The LLM only reports what the claim says
Stage one pulls the stated revenue, timeline, tools, and costs out of the source. Missing data stays null — it is never invented or estimated by the model.
← BACK TO INDEX
EXTRACTED
DETERMINISTIC
PUBLISHED
Source Intake
LLM Extraction
Price Table Lookup
Five-Axis Scoring
Hype Penalties
Hard-Fail Overrides
Verdict & Publication
VERIFIED
MODELED
UNVERIFIED
WHY IT HOLDS UP
THE CHALLENGE PROCESS
THE METHODOLOGY
How RealityScore™ scores a claim
Two stages, strictly separated. First, an LLM extracts what the claim actually says — revenue, timeline, stack, costs — and preserves nulls where the claim is silent. Second, deterministic arithmetic scores those numbers against version-stamped price tables, platform payout latency, and market price bands. The model never scores. The math never guesses.
The score is pure arithmetic
Stage two runs fixed formulas over version-stamped price tables, payout latency data, and market price bands. Same inputs, same score, every time.
Every teardown shows the workings
The extracted inputs, the table versions used, and the arithmetic behind each axis are published with the verdict. Anyone can recompute the score.
PROCESS FLOW
From source claim to published verdict
The pipeline keeps interpretation and judgment strictly apart: the model reads, the tables price, the arithmetic decides.
Public posts from X, Reddit, and YouTube are normalized into one claim packet with URL, source text, and metadata.
The model extracts stated revenue, timeline, workload, stack, and costs. Anything the claim does not state stays null.
Every named tool and API is priced from version-stamped tables, alongside platform payout latency and market price bands.
Deterministic formulas score Temporal Possibility, Workload Cost, Stack Completeness, Price Realism, and Net Margin.
Documented deductions (max −15) for guru-marketing patterns: urgency language, income promises, upsell anchoring.
Claims that are arithmetically impossible — payouts faster than the platform pays, negative margin at stated prices — go straight to DEBUNKED.
The score lands in a published band and the full workings — inputs, table versions, arithmetic — ship with the teardown.
TRUST MODEL
Every number carries an honesty label
Nothing in a teardown or replication audit gets to sound more certain than it is. Each figure is tagged with how it was obtained.
Probed directly
Confirmed with a cheap real-world probe: an actual API call, a live price check, a signup we ran ourselves.
Computed from price tables
Derived arithmetically from version-stamped price tables, payout latency data, and market price bands. Reproducible, but not directly observed.
Stated, not checked
The claim asserts it and nothing in our tables can confirm or price it. Labeled plainly so it never masquerades as evidence.
HOW THE MODEL WORKS
A claim does not get a score by vibe.
Stage one extracts. Stage two computes. The score is 100 points across five axes, minus documented hype penalties, subject to hard-fail overrides. Nothing in the pipeline exercises judgment after extraction.
STAGE 01
Extraction, with nulls preserved
An LLM reads the source and records exactly what the claim states: revenue, timeline, workload, stack, prices. Where the claim is silent, the field stays null. The model never fills gaps and never scores.
STAGE 02
Deterministic scoring over version-stamped tables
Fixed formulas price the extracted claim against real API and tool costs, platform payout latency, and market price bands, then apply hype penalties (max −15) and hard-fail overrides.
VERDICT
Published bands, published workings
FEASIBLE at 75+, STRAINED at 45–74, IMPLAUSIBLE at 20–44, DEBUNKED below 20 or on any hard fail. The full arithmetic ships with every teardown.
The score is constrained on purpose
Determinism is the trust layer: same extraction schema, same table versions, same formulas, same bands. Two people running the model on the same claim get the same number.
- The LLM extracts; it never scores or fills gaps
- Version-stamped price tables, published weights, documented penalties
- Every figure labeled VERIFIED, MODELED, or UNVERIFIED
Scores can be challenged, and they move
Every score is an opinion based on this published methodology and the disclosed math. If a creator — or anyone — supplies better public evidence, corrected costs, or a table error, we rerun the same arithmetic against the new inputs and republish. Challenge a score here.
Verdict Bands
The final score maps to one of four published verdicts. Bands are fixed; nobody nudges a claim across a boundary.
FEASIBLE75–100
The stated numbers survive the math: the timeline is possible, the stack is priced and complete, and net margin is positive at real prices.
STRAINED45–74
The claim is not impossible, but the math only closes under generous assumptions: thin margins, tight timelines, or missing costs.
IMPLAUSIBLE20–44
Multiple axes fail. The economics require prices, speeds, or workloads well outside published bands.
DEBUNKED<20 or hard fail
The claim is arithmetically impossible on public data, or trips a hard-fail override. These land in The Graveyard.
The Five Axes
The base score is 100 points across five axes. Each axis is a fixed formula over the extracted claim and the version-stamped tables — not a judgment call.
AX-01
20 PTS
Temporal Possibility
Could the stated result physically happen in the stated window? Platform payout latency alone kills many "paid in 48 hours" claims.
AX-02
25 PTS
Workload Cost
The compute, API, and tooling bill for the described workload, priced from version-stamped tables. Claims that ignore their own run costs lose here.
AX-03
15 PTS
Stack Completeness
Does the described stack actually cover every step from input to payout? Missing pieces — hosting, payment rails, distribution — are scored, not assumed.
AX-04
20 PTS
Price Realism
The stated selling price is checked against market price bands for comparable work. Prices far outside the band cost points.
AX-05
20 PTS
Net Margin
Revenue minus every priced cost — tools, fees, platform cuts. The claim earns points only if the arithmetic leaves real margin.
Hype Penalties
After the five axes are scored, documented deductions apply for guru-marketing patterns. Penalties are capped at −15 total, so hype dents a score but never replaces the math.
Course / Upsell Anchoring
penalty
The claim exists primarily to sell a course, community, or tool rather than to document the business itself.
Urgency & Scarcity Language
penalty
"Limited spots," "last chance," "secret method" — pressure patterns that correlate with claims the math cannot support.
Income Promises
penalty
Guaranteed-income or risk-free framing. Real unit economics carry risk; language that denies it is penalized.
All hype deductions combined cannot exceed −15 points. A hyped-up claim with sound math still scores; a calm claim with impossible math still fails.
Hard-Fail Overrides
Some failures are not a matter of degree. If any override trips, the claim is DEBUNKED regardless of its axis score.
IMPOSSIBLE TIMELINE
The claimed payout arrives faster than the platform's documented payout latency allows. No workflow fixes that.
NEGATIVE MARGIN
At real, version-stamped prices, delivering the claim costs more than it earns at the stated price.
NONEXISTENT STACK
The claim depends on a tool, price tier, or capability that does not exist as described in the stamped tables.
Limitations
- Scores are based on publicly available information and published price tables only.
- We do not audit bank accounts, private dashboards, or unpublished receipts.
- Scores can change when evidence or table versions change; teardowns note which versions were used.
- Scores are opinions based on this published methodology and the disclosed math — a research product, not financial or legal advice.
Want the week's verdicts every Friday?
We score the week's loudest AI income claims and send the verdicts — with the math — every Friday.