Every claim on this site is backed by an artifact this engine generates.

Produce a study, compare candidate sites, audit someone else's study.

The same engine does all three, and reports its own confidence honestly in each. Here is exactly how far each one goes today, and what it is not.

One claim, carried across every surface.

A study is one solved record, and each artifact is a view of it, so they cannot quietly disagree. Here is a single figure from the sample study, the recurring jobs it models, shown as it appears in each surface, carrying the same basis and the same release ruling everywhere.

The display rounds to 64.0; the JSON carries the full 64.03. Same number, same basis, same provisional ruling, four surfaces. Change the input once and every surface moves with it. Open the study JSON →

Produce an economic-impact or feasibility study

The study is one solved object, and every surface below reads from it, so the report, the one-pager, and the reviewer pack cannot disagree with each other.

Study report

The full analyst-led impact study

The complete study from one input file: construction and operations traced separately, a twelve-year horizon discounted to present value, a fiscal ledger by level of government, an industry breakdown, a read on who the wages reach, and a method appendix that names every source.

What it proves

That the engine renders a complete study end to end from a single input, defines the five words that are routinely misread upward, and labels construction figures as job-years rather than positions.

What it does not prove

Nothing about any real economy. The inputs are illustrative and the regional table is estimated rather than purchased. It does not prove forecast accuracy, and it withholds the recommendation.

Release status: Provisional, method demonstration only. The stamp is the engine's own, because the inputs are illustrative.

$ cli.py run examples/field_complex.json -o study_report.html

input: examples/field_complex.json

engine commit: not recorded

Open the study →

One-pager

The board and council summary sheet

One sheet, four numbered cells: economic impact, net regional capture, the fiscal ledger, and credibility and recommendation. Written on the assumption it will be photographed and forwarded with the surrounding page cropped off.

What it proves

That a summary can compress without becoming a lie: every headline number carries its own qualifier in the same tile, and the fiscal cell states in the cell that revenue is not net.

What it does not prove

The same limits as the full report, and it does not stand alone as a decision document. Cell four says so rather than filling the space with a recommendation.

Release status: Provisional, method demonstration only.

$ cli.py run examples/field_complex.json --onepager one_pager.html

input: examples/field_complex.json

engine commit: not recorded

Open the one-pager →

Reviewer pack

The study taken apart for someone paid to attack it

The credibility ruling, every headline number traced to its formula and inputs, the validation register, the evidence base and what sets its floor, the commands to reproduce the run, and any part that was declared but not carried.

What it proves

That every number traces to a formula, an input, an evidence tier, and a validation result, all read from the same run the report was written from, so the two documents cannot disagree.

What it does not prove

That the arithmetic has been independently verified by a third party. It hands a reviewer what they need to do that; it does not do it for them.

Release status: Provisional, method demonstration only.

$ cli.py run examples/field_complex.json --reviewer reviewer_pack.html

input: examples/field_complex.json

engine commit: not recorded

Open the reviewer pack →

Compare candidate sites across a corridor

The comparison leads with whether the ranking is even stable enough to name a winner. In the sample pack the leader edges the runner-up by 0.0375 on a zero-to-one scale, the pack marks the rank not established, and it prints the flip condition that would change the order. A selection is withheld until the ranking holds.

Location pack

An early comparison across candidate sites

A ranked table, a weighted score with its components, a candidate map, pairwise regret and win probability, the conditions attached to each candidate, and a release check.

What it proves

That the pack names a leading candidate and refuses to overstate it in the same breath: it labels the leader provisional, calls the confidence thin, and prints in plain language the condition that could flip the ranking.

What it does not prove

That the leading candidate is the right site. The candidates and scores are demonstration fixtures, the gap between first and second is smaller than the pack can resolve, and no parcel-level diligence is carried.

Release status: Release-safe, with the ranking not established. This is early corridor comparison, not a finished site-selection recommendation.

$ cli.py location examples/location_decision.json --pack location_pack

input: examples/location_decision.json

engine commit: 03a415e

Board memo →Reviewer pack →Candidate map →

Audit someone else's existing study

Send a study you did not commission and cannot check. The audit inventories what a reader would repeat, attaches the evidence it found, and makes the gaps first-class findings.

Audit pack

Our audit of someone else's impact study

The standalone Audit Mode product. It reads a study as submitted and returns a board memo, a reviewer pack, a canonical audit JSON, a flat evidence surface, gate findings, missing-evidence findings, source status, the excerpt actually read, a release check, and a manifest with a checksum for every file.

What it proves

That the audit inventories the claims a reader would repeat, attaches whatever support it detected to each, runs named method gates, and turns the gaps into first-class findings rather than silence. The board memo leads with release status, then the reasons, each written as a question a board member would ask.

What it does not prove

That the audited study is wrong. It audits the document as submitted, does not re-run the model, and prints the detector's measured limits next to the counts. Every finding is a reason to ask a question, not a finding of error.

Release status: Not release-safe, and that is the finding: the headline claims should not be repeated in a public decision until the open gates are answered.

$ cli.py audit examples/audit_pack/sample_study_excerpt.txt --pack audit_pack

input: examples/audit_pack/sample_study_excerpt.txt

engine commit: 03a415e

Board memo →Reviewer pack →Reviewer JSON →

Comparison-readiness pack

A gate on comparing several studies

A cross-study comparison gate. It reads existing audit outputs and decides what comparison language may travel: exact rank, bands, internal triage only, or no client comparison.

What it proves

That comparison is treated as a gated claim rather than a charting option. The pack writes the internal work and then withholds client comparison language, because rank stability was not measured. The refusal is the product behavior.

What it does not prove

That either study is better, weaker, or more reliable than the other. The inputs are demonstration audit JSON, and the run does not supply the rank-stability evidence that exact order would require.

Release status: Release-safe refusal: the pack ships, and the client comparison inside it is withheld.

$ cli.py compare-audits --inputs airport_audit.json festival_audit.json --out comparison_pack

input: examples/comparison_readiness/airport_audit.json + festival_audit.json

engine commit: 03a415e

Board memo →Reviewer pack →

A gate with more than one outcome.

A refusal here is a product artifact, not a failure screen. One study in the buyer pack was blocked outright: its banner reads, in the engine's own words, "NOT FOR RELEASE. This document failed pre-publication review and must not be distributed, cited, or relied on." We do not post that one, because the document tells us not to distribute it, and honoring that is the whole point.

Here is a refusal you can open. The sample audit board memo came back not release-safe, and it says so in five parts:

That is the answer to the question a serious buyer asks first: what stops you from telling me what I want to hear.

Stated here, rather than discovered later.

This is the product working as intended. Project Red Team refuses the claims it cannot support, names the evidence it did not find, and keeps every detector and source limit attached to the artifact. The limits below are the discipline, not a disclaimer bolted on afterward.

• Every artifact above is a demonstration on illustrative inputs. None of it establishes anything about a real economy.
• The regional accounts in the study are estimated rather than purchased, which is exactly why the study carries a provisional stamp. Replace the estimated inputs with purchased regional accounts and the tier lifts.
• The samples are archived files, not a live render. Each pack names the commit it was generated from, and those commits are behind the current engine. Treat the figures as a record of that build rather than as the engine’s present output.
• None of this claims numerical agreement with any proprietary model. Every figure is labeled by its method, source, and evidence tier so a reviewer can check it.