The Cloud Subscription Trap
How $20–$200 AI tiers use honeymoon quotas, classifier routers, and sliding-window pruning to protect margins while degrading output.
Read Report →The Engine vs. The Wrapper
Why flat-rate cloud AI subscriptions rely on silent downgrades, context truncation, and data harvesting—and why developers are returning to dedicated local GPUs.
Technical Reports
The Capex Squeeze
Why trillion-dollar data center spending only pays off if your devices become terminals. The toll booth business model, the memory paradox, and data gravity as lock-in.
Read Report →Why Context Windows Lie
1M-token promises vs. 30K reality. How web wrappers silently compress, summarize, and drop middle-prompt constraints to break long-term consistency.
Read Report →Project Showcase
Multi-File Code Refactor Agent
Processes 12k+ lines across 8 modules. Maintains cross-file constraints, preserves existing logic, and outputs validated diffs. Runs fully offline.
100% Verbatim Context
Zero Telemetry
Structured Log Analyzer
Ingests raw server logs, extracts patterns, generates compliance summaries, and flags anomalies. No external routing, no prompt leakage.
Deterministic Output
Local KV Cache
Technical Documentation Synthesizer
Aggregates RFCs, API specs, and internal docs. Produces version-controlled markdown with cross-references and constraint tracking.
Full Prompt Retention
No Vendor Lock-in
Core Principles
Consistency Over Convenience
Cloud wrappers promise uniformity but deliver throttled routing, context pruning, and silent model swaps. Local inference guarantees the same weights, same context window, and same constraints every turn. You own the instrument. You control the workflow. No telemetry. No hidden downgrades. Just deterministic, repeatable AI that works the same today as it did yesterday.
About
Methodology & Scope
The patterns described in our reports are observable across multiple commercial tiers. The specific routing percentages and degradation metrics reflect aggregated behavior observed in independent testing, not a single vendor's architecture.
Half Curtain is an independent, self-hosted research space. No analytics, no tracking, no cloud dependencies. All content is written and maintained by a single operator running fully local AI infrastructure.
Contact
Due to the volume of automated scraping and unsolicited mail, the contact channel is temporarily closed. Reports and findings will continue to be published here as they are completed.