Ben Michalek / BoggersTheFish
A living archive of systems that make reasoning inspectable.
Verifier-first reasoning. Deterministic language. Tension-driven model experiments. Verified computing.
The code may change or disappear. The questions, evidence, failures, and useful mechanisms remain recorded here.
current architecture boundary
human language
→ semantic graph
→ typed verification
→ accepted state
→ response plan
model proposals never become proof by confidence aloneTS-Reasoner
Typed verifier authority, containment, repair, memory, risk gates, replay, and receipts.
TS Chat Language
Human language to inspectable semantic state without using an external LLM as the language substrate.
Ten-SON-LM
Tests whether tension is a useful causal control signal inside a recurrent semantic workspace.
TensionForge
Custom RX 480 OpenCL training with forward, backward, optimizer, and recurrent parity evidence.
Quick Links
Boggers Credits (BGC)
Support the work without financial promises.
Donations are donations. Boggers Credits (BGC) are experimental off-chain community credits: not cryptocurrency, not an investment, not legal tender, not equity, not redeemable for cash, may remain off-chain forever, and not financial advice.
not crypto / not investment
Learn more about BGCCurrent architecture
Four programmes, one evidence rule
Each programme has a different job. None may convert fluency, confidence, naming, or version count into proof.
Reasoning and language
ts-chat-language compiles human input into explicit semantic state. TS-Reasoner decides what may be accepted, repaired, remembered, or rejected.
Open programmeModel science
Ten-SON-LM asks whether learned tension causally improves a recurrent semantic workspace. TensionLM remains the earlier attention-based evidence line.
Open programmeAdaptive inference
TSQ explores precision escalation driven by tension and verifier failure. It does not claim an efficiency win until distinct numerical paths are measured.
Open programmeVerified computing
bogbin develops deterministic state, capabilities, persistence, and rollback. TensionForge tests verified training on legacy commodity hardware.
Open programmeCurrent evidence
Field notes, including the failures
The archive records what happened, not just what sounds impressive.
Ten-SON Milestone 1: Partial
Delayed recall, balanced brackets, and a synthetic next-token task learned. Copy missed threshold. Tension evidence is promising on one task and mixed elsewhere.
TensionForge: correct, currently slow
The RX 480 path has forward, backward, optimizer, and recurrent parity receipts, but the tested runtime is substantially slower than PyTorch CPU.
TSQ: runtime before quantization proof
Routing, repair, evaluation, data, and adapter tooling exist. Genuine adaptive-precision evidence waits for materially different low/high numerical backends.
TS Habitat v3: verified symbolic agent loop
The bounded simulated habitat now maintains explicit goals, uses deterministic graph tension to allocate capped effort, navigates declared topology, verifies and executes one symbolic action at a time, checks observed effects, replans stale assumptions, and emits whole-run TurnReceipt v3 evidence. It uses no external model or real-world actuation.
Permanent lineage
52 repositories become one readable history.
Old repositories can be deleted without pretending they never existed. The lineage archive records each meaningful idea, what survived, what failed, and which current project inherited it.
Explore the project lineageClaim boundary
Runnable mechanisms are real. Broad conclusions still need proof.
Not a finished AGI.
Not a general theorem prover.
Not a production operating system.
Not an adaptive-quantization win yet.
Verifier authority is separate from model confidence.
Negative results and regressions stay visible.
Choose your path
A cleaner way into the archive
Different visitors need different first moves. This section keeps the serious routes obvious.
I'm new
Read the plain-language map before opening source repositories.
Open routeI want current work
Inspect the active programmes, evidence boundaries, and next experiments.
Open routeI want the full history
See how 52 repositories became the current modular stack.
Open routeI want evidence
Start with claims, setups, results, limitations, and replay notes.
Open route