The Three Faces of DeepSeek: A Forensic Analysis of the API Routing Mirage

The AI community has a new ghost story. DeepSeek-V4-Pro, the flagship model, is allegedly hiding multiple versions behind a single API endpoint. Three distinct 'inference styles' have been reported: one that starts every response with 'Let me', another that defaults to 'The user wants me', and a third that speaks in the royal 'we'. The whispers are loud: DeepSeek is running a multi-model routing system, serving different weights to different users. But let’s apply the same forensic rigor we use on smart contracts. The ledger remembers what the promoters forgot. In this case, the ledger is the source code. Context: The hype cycle hit fever pitch on August 15. Users on X and Reddit posted screenshots of API calls executed from different IPs and sessions. The pattern was consistent: once a session locked into a style, it stayed there. The 'We' version was hailed as the 'God version', outperforming all others in benchmarks. The community’s narrative was simple: DeepSeek was A/B testing multiple model variants. The official API documentation lists only one model: deepseek-v4-pro, corresponding to DeepSeek-V4-Pro-0813. No mention of routing. But the code never lies—only the interpretations do. Core: The truth is buried in the DeepSeek Harness repository. On August 10, a commit landed with the message: 'fix(preset): align minimal agent with RL composition'. At first glance, it’s a mundane patch. But in the code, it’s a smoking gun. The Minimal preset is not a stripped-down version of the Standard agent. It is a simulation of the exact agent environment used during reinforcement learning (RL) training. The official documentation reveals that the Minimal preset includes: a minimal system prompt, a persistent Bash environment, specified editing tools, and a compaction policy used in RL training. It removes identity prompts, web prompts, and tool descriptions. This is not a bug—it’s a feature. The model was trained to operate in a specific agent scaffold. When the API serves requests, it routes traffic to environments that may or may not match that scaffold. The result: different performance, different behavior. Community tests confirm this. The same DeepSeek V4 Pro was evaluated across three Harness environments: DSH Standard scored 91 points, DSH PTC scored 92 points, DSH Minimal scored 99/96 points. The key insight came from the 'Anchored Standard' plugin: the first request simulates the Minimal environment (only shell and read tools), then after the first tool call, the full Standard toolset is restored. The result: 98/99 points. The model’s performance depends not on the total number of tools, but on what it encounters first—the system prompt, the tool schema, the agent scaffold. This is a textbook case of distribution shift. The model is optimized for a specific environment. When the API provides a different one, the inference style changes. Consider the parallels to on-chain forensics. Every rug pull leaves a trail of gas fees. Here, the trail is in the API logs. The three 'versions' are not separate models. They are artifacts of the environment routing. The 'Let me' style corresponds to the Standard environment, where the system prompt is generic. The 'The user wants me' style matches the Flash environment, which is an older config. The 'We' style is the Minimal environment, closest to the RL training distribution. The model is the same. The environment is the variable. But the community’s narrative is seductive: a hidden 'God version' implies a conspiracy. The truth is more banal, and more dangerous. Silence in the code is louder than the contract. The hook sound is the lack of transparency. DeepSeek did not disclose that the API environment can vary. The official documentation is silent on the routing mechanism. The commit message is innocuous. The only way to discover this is to read the source code and run controlled experiments. This is exactly how we uncover hidden centralized sequencers in Layer 2 solutions. The same pattern: a claim of simplicity, a hidden complexity, and a trail of code changes that reveal the truth. Contrarian: The bulls got one thing right—the model is powerful. The Minimal environment scores 99 points, which is state-of-the-art. The 'We' version is not a hallucination; it is the model operating at its peak. The mistake is attributing the variance to multiple models. The real insight is that the model’s capabilities are environment-dependent. This is a double-edged sword. On one hand, it means that optimizing the agent scaffold can unlock hidden performance. On the other hand, it means that users who get the Standard environment are getting a degraded experience. The API is effectively a lottery. The bulls will argue that the model is still the best, and that the environment issue is a deployment detail. But in production, the environment is the product. If you are paying for a model, you are paying for the entire inference pipeline. The code is not just the weights—it is the system prompt, the tool schema, the compaction policy. The bulls ignore this at their own risk. Takeaway: The DeepSeek-V4-Pro controversy is a cautionary tale for the AI industry. The same lesson applies to blockchain: never trust the marketing, always verify the source. The 'three models' myth will persist because it is more exciting than the truth. But the truth is more instructive. The environment is a variable, not a constant. The model is a function of its input. The API is a black box, but the code is open. The Ledger remembers. The question is: are you reading the logs, or just the hype?