architecture field log // 2026—ongoing

Generational tracking: how the K4 machine evolves.

This is not one system endlessly guessing. It is a research project that keeps being rebuilt as models, local hardware, and our evidence improve. Each generation preserves what was reproducible, retires what was wasteful, and turns failure into constraints for the next architecture.

07architecture generations
01unchanged solve standard
∞current local horizon

The architecture changes, but the acceptance rule does not: reproduce the ciphertext exactly, satisfy the held-out Kryptos clues at their exact positions, disclose the method and key material, and leave an auditable trail. A persuasive story or high language score is never a solve.

Scroll the line · select any node to jump

01

deterministic foundation

Generation 01: The reproducible cipher lab

absorbed
cipher plugins → candidate generator → exact verifier → SQLite ledger
What we tried

Finite implementations of classical cipher, route, key-schedule, and transform families running directly on the Mac.

What we learned

Search volume is not evidence. Every candidate needs a parameter signature, reversible implementation, and exact round-trip check.

What survived

The verifier, experiment ledger, deduplication, private solve inbox, and reproduce-or-reject policy remain the foundation today.

02

adaptive execution

Generation 02: Test small, then earn more compute

evolved
experiment catalog → finite micro-test → promote / stop / review
What we tried

A broad plugin catalog with planner scoring, small probes, stopping rules, failure mining, and promotion gates.

What we learned

Cheap falsification beats blindly extending every family. Exhausted parameter signatures must become durable negative evidence.

What survived

Bounded experiments, resumable checkpoints, route deduplication, and the rule that new compute must buy genuinely new coverage.

03

seven-day frontier chain

Generation 03: Use frontier reasoning to choose the search

retired
evidence → frontier proposer → frontier critic → planner → verifier
What we tried

A chained frontier model campaign that proposed three bounded mechanisms, criticized them independently, and compiled one into code-owned search.

What we learned

Strong models help most with hypothesis selection and adversarial review—not candidate enumeration. Serial public calls also made quota an availability dependency.

What survived

The proposer/critic separation, strict structured outputs, blind held-out clues, and the declarative boundary between model judgment and execution.

04

evidence discipline

Generation 04: Make the machine argue against itself

absorbed
negative evidence + null controls + coverage ledger → blinded model context
What we tried

Failure-aware prompts, known-cipher calibration, false-positive controls, multiple-testing correction, and explicit research-productivity telemetry.

What we learned

A diagnostic maximum such as 0.49 is not progress or probability. With millions of trials, interesting-looking noise is expected.

What survived

Campaign-adjusted evidence, held-out validation, novelty accounting, and public language that distinguishes activity from signal.

05

hybrid routing

Generation 05: Frontier judgment, local mechanics

retired
frontier proposal + critique → local Qwen planning → deterministic sweep
What we tried

Reserved paid frontier calls for research judgment while moving bounded planning and basic language work to local Qwen.

What we learned

Model routing should reflect the value of the decision. Quota exhaustion, circuit state, and model health need to be visible—not silent failure modes.

What survived

Role-specific models, stage-by-stage health, circuit-breaker telemetry, and deterministic fallback boundaries.

06

continuous local research

Generation 06: Big reasoning where it matters. Local everywhere.

retired
local 120B proposer → local 120B critic → local Qwen planner → verifier
What changed

gpt-oss:120b now handles the two judgment-heavy roles. qwen2.5:14b compiles bounded plans. Public LLM routing is disabled.

What we measured

Medium reasoning completed reliably where high effort exhausted its answer budget. Canonical H1/H2/H3 IDs removed a meaningless model-copying failure.

Why it matters

The loop can run without API credits. Models unload between roles to fit the Mac, while deterministic code remains the only path to acceptance.

07

clean-sheet rebuild

Generation 07: Build a discovery laboratory, not a longer conversation.

current
evidence vault + sealed gym → independent local agent islands → typed programs → constraint workers → proof gate
What changed

Agents now need research certificates, finite budgets, controls, falsifiers, and typed executable programs before compute is authorized.

What we measure

Exact benchmark recovery, information gained, semantic families opened or closed, duplicate rate, and independent replication—not raw candidate volume.

Why it matters

Large models work in residency-aware batches while deterministic workers run independently. The $1 plaintext check remains human-approved and rare.

Open K4 Forge

the next architecture

It has to earn its place.

We will keep changing the stack as foundation models and local inference improve—but only when a new design increases reliable novelty, reduces repeated work, or strengthens falsification. More agents, more tokens, and more candidates are not automatically more progress.