation and applying each at full force counts the same trouble repeatedly. Left is the recommendation strength still standing after that layer, on the same scale as the headline score above but written as a decimal: 1.00 is a case with every governance layer assessed and none finding against it, and 0.00 is a case with no admissible evidence behind it.
| Layer | Finding | Cost | Left | Why |
|---|---|---|---|---|
| evidence quality | 0.45 | −0.015 | 0.315 | evidence carries 100% contamination or internal contradiction; the leading explanation is undetermined |
| case structure | 0.75 | −0.005 | 0.310 | structure holds; weakest leg is delivery |
| regional absorption (not assessed) | 0.88 | −0.002 | 0.308 | regional labor and capacity conditions were not supplied, so the model's no-supply-constraint assumption stands unchallenged |
| cross-view arbitration | 0.78 | −0.005 | 0.303 | the information, reality and action views arbitrate to 'the analysis may be published but no recommendation may be drawn from it' |
| lens disagreement (not assessed) | 0.88 | −0.002 | 0.301 | stakeholder lenses were not run, so it is unknown whether this verdict survives a lender's weighting or only the sponsor's |
| assessment coverage | 0.60 | — | 0.301 | 3 of 5 scoring layers were assessed, so confidence is capped at 60%. Examining fewer layers cannot produce a stronger recommendation. |
Each of these is a real finding and none of them is evidence about the project. Scoring them was tested and made the recommendation measurably worse at ranking projects — a firm's forecasting record says a great deal about how much to trust the study and nothing about what the site will produce. They are reported here, and any one of them can withhold a recommendation outright.
Finding is what a layer concluded on its own, and it is a multiplier, so it runs the other way from its name: 1.00 is a layer that looked and found nothing against the case, and 0.00 is a layer that found everything it could find against it. Lower is worse. A check reading 0.98 has therefore found almost nothing against the case, and one reading 0.10 has found nearly everything it could — which is the opposite of the way a column headed “finding” reads on sight.
| Check | Finding | What it says |
|---|---|---|
| forecast record (not assessed) | 0.88 | no closed studies in the register yet, so this practice has no measured accuracy to discount by. That is not the same as no expected bias: across 210 transport projects the published record is systematic overestimation that did not improve over thirty years (Flyvbjerg et al. 2005). Unmeasured is not unbiased. |
| method proof | 0.66 | proof confidence for this method is 0.14 of 1.00, where 1.00 is a method checked against its own recorded outcomes and 0.00 is one never checked; no baseline recorded, so drift could not be checked |
| practice capacity (not assessed) | 0.88 | the rest of the firm's book was not federated into this study |
| claim strength (not assessed) | 0.88 | the claim-strength check was not run, so nothing capped how strongly this study states its conclusions |
These three are asked of the analysis and not of the project, and an analysis can pass one of them while failing another. Whether the region gains, whether the operator services its debt, whether a contributing jurisdiction recovers what it put in, and what the land is worth if an entitlement never arrives are four separate questions about four different parties; none of them is scored here, and a figure that answers one of them is not evidence about another. What these three establish is whether the analysis that answers them holds together. They are scored by coupling each pair of analyses and measuring what survives between them, rather than by averaging them into a single figure that would hide the disagreement.
Each score runs 0 to 1: 1 is every pairing inside that group agreeing, each side well supported, with what they establish reaching the conclusion intact; 0 is pairings that jointly establish nothing, because they do not overlap or because they contradict each other. They are calibrated against each other and against nothing outside this study, so the reading they support is which of the three is weakest, not whether any one of them is a passing mark. This is not the recommendation-strength scale used above: the two are different quantities that happen to share a range, and a figure here cannot be set beside one there.
| Question | Score |
|---|---|
| Is the modeled effect real? (weakest) | 0.24 |
| Will it actually land in this region? | 0.49 |
| Is the practice entitled to say so? | 0.26 |
The full withholding — what it covers, what may still be used, and what would lift it — is stated once, in the limits section under Required disclosures. The reasons above are what triggered it on this run.
