We write history — and we are part of it.

January 2026 marked a bifurcated start to the year. On the core platform front, PrimeForge initiated a strategic shift towards unbundling into a "Micro-App Constellation," moving from a monolithic product to an "Infrastructure Factory." Simultaneously, the company launched **SaidMe**, a completely independent, privacy-focused consumer application. This dual-track approach demonstrates our capacity to scale enterprise infrastructure while delivering secure, standalone consumer utilities.

December focused on internal operational optimization by adopting a "Multi-Agent Loop" development workflow, utilizing tools like Codex, Cursor, and Claude Code. This shift to an "Orchestrator" role has significantly accelerated PrimeForge's iteration velocity, with the "Proof Chain" module now operating successfully at scale.

October marked the completion of the research phase with the formal filing of the UK patent (Application 2517987.0) for the dynamic routing system. With the intellectual property secured, the strategic focus shifted immediately from invention to commercial adoption, rebranding the system as "PrimeForge" and prioritizing sales execution to address revenue lags caused by the intense R&D period.

August focused on the implementation of the ILP Solver, the core engine that converts the Cost Matrix into an executable plan. The system architecture closes the loop between observation and execution, using the solver to minimize costs under constraints like rate limits and concurrency. A key architectural decision was to decouple the metric aggregation from the core routing logic to prevent performance bottlenecks.

June concentrated on solidifying the intellectual property strategy by defining the patent claims around the ILP routing mechanism. Simultaneously, the system underwent its first operational stress test with the launch of the "Right to Work" check app, successfully demonstrating that the new architecture decouples client operations from upstream provider failures. This month was about converting the routing work into something defensible (legally) and undeniable (operationally).

April was the first month of running the ILP planner prototype in production, revealing significant failure modes including high solver latency, frequent infeasibility due to hard constraints, and system oscillation. The root causes were identified as treating real-time planning like an offline batch job and lacking state awareness. Mitigation strategies focused on caching, soft constraints, and hysteresis to ensure stability and graceful degradation.

February focused on solving the integration scalability problem by adopting the Model Context Protocol (MCP) as the standard interface between AI clients and tools. A prototype MCP server for Companies House was built and successfully tested with Claude Desktop, proving MCP's viability as a stable contract. However, the experiment highlighted that MCP is just an interface, and a robust policy layer for authentication, rate limiting, and observability must be built on top.

December's annual review highlighted solid market growth and the stability of the new Go-based infrastructure, but identified the "Dynamic Routing" system as a critical failure. The fallback rule-based approach has become unmaintainable due to combinatorial complexity. The strategic correction for the upcoming year is to abandon heuristic rules in favor of mathematical optimization methods like Operations Research.

October paused feature development to address the "Access Storm" instability by researching Dynamic GraphQL Query Planning. The focus was on designing an architecture that parses queries into an Execution Plan (AST) and optimizes it continuously based on real-time API state, effectively treating the system as a distributed database engine. Feature development was paused to address the "Access Storm" stability issue. Research focused on "Dynamic GraphQL Query Planning."

August focused on developing the "Corporate Health Check" module, employing a Proxy Swarm architecture to aggregate compliance data from various sources including Companies House and the ICO. A Puppeteer-based headless scraper was implemented to handle sources without public APIs. While the system is functional, it remains fragile and maintenance-heavy due to the inherent instability and latency of scraping operations. Development of the "Corporate Health Check" / Due Diligence Report. Objective: Aggregate external data to assess company compliance status.

June focused on optimizing operational efficiency by implementing an Agentic Architecture with an "Intelligent Routing Agent." This system uses a meta-cognitive layer to classify request complexity and dispatch queries to the most appropriate model (GPT-3.5, GPT-4o, or GPT-4), preceded by a semantic cache. This hierarchical approach effectively balances performance and cost, marking a shift towards a true Multi-Agent System.

April was dedicated to experimenting with fine-tuning Llama 3 8b Instruct for job description generation. Despite technical success in setting up a QLoRA training pipeline on Apple Silicon, the project faced severe overfitting issues, with the model memorizing the small dataset rather than generalizing. This failure highlighted the distinction between knowledge injection and style transfer, leading to a strategic pivot away from supervised fine-tuning towards RAG and few-shot prompting. April focused on fine-tuning **Llama 3 8b Instruct** using the curated dataset.

February marked the strategic expansion into the Chinese market with a dual-platform approach utilizing both a WeChat Service Account and a dedicated web platform. A critical technical challenge regarding user identity fragmentation was solved by implementing a `UnionID` based authentication system, enabling a unified user experience. This successful integration has driven early traction, resulting in over 5,000 users and 200 paid subscriptions.

December achieved a major commercial milestone with the onboarding of our first trial user, validating initial market interest. On the technical front, we accelerated and deployed the RAG v1 AI Assistant using GPT-4 and LangChain.js to establish product differentiation. While the system is functional, initial feedback highlights that low data volume limits the utility of the retrieval-augmented generation, identifying a key area for future improvement.

Jian MA has officially arrived in the UK to commence full-time operations. Darkhorseone Limited (Company number: 15002342, incorporated July 13th) has transitioned from a dormant entity to an active operating company. The primary objective for October was to eliminate all administrative impediments to allow for pure product focus in the subsequent months.