Benchmarks/Next-move identification

Benchmarking Mimic One on Real-World Actors

Mimic One, the latest model in Hidden Systems’ Mimic series, follows Mimic Zero, released earlier in 2026. It reads the situation around a person, government, company or institution and forecasts the moves it is most likely to make next. On a sealed test of real events it identifies the next move more often than Claude Sonnet, GPT, DeepSeek and Gemini, despite a much smaller model size.

Next-move identification · judged · %
Mimic One: 56.0
Claude Sonnet 5.5: 53.7
GPT-6.1 Sol: 50.0
DeepSeek V4 Flash: 49.3
Gemini 3.8 Flash: 44.8
01

What Mimic One Does

Given a moment in a developing story and an actor in it, Mimic One returns a ranked list of likely next moves, each in a single concrete line, with a probability attached. Within a simulation, those predicted moves become possible branches: each one changes the situation the other actors face, and the next round of moves follows from there.

The output is a short list of materially different moves, not a narrative. Restating the same move in different words does not help a simulation explore, so the list is built to cover distinct possibilities.

02

The Test

The sealed test holds 609 forecast points drawn from 68 real news stories that were never used for training or tuning. Each point fixes a moment, an actor and the public record available at that moment. Everything that happened afterwards is withheld from every system and used only for scoring.

Every system receives the same context and the same instruction, and returns exactly 12 moves. Fixing the list length means no system can improve its score by listing more.

The headline measure asks one question per forecast point: does the list contain the move the actor actually made next? We use the 402 next moves that are clearly documented in the record, so the target itself is not in doubt.

A Heathrow expansion case from a separate evaluation illustrates the task. Asked what Andy Burnham would do next, the 4B model included “Andy Burnham opposes the Heathrow third runway” among its 12 predictions. The recorded move was his refusal to back the third runway, so the list scored a match.

Next-move identification · judged · %
Mimic One: 56.0
Claude Sonnet 5.5: 53.7
GPT-6.1 Sol: 50.0
DeepSeek V4 Flash: 49.3
Gemini 3.8 Flash: 44.8
Fig. 1. Share of next moves contained in each system’s 12-move list.

Mimic One identifies 56.0% of next moves, ahead of Claude Sonnet 5.5 at 53.7%, GPT-6.1 Sol at 50.0%, DeepSeek V4 Flash at 49.3% and Gemini 3.8 Flash at 44.8%. Its lead over Gemini is statistically clear; the leads over Sonnet, GPT and DeepSeek are within the test’s margin, so the fair reading is that a 4B model matches or beats frontier systems on this task.

03

Built for Simulation

Mimic One runs on a single 4B-parameter base model. The systems it matches or beats are estimated at about 70 to 375 times larger.

Model size · billions of parameters
Mimic One4B
Claude Sonnet 5.5outside estimate~1000B
GPT-6.1 Soloutside estimate~1500B
DeepSeek V4 Flashpublished284B
Gemini 3.8 Flashoutside estimate~275B

Total parameters; each cell is 15B, and Mimic One’s 4B is drawn as one cell. Anthropic, OpenAI and Google do not publish model sizes; their figures are rough outside estimates. GPT-6.1 Sol’s ~1.5T follows a corrected knowledge-probe estimate of GPT-5.5 (~1.5T), and Claude Sonnet’s ~1T a knowledge-probe estimate of Claude Sonnet 5 (~0.97T).

Fig. 2. Model size in billions of parameters. DeepSeek V4 Flash has 284B total parameters and uses 13B per token [1].

Size matters because of how simulation works. Every move an actor makes opens new branches, and every branch needs fresh decisions from every actor in it. A single scenario can require thousands of these calls. A model that is compact and fast makes deep, wide exploration affordable; a frontier model at every step would not.

04

The Scenarios

The 68 stories are sampled from live news across many domains: monetary policy and central banking, financial markets and crypto, technology companies and AI labs, aviation and corporate disputes, courts and prosecutions, elections and party politics, sport, diplomacy and armed conflict. Each story contributes forecast points for the actors taking part in it.

That gives 194 distinct actors. They include central banks such as the Bank of England and members of the Federal Reserve; regulators including the CFTC, SEC commissioners and the FAA; companies from NVIDIA, Meta and Saudi Aramco to Ryanair, Coinbase and Flutter Entertainment; football clubs and their players; prosecutors, police forces and political parties; governments and ministers; and individuals from technology founders to broadcasters and musicians.

The moves are just as varied: rate decisions, product launches, lawsuits, transfers, resignations, public statements, sanctions and strikes. Many forecast points concern a named individual rather than an institution.

05

Behaviour — A Context Problem

Predicting what an actor does next is not mainly a test of general intelligence. It depends on knowing how this actor behaves: what it tends to say, which moves it repeats, how it responds to pressure, and what the situation now allows. That knowledge comes from the context and from training on how actors actually behave.

It matters most in complex situations, where many actors are interacting, several threads of the story run at once, and far more moves are open to each actor. We split the test into thirds three ways: by how much is happening in the story, by how many actors are actively involved, and by how much of the context concerns the actor being forecast.