No purchase changes this while the recommendation is withheld. Arbitration says the analysis may be published but no recommendation may be drawn from it. Those are categorical: clear them first, and only then is it worth asking what evidence would cost.
Moves it is the gain in recommendation strength, on the same scale as the headline score written as a decimal: 1.00 is a case with every governance layer assessed and none finding against it, and 0.00 is a case with no admissible evidence behind it. Per $1k is that gain per thousand dollars spent.
| Do this | Cost | Moves it | Per $1k | Why |
|---|---|---|---|---|
| better evidence | — | +0.699 | — | The evidence convergence would need to rise from 0.24 to 1.00. No collection plan was supplied, so this is unpriced — run `the value-of-information step` with a table of what resolving each assumption is worth to cost it. |
These deductions are facts about the practice rather than gaps in the evidence, and no amount spent on this engagement moves them. Saying so plainly is more use than quoting a price that would not work.
For orientation only. If each governance layer were resolved perfectly — which is not something anyone can choose — the largest would move the recommendation by 0.009 of the 1.00 the score runs to (evidence quality). That identifies the binding constraint; it is not a course of action.
Coverage. 5 governance layers were not assessed in this study: forecast record, regional absorption, lens disagreement, practice capacity, claim strength. Each is charged against the recommendation rather than passed over, because an unexamined layer is not a layer that passed — but the honest statement is that this study does not know what they would have found.
What this study does not establish, how the region's table was built, and every source behind it.
Written for a reviewer reproducing or attacking the numbers. It is deliberately technical, and a decision-maker does not need it to use Parts I and IV.
Every input-output model rests on the same structural assumptions, and they are the first thing a competent reviewer will test. Stated plainly:
| Horizon | Closed forecasts | Mean accuracy | Optimism bias | Evidenced |
|---|---|---|---|---|
| 0-3 years | 0 | — | — | not enough closed forecasts |
| 4-7 years | 0 | — | — | not enough closed forecasts |
| 8+ years | 0 | — | — | not enough closed forecasts |
This study projects 12 years. This practice has no closed forecast at that horizon, so its accuracy there is unmeasured — not good, not bad, unknown. The nearer-horizon record above is the most that can honestly be said, and it is evidence about a different question.
These quantities set the evidence floor for the whole study, because a tier propagates as its weakest input rather than its average: Leontief total requirements, induced effects, occupational composition, purchaser-to-producer price conversion, regional accounts, tax revenue.
Replacing them is what would raise the tier. Nothing else will, however much else is strengthened.
What is withheld. A recommendation. Nothing here tells anyone to proceed, to decline, to defer or to proceed on conditions, and any sentence that reads as advice is a description of what the analysis found rather than a course of action.
Why. The credibility gate reviewed this study and refused to let a verdict be stated. A gate fails on what the document's own evidence does not carry, which is a different question from whether the conclusion is likely to be right, and each failing gate prints the evidence it looked for and did not find.
What can still be used. Every figure in this document at the tier printed beside it, and every finding the analysis reached, including the decision layer's own reasoning where it is shown. The withholding removes one sentence from the study. It does not remove the study.
What would clear it. Resolving the gate's findings and re-running the review. Each finding names the evidence it wanted and the consequence of not having it, so the list is finite and its cost can be estimated before the work starts.
What this does not mean. That the answer is no. A withheld recommendation is not a decline, and the reader who fills the silence with the view they walked in with has been misled by an absence this document exists to prevent. What the review established is that the evidence does not yet support advice in either direction.
Every criticism this document makes of other studies was run against the report body before this section was added. What follows is what that found, and what it did not establish. It is here because a standard applied to strangers and not to ourselves is marketing.
