September delivered the rules engine, the board generator, and four families of AI opponent. October made it real. Seventeen commits between October 3 and October 27 added 16,709 lines across 146 files. The largest single change: a full 3D tactical window that renders the hex board in perspective, with terrain elevation, unit counters, and line-of-sight visualization layered over the procedurally generated maps.
The AI opponents advanced three generations in a single month. Agent v4 introduced shaped reward: instead of a sparse win/lose signal, the training loop now credits territory control, unit preservation, and tactical position at every turn. Agent v5 added imitation learning, where a neural network trains on game traces from strong play before switching to self-play reinforcement learning. Agent v6 introduced the gauntlet: a rotating evaluation bracket where each new agent generation must beat the previous three to graduate.
A browser client built on TypeScript and Vite now runs alongside the pygame-ce desktop client. Browser tests verify the client against the same rules core. Sound design arrived: weapon reports, movement cues, and status tones keyed to game events. Game-play videos can be recorded and exported for post-game review or demonstration.
Local inference cluster support means the prompted and tuned model agents can now distribute work across multiple machines. A packaging tool bundles the project for distribution.
“Each new agent generation must beat the previous three to graduate.”
— Gauntlet evaluation protocol
Spoor_AI is a new project that asks whether attack patterns in honeypot logs can be found by evolution rather than by hand-written rules. The platform reads a honeypi archive spanning 309 days and more than 68 million events, then trains, measures, and evolves detectors using a genetic algorithm. Slow-scan detection is the first attack class under study.
The system works in three stages. The visualizer renders event streams into images. The pattern detector evolves “picture genomes” that compete on recall and precision against labelled ground truth. The characterization stage (not yet built) will classify what the detector found. Early results are honest: the evolved picture genome and the hand-tuned baseline achieve the same recall, and a shuffled control also matches—meaning the picture layout adds no measured information yet. The project ships 76 Python source files and 6 test files under GPL-3.0, with 12,929 lines of source code.
Volition_AI is a controlled experiment in making language models exhibit volitional behavior: commit-then-execute discipline, goal shielding against distractors, pre-set if-then implementation rules, knowing when to stop, and honest objection when a task conflicts with stated values. The study uses three arms—prompt engineering alone, external scaffolding, and LoRA fine-tuning—and measures each against the same battery of tasks in simulated in-memory worlds.
A dual-band conversational design separates the visible dialogue from a hidden side channel where the model records its internal commitments. The project ships 224 Python source files, 30 test files, and 9,141 lines of source code.
Bentley_AI asks a deceptively simple question: if you compress every weight matrix in a transformer into a codebook of shared vectors, how small can the codebook get before the model breaks? The project tests SmolLM2-135M, Qwen3-0.6B-Base, and Llama-2-7B using k-means clustering, product quantization, minimum description length with three compression tiers, and E8P lattice quantization, with and without Hadamard rotation. KL divergence fine-tuning recovers quality after compression.
The compute runs on an RTX 5090 workstation plus a local cluster of four AMD Ryzen AI Max+ 395 machines, each with 128 GB of unified memory. The project ships 66 Python source files, 18 test files, and 8,448 lines of source code. A pre-registration document and research survey are dated October 28.
RAD_PhotoFrame_AI takes a PF1007L Android photo frame (Allwinner A33, 1 GiB RAM, 1280x800) and repurposes it as a wall-mounted status display. The frame gets root through USB, the vendor software is disabled, and a kiosk app polls a Python service on the workstation every ten seconds for HTML dashboard fragments. Planned panels include AI work in progress, VIBE project status, host health, and a family clock with calendar and weather. The frame holds no credentials and no source documents. Status: design documents only, no code yet.
SensorGrind_AI, the self-evolving honeypot-log analysis pipeline (now spanning 1.1 billion events across 662 active days with 36 or more analyzer stages), received a commit adding neuromorphic processing capabilities. The spike-timing approach complements the existing statistical and ML stages in the pipeline.
Transcript2Symbol_AI received three October commits advancing from architecture to initial build-out. A home-host amendment accepted the RTX 5090 workstation as the development target. The symbolizer pipeline now runs on dev talk transcripts: 24 files, 2,610 turns, 2,710 frames produced in 1,005 seconds of wall time.
··· “Frustrating adversaries since the dial-up era” · GitHub: rondilley · 62 Repositories and Counting ···
September asked how to represent signals so machines can reason and humans can audit. October asks four sharper questions: Can attack patterns be found by evolution rather than by rule? (Spoor.) Can fine-tuning produce volitional behavior that prompting cannot? (Volition.) How many distinct weight vectors does a transformer actually need? (Bentley.) Can a $50 photo frame become a useful operations display? (RAD PhotoFrame.) Each project carries its own success criteria, its own evaluation method, and its own conditions for stopping early. Spoor has already reported its first honest null: the evolved picture layout adds no measured information yet. That is a result, not a failure.
October’s hardware story is the emergence of a local inference cluster. The RTX 5090 workstation anchors Squad Leader’s training runs and Bentley’s compression experiments. Four AMD Ryzen AI Max+ 395 machines with 128 GB of unified memory each provide distributed inference for the prompted and tuned model agents. Transcript2Symbol accepted the RTX 5090 as its home host via a formal amendment. The cluster runs local models for Squad Leader, Bentley, and Transcript2Symbol without sending game states, weight matrices, or transcripts to any cloud endpoint.
October 3: Squad Leader 3D window and agent v4. October 3–7: Transcript2Symbol home-host amendment and initial build-out. October 14: Volition AI and RAD PhotoFrame initial commits. October 22: Spoor AI initial commit. October 23: SensorGrind neuromorphic capabilities. October 27: Squad Leader browser client, sound, agent v5–v6, packaging tool. October 28: Bentley AI research survey and pre-registration. October 30: Bentley AI initial code import.
Agent v4’s shaped reward replaces the sparse win/lose signal with a per-turn credit that accounts for territory control, unit preservation, and tactical positioning. The dense reward allows the PPO training loop to learn from every game turn rather than waiting for the final outcome. The challenge is reward hacking: an agent that optimizes for the proxy score while losing games. The gauntlet in v6 addresses this by requiring each generation to win actual matches against its predecessors.
Spoor_AI’s detector evolution works by treating each detector configuration as a genome. A population of genomes competes on recall and precision against labelled honeypot events. Crossover and mutation produce the next generation. The honest early finding—that a shuffled control matches the evolved detector’s recall—means the current fitness landscape is flat. The next step is changing how events map to the picture representation, not tuning the evolution parameters.
Bentley_AI tests five compression approaches: k-means clustering (the baseline), product quantization (split each vector into sub-vectors and quantize independently), minimum description length at three granularity tiers, E8P lattice quantization (a mathematically optimal packing), and all of the above with and without Hadamard rotation to spread information across dimensions before quantizing. KL divergence fine-tuning after compression recovers some of the lost model quality.