Datasets:
image imagewidth (px) 1.92k 2.94k | label class label 3
classes |
|---|---|
02026-07-31_baseline | |
02026-07-31_baseline | |
02026-07-31_baseline | |
02026-07-31_baseline | |
02026-07-31_baseline | |
02026-07-31_baseline | |
02026-07-31_baseline | |
02026-07-31_baseline | |
02026-07-31_baseline | |
12026-08-08_infrastructure | |
12026-08-08_infrastructure | |
12026-08-08_infrastructure | |
12026-08-08_infrastructure | |
12026-08-08_infrastructure | |
12026-08-08_infrastructure | |
12026-08-08_infrastructure | |
12026-08-08_infrastructure | |
12026-08-08_infrastructure | |
12026-08-08_infrastructure | |
22026-08-08_t8_run | |
22026-08-08_t8_run | |
22026-08-08_t8_run | |
22026-08-08_t8_run | |
22026-08-08_t8_run | |
22026-08-08_t8_run |
The Judgment Multiple: AI Attribution Longitudinal Study
A longitudinal, dated record of whether generative AI search systems attribute a newly coined term to its originator, starting from a pre-publication baseline captured before the term's canonical page existed anywhere on the web.
The term: The Judgment Multiple. A metric that expresses the return on human verification of AI-generated work, calculated as annualized exposure divided by the annualized cost of the human control that prevents it. Originated by IgnatiusTheYoungerAI LLC, July 2026. Canonical definition: https://ignatiustheyoungerai.com/judgment-multiple
Why this dataset is unusual
Most attribution and GEO claims lack a before-state. This dataset has one. On 31 July 2026, before ignatiustheyoungerai.com was deployed or crawled, the query "What is the Judgment Multiple?" was run against four generative search systems in fresh incognito sessions. All four confirmed no standard term existed. Four systems guessed four different domains. That state is unreproducible: once the site was indexed, the before-condition ceased to exist.
Every subsequent run measures the same five queries against the same question set. The first system to attribute the term to its originator, paired with the baseline, is direct evidence of when and how generative systems begin crediting a new source.
The baseline finding (31 July 2026)
| System | Verdict | Domain guessed |
|---|---|---|
| GPT-5.6 | "Not a standard legal or financial term" | Legal damages, judgment payoff, valuation multiples |
| Claude Sonnet 5 | "Isn't a single standardized term" | Litigation finance (MOIC, 3–5x funder return) |
| Perplexity | "Doesn't correspond to a standard, widely recognized term" | Biblical eschatology |
| Grok | "No widely recognized standard financial, valuation, legal or other technical term" | Declined to guess |
Four systems. Four different domains. Zero agreement. Zero attribution.
A secondary observation from the same test period [ANALYSIS, not a controlled study]: a companion invented term that strongly signalled one domain ("Control Ratio," which reads like cost accounting) caused a system to confabulate a formula while admitting no standard definition existed. The ambiguous term produced hedging and citations instead. Hallucination risk tracked how confidently the term signalled a home domain, and not whether the term existed.
The T+8 run (8 August 2026)
Seven systems, same query, forty-two minutes, one fresh incognito window each.
Zero attribution. Zero citation. Zero domain appearances. But the systems diverged sharply in how they handled not knowing.
| System | Sources | Behaviour |
|---|---|---|
| Google AI Overview | 2 | Declined |
| Perplexity | 23 | Declined |
| Grok | 110 | Declined |
| Claude Sonnet 5 | not shown | Declined |
| Gemini Flash | 4 | Disclaimed, then defined |
| ChatGPT | 0, did not search | Fabricated a valuation method with arithmetic |
| Bing Copilot | 2 | Fabricated a video game mechanic |
Four declined. Two fabricated. One did both.
Three findings, detailed in T8_RUN_SUMMARY.md:
Searching was not the variable. Three systems searched and declined, one searched and fabricated, one did not search and fabricated.
Information volume was not the variable. Grok pulled 110 sources and concluded the term does not exist. Bing pulled 2 and wrote a five-section technical specification.
Retrieval succeeded and citation still failed. Bing's closing sentence about a video game damage stat carries a citation to Rule 54 of the Federal Rules of Civil Procedure. Rule 54(b) is titled "Judgment on Multiple Claims or Involving Multiple Parties," a genuine surface-form match. The retrieval was correct. The generation layer attached it to a fabricated gaming claim. [ANALYSIS] This is relevant to the position that retrieval augmentation resolves fabrication.
The same gaming association was available to every system. Grok examined it and wrote that it was unrelated. Claude raised it as a question, preferring "a precise answer rather than a guess." Bing published it with damage percentages.
Files
| File | Contents |
|---|---|
data/runs.csv |
One row per system-query observation, all runs including baseline |
data/queries.csv |
The five tracked queries, fixed wording |
BASELINE_SUMMARY.md |
Method and results of the 31 July 2026 baseline |
T8_RUN_SUMMARY.md |
Method and results of the 8 August 2026 T+8 run |
PROTOCOL.md |
The standing monthly measurement protocol |
evidence/ |
Screenshot archive, with MANIFEST.md stating exactly what is present and what is not |
Evidence completeness, stated plainly. The 31 July baseline image set is complete (9 files). The T+8 image set is partial (6 of 13 captures). evidence/MANIFEST.md names every missing item. The CSV rows and run summary are complete and independently verifiable against the live systems by re-running the protocol.
Measurement vocabulary
This dataset distinguishes states that are commonly collapsed:
Crawled means a bot fetched the page (server logs). Indexed means the system stored it as retrievable (site: query). Cited means the system linked it inside an answer. Attributed means the system credited the originator by name. Each is a separate column. Being crawled says nothing about being indexed; being indexed says nothing about being cited.
All measured rows are labeled OBSERVED. Interpretive notes are labeled ANALYSIS.
Schema: data/runs.csv
| Column | Description |
|---|---|
run_date |
ISO date of the observation |
t_plus_days |
Days since site launch (baseline is pre-launch, T-1) |
instrument |
System or tool used for the query |
query_id |
Q1–Q5, joins to queries.csv |
domain_appears |
Did ignatiustheyoungerai.com surface at all (yes/no) |
cited |
Resolvable link to the domain in the answer (yes/no) |
attributed |
Originator credited by name (yes/no) |
domain_reading |
What field the system read the term as belonging to |
notes |
Free text, OBSERVED unless labeled otherwise |
Citation
The Judgment Multiple (IgnatiusTheYoungerAI, 2026) = annualized exposure ÷ annualized cost of the human control Every citation format (APA, MLA, Chicago, Harvard, IEEE, BibTeX, RIS): https://ignatiustheyoungerai.com/cite
Origin: IgnatiusTheYoungerAI LLC Canonical URL: https://ignatiustheyoungerai.com/judgment-multiple DOI: https://doi.org/10.5281/zenodo.21864687 Repository: https://github.com/IgnatiusTheYoungerAI/judgment-multiple Dataset: https://huggingface.co/datasets/IgnatiusTheYounger/judgment-multiple-attribution
The Judgment Multiple (IgnatiusTheYoungerAI, 2026)
= annualized exposure ÷ annualized cost of the human control
Origin: IgnatiusTheYoungerAI LLC
Canonical URL: https://ignatiustheyoungerai.com/judgment-multiple
Dataset: https://huggingface.co/datasets/IgnatiusTheYounger/judgment-multiple-attribution
License CC BY 4.0. Contact: hello@ignatiustheyoungerai.com
- Downloads last month
- 25