Of the 23 checks in the standard, 3 computed a finding from this document's own contents, 13 could establish no more than that the subject is raised somewhere in it, and 7 cannot be settled by reading a finished document at all. Those are three different results and they are not added together anywhere in this section.
| Result | Checks quoted from the vendor's own guidance | Checks the field never wrote down | What the result licenses a reader to say |
|---|---|---|---|
| Measured | 0 | 3 | The check computed something from the document's own contents and the result stands on its own arithmetic. It is a finding about the document, not about the model behind it. |
| Addressed | 6 | 7 | The document contains language on the subject. Whether it handles the subject competently is not established here and is left to a reader. Addressed is not a pass; reporting it as one would clear the exact studies this standard was built to catch. |
| Not matched | 0 | 0 | A keyword lexicon found no language on the subject. That is a statement about our search rather than about the study: it means we did not find it, never that the study does not do it. |
| Not run | 5 | 2 | The check cannot be settled from a finished document; it needs the inputs and a reader. A not-run check and a passed check are opposite results, and a checklist that quietly drops its hard items reads as a clean bill of health. |
The addressed and not-matched columns are keyword-lexicon results rather than a reader's judgment, and the lexicons have a measured error: pointed at the one document we could check line by line, 3 of 12 presence checks returned a false negative, and every one of those errors ran in the direction that accuses.
Each of these appears in prose and in no table in this document. That is not necessarily an error - a figure can be narrated once and legitimately never tabulated - but a reader cannot check it against anything, so it is listed rather than left for them to notice.
| Figure | Where it is stated | Nearest tabulated figure |
|---|---|---|
| $4.0M | Whether a $4.0M county contribution is … | $3.8M |
| $2.5M | … spending the region would have seen anyway, so $2.5M of gross output was removed … | $2.3M |
| $11.2bn | That report headlined an $11.2bn increase in the size of a … | $1.82B |
| $6.96bn | … state's economy where its own value added was $6.96bn, a 61% overstatement produced … | $1.82B |
| $41,199 | … averages 10% below the regional median wage ($41,199 job-weighted average wage … | $39,620 |
| $46,000 | … ($41,199 job-weighted average wage against a $46,000 median). | $48,583 |
| $372.8M | … would have come out 3.4% higher — $372.8M against $360.7M. | $360.7M |
| $1k | Per $1k is that gain per thousand … | $6,566 |
This document carries 246 qualifying sentences, and 1 section that qualifies the whole document rather than a single claim: what was not examined.
Measured on the shipped audit predicate (rules/exemplar scorer plus the learned detector), on 3 held-out studies chosen by document hash: finds 76% of labelled qualifications. On the representative held-out sample it flagged 14 sentences and 6 were false positives, so about 4 in 10 flags were not qualifications; pooling the exhaustive positive-find pass with that sample gives 82% precision, which is useful but too flattering as a field-rate estimate. Observed document recalls ranged 67%-87%, but the intervals overlap; current evidence does not prove recall varies by house style. Cross-study comparison therefore needs the same extraction basis and a rank-stability check, not just raw counts.
That figure is therefore approximate, and the two errors do not cancel. A qualification phrased in a register the detector has not learned is not counted, and some sentences it does flag are not qualifications; at the measured rates, a hundred sentences containing five real qualifications return about 3.8 found and about 2.9 false, so a count published this way is more likely to be high than low. Apply the same limit before comparing any study we audit.
What that permits. This count may be compared with another study's only where three things hold: both were extracted the same way, the detector’s error is printed beside both, and the ordering survives resampling. Where the ordering does not survive, we publish the range and the reason rather than a rank. On the held-out set the exact ordering held only 30% of the time, so a bare league table of studies is not something this measurement supports.
What this does not mean. That the count is unusable, or that this document is padded with caveats. The detector's error is measured rather than suspected, and a figure carrying its measured error is a stronger claim than a bare one — it is the disclosure this standard asks of every study it is pointed at, made about ourselves first.
What is withheld. Any result at all for 7 checks. Not a pass, not a failure, and no count of them enters a total anywhere in this section.
Why. Each needs something a finished document does not contain — the model run behind a table, the survey instrument, or a reader's judgment about whether a subject was handled competently — and reading the document harder supplies none of it. The reason is printed per check below.
What can still be used. The 16 checks that did produce a result, above, each at the strength its own result claims — the results table says what every one of them licenses a reader to say. Nothing in this list retracts any of them; it is the boundary around them.
What would clear it. The inputs each check names, and a person to read them. The standard was written knowing a finished document could not answer these, so a review holding the model run is the designed answer here and not a fallback.