Next-move identification by scenario complexity · %
Mimic OneClaude Sonnet 5.5GPT-6.1 SolDeepSeek V4 FlashGemini 3.8 Flash
304050607059.8 Mimic One48.5 Claude Sonnet 5.543.9 GPT-6.1 Sol44.7 DeepSeek V4 Flash31.8 Gemini 3.8 FlashUp to 129 events130–459 events460+ events6063645760Up to 129 events4849424642130–459 events6049444532460+ events
Fig. 3. Next-move identification by the number of events in the story’s timeline at the forecast point. Each third holds about 134 judged next moves.

On simple stories every system scores around 60%. On the most complex third the frontier models fall to between 32% and 49%, while Mimic One holds at 59.8%: a lead of 17.6 points over their average (95% interval 10.4 to 27.1).

Next-move identification by active actors · %
Mimic OneClaude Sonnet 5.5GPT-6.1 SolDeepSeek V4 FlashGemini 3.8 Flash
304050607057.1 Mimic One48.4 Claude Sonnet 5.543.7 GPT-6.1 Sol49.2 DeepSeek V4 Flash34.1 Gemini 3.8 Flash2–7 actors8–15 actors16–35 actors58595251502–7 actors53535447498–15 actors574844493416–35 actors
Fig. 4. Next-move identification by the number of actors actively involved in the situation. Each third holds about 134 judged next moves.

The same pattern appears as more actors become involved. Mimic One remains between 53% and 58% across all three groups, while the other systems fall as low as 34%.

Next-move identification by context about the actor · %
Mimic OneClaude Sonnet 5.5GPT-6.1 SolDeepSeek V4 FlashGemini 3.8 Flash
304050607050.0 Mimic One43.3 Claude Sonnet 5.534.3 GPT-6.1 Sol44.8 DeepSeek V4 Flash26.1 Gemini 3.8 FlashUp to 16 lines17–62 lines64+ lines5965595956Up to 16 lines595356445217–62 lines504334452664+ lines
Fig. 5. Next-move identification by how many lines of the context concern the actor being forecast. Each third holds about 134 judged next moves.

Context about the actor itself matters too. Every system finds the actors with the longest records harder to call, because they are the busiest actors in the busiest stories. Mimic One loses least: with 64 or more lines about the actor it leads Claude Sonnet by 6.7 points, GPT by 15.7 and Gemini by 23.9.

The two effects overlap, since busy actors sit in busy stories, but each holds when the other is fixed. In the most complex stories Mimic One leads the other systems by 17 to 19 points whether the actor’s own record is thin or dense, and within complex stories its lead grows as the record about the actor grows.

06

Scoring

Each system’s list is judged against the actor’s actual next move by two independent passes of Claude Sonnet, with Claude Opus settling any disagreement. The judge sees only the list and the true move, and the same procedure and items are used for every system. Language-model judges have been studied for evaluating open-ended outputs [2].

Intervals are computed by resampling whole stories, keeping forecast points from the same story together rather than treating them as independent observations [3]. Mimic One leads Claude Sonnet by 2.1 points (95% interval −5.4 to +9.4), GPT-6.1 Sol by 5.8 (−1.9 to +12.7), DeepSeek by 6.7 (0.0 to +13.3) and Gemini by 11.0 (+2.2 to +19.2).

07

Reading the Results

The test is drawn from real stories, so it measures the task Meridian needs: forecasting what specific actors do next from the public record at the time. It is not a general benchmark of intelligence.

This evaluation uses a 4B model trained on fewer than 1,000 scenarios, with 60 reinforcement-learning steps. Absolute recall is constrained by scenario complexity and the available context, so the central comparison is how well the models forecast given the same information. These results do not establish a performance ceiling. See Further Research for training details and next steps.

Across all evaluated models, the combined set of correct predictions covers more than 80% of scored next moves, even with the dataset’s imperfect context. This counts a move as covered when at least one model predicts it correctly; it is the union of their hits, not the score of a deployed ensemble. The models recover complementary moves, leaving substantial room to improve single-model recall.

Ground truth comes from a dated news record. Moves the record misses count against every system equally, and retrospective tests cannot remove knowledge already held in a model’s parameters. Mimic One records the highest recall in this comparison. Its lead over Gemini is statistically clear; the remaining comparisons do not establish a clear difference.

08

Further Research

These results come from a 4B-parameter model trained on fewer than 1,000 distinct scenarios that took place over a period shorter than three months. The reported training run used 7.9 H200 GPU-hours at a cost of $30. Reinforcement learning was limited to 60 steps by the number of available training states.

Further research will expand the breadth and time span of training scenarios, increase the number of training states and reinforcement-learning steps, and improve the context supplied at each forecast point. Larger models are another direction, including mixture-of-experts models with 35B total parameters and 3B active parameters per token.

09

References

  1. DeepSeek-AI (2026). DeepSeek-V4-Flash model card. Model architecture and total and activated parameter counts. Accessed 26 October 2026.
  2. Zheng, L., Chiang, W.-L., Sheng, Y., Zhuang, S., Wu, Z., Zhuang, Y., Lin, Z., Li, Z., Li, D., Xing, E. P., Zhang, H., Gonzalez, J. E., and Stoica, I. (2023). Judging LLM-as-a-Judge with MT-Bench and Chatbot Arena. Advances in Neural Information Processing Systems, 36, 46595–46623.
  3. Field, C. A., and Welsh, A. H. (2007). Bootstrapping clustered data. Journal of the Royal Statistical Society: Series B (Statistical Methodology), 69(3), 369–390.