A reproducible LLM-as-a-judge length-bias experiment · 11 graders · 15,840 grades
I padded 120 public, human-scored essays with sentences that add no content, then had each grader score both versions. Small graders raise the grade for padding. Claude, GPT, Gemini, DeepSeek and others lower it by 0.7–1.4 points, which is 2–9× what they take off for deleting 40% of the essay's argument. Telling the grader to ignore length barely helps.
|
Multi-tenant AI commerce platform Conversational AI for orders, customer service, and business automation across WhatsApp, TikTok, Facebook, and web — one backend, many brands. |
Bulk messaging platform SMS infrastructure for operators at scale — campaigns, provider failover, queueing, delivery analytics. Dockerized monorepo. |
|
Specialist software studio — site & brand system SvelteKit 2 + Svelte 5 runes. Runtime theme engine swaps accent palette and logo variants without rebuilds. No UI framework. |
Credential & inventory application Web app for managing and sharing game accounts — credentials, in-game state, skins, battle passes, ban status — with audit trail. |
fsh-lib bridge — one codebase, any server stack.
|
Core bridge — framework, inventory, UI, dispatch, phone, MDT, garage, keys. Auto-detects QB / ESX / Standalone and 9+ inventory systems. |
Skill-check crafting with XP/leveling, placeable benches, nonce-based anti-replay, oxmysql persistence. |
Per-weapon damage/recoil/spread tuning with React admin UI, presets, and audit log. |
|
Placeable drug-breakdown scale with risk & reputation. Zero-dependency standalone. |
Money-laundering machines with wear, parts durability, placeable props. |
Phone-booth supplier encounters with trust/heat. 4-backend audio abstraction. |
|
Street drug sales with employee system and vehicle buyer spawns. |
Hidden vehicle stashes with key-script integration. Persistent oxmysql storage per-vehicle. |
QBCore resource forks, Discord stat bots, NPWD (TypeScript), GTA V audio occlusion tools, FiveM Loki logging setups. |
A typical multi-tenant AI backend — isolation at every layer, LLM calls policy-gated (OpenAI · Anthropic · Gemini), channel fanout through a durable queue.
|
|
|
|
|
|