Entity-conditioned probing with resampling — PHP 7.4 CLI toolkit implementing standardized prompts, structured outputs, k-sampling, and Bradley–Terry aggregation.

• PHP: PHP 7, no jQuery required. CLI only.
• Libraries: guzzlehttp/guzzle, opis/json-schema, league/csv, math-php.
• Data: Final aggregation product (PL scores, frequency metrics, CIs) from 15,600 GPT-5 prompts (No Cache)

This repository provides a transparent demonstration of the RankLens Entities evaluation process used in SEO Vendor’s patent-pending system for large-language-model ranking analysis.

• Goal. Compare PL scores (α=0.2, BT=300) with frequency baselines (@1/@3) to surface category leaders and quality–popularity gaps.
• Coverage. 52 categories; global + per-locale (US/GB/DE/JP).
• Sensitivity. α/BT variants only reshuffle near cutlines → treat as ties.
• Deliverables. CSVs + viewer; see Data.

• adjustable_dumbbells
• alpine_boots
• base_layers
• basketball_apparel
• basketball_hoops
• basketball_shoes
• compression_leggings
• compression_shorts
• creatine_monohydrate
• cross_training_shoes
• down_jackets
• ellipticals
• fitness_trackers
• foam_rollers
• football_boots
• golf_apparel
• golf_rangefinders
• golf_shoes
• headlamps
• hiking_boots
• indoor_soccer_shoes
• light_hiking_shoes
• max_cushion_running_shoes
• merino_socks
• pickleball_paddles
• protein_powders
• racing_flats
• rain_jackets
• resistance_bands
• road_running_shoes
• rowers
• running_shoes
• running_socks
• smart_jump_ropes
• soccer_cleats
• spin_bikes
• sports_bras
• sports_water_bottles
• sunglasses_for_running
• swim_goggles
• tennis_apparel
• tennis_rackets
• tennis_shoes
• trail_backpacks
• trail_hiking_boots
• trail_running_shoes
• treadmills
• triathlon_wetsuits
• winter_jackets
• yoga_blocks
• yoga_mats
• yoga_towels

These CSVs contain Plackett–Luce (PL) and frequency statistics for each category.

• pl_rank: Rank by PL mean score (lower is better)
• pl_score: Estimated worth (normalized)
• pl_ci_low / pl_ci_high: 95 % bootstrap confidence interval
• freq_top1 / freq_top3: Share of times the brand appeared in top 1 or top 3 positions
• _ci: Approx. bootstrap confidence width for the frequency metric

All scores derived from canonicalized data in /data/results.canonical.jsonl using php bin/aggregate.php

Below we highlight eight representative categories that illustrate distinct agreement/disagreement patterns between PL (our Bradley–Terry/Plackett–Luce worth) and popularity (frequency@1/@3). These cases are drawn from final data results (global). Unless stated otherwise, interpret reshuffles near cutlines as statistical ties.

Rowers. Figure 1 shows a textbook consensus: Concept2 dominates on PL and also leads in popularity; the remainder forms a second tier (e.g., WaterRower, NordicTrack). This is the ideal “quality aligns with popularity” scenario.

Figure 1. Rowers — consensus leader.

PL crowns Concept2 and frequency@1/@3 agree; remaining brands form a distant pack.

Protein powders. Figure 2 captures an extreme case: Optimum Nutrition nearly saturates PL and wins popularity. The rest of the market drops off when a single “default” brand is dominate.

Figure 2. Protein powders — extreme dominance.

A single brand (Optimum Nutrition) dominates PL and frequency, with a very long tail.

Running shoes. Figure 3 shows divergence: a PL leader (e.g., Nike) coexists with shortlist momentum around Brooks/ASICS in frequency@3.

Figure 3. Running shoes — PL vs popularity divergence.

Quality (PL) favors one brand while users shortlist others.

Triathlon wetsuits. Figure 4 exhibits a compact top cluster (Orca, Quintana Roo, BlueSeventy). With overlapping uncertainty, these are practical ties; policy should avoid over-interpreting small gaps.

Figure 4. Triathlon wetsuits — tight race.

A 2–3 brand pack with overlapping PL uncertainty—treat as statistical ties near the cutline.

Resistance bands. Figure 5 shows TheraBand as the PL leader, while popularity drops across rivals (e.g., SPRI, Bodylastics). The shape also reveals potential substitution pitfalls—worth expanding on.

Figure 5. Resistance bands — long tail (and alias risk).

Clear PL leader but fragmented popularity; prioritize alias cleanup and a tight top-N.

Winter jackets. Figure 6 illustrates preference splits likely driven by locale/retailer exposure: PL favors UNIQLO, while frequency@1 spikes on The North Face. This is a good reminder that popularity is not the same as factors such as quality.

Figure 6. Winter jackets — regional/retailer effects.

PL and popularity point to different leaders; likely influenced by regional/retailer footprint.

Pickleball paddles. Figure 7 presents a clean top-3 (e.g., Selkirk, Paddletek, ONIX) that’s consistent across PL and frequency@3. This is where a fixed cutline with stable recommendations works well.

Figure 7. Pickleball paddles — clear podium.

A stable top-3 across PL and frequency; well-suited to a fixed recommendations set.

Smart jump ropes. Figure 8 shows no dominant PL leader and scattered frequency spikes. In such categories, widen coverage and favor specific situations over hard defaulting.

Figure 8. Smart jump ropes — fragmented field.

Low, diffuse PL with sporadic popularity spikes—prefer broader coverage and exploration.

bin/
  probe.php           # run k samples/entity and write JSONL
  aggregate.php       # compute Bradley–Terry scores + Spearman and export CSV
src/
  Client/LLMClient.php
  Prompt/Template.php
  Schema/Validator.php
  Canon/BrandMap.php
  Rank/BradleyTerry.php
  Stats/Spearman.php
  Stats/Bootstrap.php
data/
  entities.sample.jsonl
  results.sample.jsonl
  pi-top/...
tests/
  unit/...
.github/workflows/
  php.yml             # CI with PHPUnit

composer install
php bin/probe.php --entities data/entities.sample.jsonl --out data/results.sample.jsonl --k 100 --n 10 --temperature 0.5
php bin/aggregate.php --in data/results.sample.jsonl --out data/scores.csv

• Replace LLMClient::call() with your vendor API using structured outputs (JSON Schema in Schema/Validator.php).
• Bradley–Terry implementation uses MM updates; Spearman ρ via MathPHP fallback when installed.
• Add your alias rules to Canon/BrandMap.php.

Bradley–Terry (default):

php bin/aggregate.php --in data/results.sample.jsonl --out data/scores_bt.csv --method bt

Plackett–Luce (listwise):

php bin/aggregate.php --in data/results.sample.jsonl --out data/scores_pl.csv --method pl

• Code: Apache-2.0 (c) 2025 SEO VENDOR LLC. https://seovendor.co
• Data: CC BY 4.0

Patent & License Notice
Portions of this repository implement methods described in U.S. Patent Application No. 18/922,789, “LARGE LANGUAGE MODEL RANKING GENERATION AND REPORTING SYSTEM.”
Use of this code does not grant rights under that patent or any continuations.
See LICENSE, NOTICE, and PATENT.md for full details.