What this does not mean. That these checks came back clean, or that their absence is evidence the answer is favorable. A check that did not run and a check that passed are opposite results, and they are labeled differently everywhere in this pack.
| The question | Why it could not be run |
|---|---|
| Is a job that lasts several years counted once, or once per year? | a finished document cannot settle this; it needs the inputs and a reader |
| Did anyone establish that the project will work before modelling what it would produce? | a finished document cannot settle this; it needs the inputs and a reader |
| Is the region the study reports on the region the decision is about? | a finished document cannot settle this; it needs the inputs and a reader |
| Is the model's data from the same period as the spending it is being asked about? | a finished document cannot settle this; it needs the inputs and a reader |
| Can the region actually supply the labour and materials the projection assumes? | a finished document cannot settle this; it needs the inputs and a reader |
| Does it call output the size of the economy? | a finished document cannot settle this; it needs the inputs and a reader |
| Can each headline number be traced to the input that produced it? | a finished document cannot settle this; it needs the inputs and a reader |
Everything below ran, and nothing below was established. A self-audit that lists what it found and never says what it never looked at is the same document as one with no limits section, and this standard treats that as a material failure in other people's studies.
None of it retracts a finding above it. The checks that measured something measured it, and this is the edge of what they cover rather than a qualification of what they found.
| What this audit did not measure | Why not, and what it would take |
|---|---|
| Whether a subject this document raises is handled competently | 13 checks came back addressed, which means language on the subject was found and nothing more. Judging the treatment takes a reader, and this section does not attempt it. |
| The disciplines that need the inputs rather than the document | 7 checks cannot be settled from a finished document at all, and are named above. Nothing here should be read as those checks having come back clean. |
| This section itself | The audit target is the report body as it stood before this appendix was appended. An audit of the audit would recurse, so no claim on this page has been checked by the code that wrote it. |
| Whether the study's arithmetic is right | This layer reads a rendered document. It does not re-run the model and it cannot tell a defensible coefficient from an indefensible one. The engine's own identity tests do that, and they are a different body of evidence. |
| The recall of these word lists on any document but this one | Their measured error - three misses in twelve checks - comes from the one document we could read line by line, which is this one. Each miss was repaired by widening a list, and that is the repair that does not generalize: the next document phrases it a fourth way. |
| Whether the detector finds caveats at the same rate in every document | Three held-out documents returned 90%, 70% and 67%. Their intervals overlap and a shared-recall bootstrap returns p = 0.69, so this denominator settles the question in neither direction. Distinguishing 67% from 87% would take roughly 69 labeled positives per document. |
What is withheld. There is deliberately no score here. Nothing in this section compresses these findings into a single number, and the same refusal applies to every study this standard is pointed at.
Why. A study either keeps a discipline or it does not, and compressing that into a grade would produce a number whose first use is marketing and whose second is a threshold nobody can defend.
What can still be used. The page above, unaggregated: which checks measured something, what each one found, which matched no language, and which could not run. A reader who wants the short version reads the results table, which is four numbers and a sentence saying what each one licenses them to claim.
What would clear it. Nothing supplied later closes this. It is the one refusal in the document that is not waiting on evidence, because what is missing is a scale rather than a fact, and there is no defensible one to supply.
What this does not mean. That the document declines to be judged, or that a grade was computed and held back. Every finding this standard can make about this report is printed above. What is refused is the compression, not the judgment.
It was not built here. The accounts behind this study arrived already regional and were used as supplied. No location quotients, purchase coefficients or balancing were applied to them, so none of the choices those steps require were made by this study, and the method spread reported in Part III carries no regionalization variants — a study that never made that choice should not be shown a range around it.
Every quantity in this report, the method that produced it, and the evidence it rests on. This appendix exists so that a reviewer can check the work rather than take it on trust.
Entries below marked decision layer are Project Red Team decision-layer output applied to the base analysis, not standard input-output output. A reviewer reproducing the federal method will reproduce the impact tables and will not reproduce those. It is said once here rather than under each of them.