Key takeaways
- Eleven general Nigerian legal materials formed the added corpus.
- The stages changed loading, summary length and summarisation tool.
- A 70% always-Dismissed majority-class reference applies to Part II accuracy charts.
What was tested & why
Extension design
GLMP means fuller general legal materials; GLMRS means model-constructed general legal materials as relational summaries. The original No Clue and With Clue Nigerian results are baselines counted once, not independent new predictions.
Source: Thesis Chapter 5 · source-reported unless otherwise noted.
Results / visual evidence
Nigerian extension · selected pooled conditions
Thesis Tables 4.1–4.2, 6.7 and 7.2 · 120 model–case observations per condition; baselines are reused, not recounted.
Source tables / thesis transcription
Inspect the evidence
Scroll wide tables horizontally. Source notes and qualifications remain with their tables.
First-stage experimental conditions (simultaneous loading)
Condition Input beyond the Purpose
case facts
No Clue None Original facts-only baseline.
With Clue Case-specific contex- Original case-specific context
tual information baseline.
General Nigerian Legal Ma- Fixed 11-material Tests broader jurisdiction-level
terials No Clue Nigerian legal corpus context without case-specific
clues.
General Nigerian Legal Ma- Fixed 11-material Tests broader context combined
terials With Clue corpus + case- with case-specific context.
specific contextual
information
GLMRS No Clue Relational sum- Tests connected compression of
maries of the fixed the same underlying source
11-material corpus corpus.Thesis Table 5.1 · source transcription; layout preserved for multi-line headings and notes. Verify against the source PDF for formal citation.
Second-stage experimental conditions (materials preloaded before the prediction task)
Condition Preloaded input Purpose
GLMP a No Clue Complete text of the fixed Run 1: fuller materials
11 materials preloaded, facts only.
GLMP a With Clue Same + case-specific clue Run 1: preloaded materials
supplied with the case combined with case-specific
context.
GLMP b No Clue As GLMP a No Clue Rerun: tests stability of the
identical experiment.
GLMP b With Clue As GLMP a With Clue Rerun: tests stability of the
identical experiment.
GLMRS-C No Clue Relational summaries, Tests compressed connected
150–300 words per source context preloaded before the
task.
GLMRS-D No Clue Relational summaries, Same, with a longer permitted
300–500 words per source summary.
GLMRS-E No Clue Relational summaries, Same, with the longest
500–750 words per source permitted summary.
GLMRS-C, -D and -E were originally reported as GLMRS-A, -B and -C. They are relabelled so
that the letters follow summary length (Appendix 12.1).Thesis Table 5.2 · source transcription; layout preserved for multi-line headings and notes. Verify against the source PDF for formal citation.
Third-stage experimental conditions (shorter relational summaries preloaded, six models)
Condition Preloaded input Purpose
GLMRS-A No Clue Relational summaries, 50– Extends the length curve below
100 words per source 150 words (shortest summaries).
GLMRS-B No Clue Relational summaries, Intermediate point between
100–150 words per source 50–100 and 150–300 words.Thesis Table 5.3 · source transcription; layout preserved for multi-line headings and notes. Verify against the source PDF for formal citation.
Fourth-stage experimental conditions (stored summaries from five tools, ChatGPT as the fixed predictor)
Summariser (tool) Lengths tested Role in the comparison
ChatGPT 100–150 and 150–300 Self-summary control: the
words predictor summarises for itself.
Claude 100–150 and 150–300 Cross-model summariser.
words Claude showed the highest
observed verdict agreement
across the two GLMP reruns.
Microsoft Copilot 100–150 and 150–300 Widely available in law-firm
words office environments. Same
OpenAI model family as
ChatGPT in a different product.
Gemini 100–150 and 150–300 Cross-model summariser with
words long-context input.
Gemini Notebook (for- 100–150 and 150–300 Source-grounded summariser
merly NotebookLM) words that answers only from
uploaded documents.
Each of the ten tool–length summary sets was preloaded into ChatGPT, which predicted the
20 cases under No Clue (run a). The ChatGPT- and Claude-summary conditions were rerun
(run b).Thesis Table 5.4 · source transcription; layout preserved for multi-line headings and notes. Verify against the source PDF for formal citation.
General Nigerian legal materials used in the added conditions
No. Legal material Role in the corpus
1 Constitution of the Federal Republic of Nigeria General constitutional framework, including
1999, as amended public-law and fair-hearing principles where relevant.
2 Labour Act, Cap L1 LFN 2004 General Nigerian labour and employment statutory
material.
3 Trade Marks Act General statutory framework for trademark rights,
registration and infringement.
4 Land Use Act 1978 General statutory framework relevant to land
interests and transactions.
5 Banks and Other Financial Institutions Act 2020 General statutory framework governing banking
(BOFIA) institutions and banking regulation.
6 CBN Consumer Protection Framework General Central Bank of Nigeria consumer-protection
and banking conduct framework.
7 Admiralty Jurisdiction Act General statutory framework for admiralty
jurisdiction.
8 Sheriffs and Civil Process Act General statutory framework relevant to civil process
and judgment enforcement.
9 NAFDAC Act General statutory framework for the national food
and drug regulator.
10 Counterfeit and Fake Drugs and Unwholesome General public-health and pharmaceutical statutory
Processed Foods Act material concerning counterfeit/fake drugs and
unwholesome processed foods.
11 Merchant Shipping Act 2007 General statutory material concerning merchant
shipping and maritime matters.Thesis Table 5.5 · source transcription; layout preserved for multi-line headings and notes. Verify against the source PDF for formal citation.