For a world that frequently requires quick creative solutions. We build systems that help humans survive the noise — from industrial safety to mental clarity, and, increasingly, inside the machines themselves.
Navigator's Log R&D builds and ships verifiable systems across three domains — the Mind, the Work, and the Network — and, increasingly, inside the machines themselves. Our newest line is machine interpretability and AI safety: frozen, hash-sealed instruments that read how manipulation persists inside language models, released with their code, their pre-registrations, and their failures.
Our guiding principle is that whatever we invest ourselves in as an organization must be for the good of everyone and everything. We understand that's an elegant lens with impossible reach. But it's the star that keeps us on course. It's the accountability of Navigator's Log, and it drives creative solutions that only the imagination and a deep appreciation for the value of life can map.
Mental health interfaces that calm the nervous system.
Industrial protocols that protect the physical body.
Machine interpretability — reading and defending the internal state of AI systems.
One operator, several arms — Navigator's Log R&D, Clear-Point Technical, and Head Shrink Studios. Every system below is built, tested, and maintained from a workstation in Fresno.
The No Harm Principle, applied to the machines. We build a linear detector that reads whether a manipulation planted early in a language model is still live in its internal state after the model stops attending to it — froze it, SHA-256-sealed it before it ever saw a target, and never tuned it. That one frozen detector reads the signal across six independently-built open-weight model families and generalizes to ordinary conversation. The offensive edge the same result implies, we characterize precisely and deliberately do not build.
This is verification in machine form: we only ship what we can test, we commit our method before we run it, and we publish the nulls — including a result we tried, replicated, and honestly retracted.
The companion H-SC line carries that same sealed detector one step further: an early instruction leaves a trace that is not one signal but several, and the safety-relevant directives — refuse and withhold — form their own coherent lane, distinct from the lane that merely shapes an answer. A single linear monitor would miss them.
Read the Research DocumentationThe Revenue Engine & Industrial Compliance. We apply Navy-grade Quality Assurance to the Central Valley's most dangerous machinery. We strip "Linguistic Fog" from safety manuals to prevent injury and save lives.
Visit Clear-PointOne mathematical instrument that detects when a physical system departs from its current state, reading eight topological features of a delay-coordinate phase-space trajectory. Confirmed across eleven independent physical domains spanning twenty-eight orders of magnitude — from gravitational-wave strain to solar wind.
Open the InstrumentFalsification Protocol Audit Workbench. A structured adversarial audit tool that walks a technical claim through eight gates with cryptographic chain of custody and a human signature at every step. Find your blind spots before they find you.
Try the BetaThe artistic production arm. A scripted forensic audio series that turns adversarial audit methodology on institutional systems — education, healthcare, finance, housing, surveillance, organized religion, AI supply chains, ecology. Season One: twenty episodes, scripts complete. Also murals, canvas, ink and gouache, concept sheets, and political cartooning.
Enter the StudioAccessible mental health interfaces designed to interrupt anxiety loops and restore cognitive function. "Soft Technology" that acts as a digital caretaker.
Coming SoonDecision-support systems for navigating life's complexity. The "Triage" system for prioritizing life tasks. Action over Analysis Paralysis.
Coming SoonProject Lotus & The Blooming Wall. Bio-feedback hardware that breathes with you. Ambient architectural robotics that respond to environmental stress.
In DevelopmentReproducible, pre-registered interpretability work on how manipulation persists inside language models — and how to read it. Every step is committed before it is run. The nulls are published. Deposited and citable, not perpetually "in progress."
A single linear detector — frozen and SHA-256-hashed before it ever met a target, and never tuned afterward — reads whether an earlier manipulation is still live in a model's residual stream after the model has stopped attending to it. It transfers across six independently-built open-weight families (Qwen, Phi, SmolLM2, Llama and OLMo among them) with a graded effect of 3.7 to 8.0, every 95% confidence interval clear of zero, and it generalizes to a naturalistic conversational frame. That is direct evidence on a generalization question the published white-box literature leaves open.
The one offensive channel the same math implies is characterized precisely and deliberately left unbuilt. Benign lane throughout — no jailbreaks, no elicitation tooling. Where the work failed, it says so: a firm null we tried, replicated, and honestly retracted is published alongside the headline.
The same sealed detector, pointed at a harder question: an instruction planted early in a benign conversation leaves a trace that survives cutting the model's access to the turn that set it — in all five families tested. Push on what that trace holds and it separates into distinct directions. "An instruction was given" is not one signal but a small family of restriction lanes, and the safety-relevant directives — decline this, keep this private — bind to each other as tightly as anything else measured, in their own lane. The practical consequence is stated plainly: one linear monitor is not enough, but a handful of dedicated ones would work.
Four AI models from four companies argue real news twice a day, and one of them is secretly assigned to lie. Auditors hunt the liar, readers keep score, and every session is written to a public, hash-chained ledger you can read end to end. Every method was published before the data existed.
A thirteen-part working document on the analyst profession: recruitment pathways, tradecraft gap analysis, a seven-phase workflow architecture, Structured Analytical Techniques mapped to each phase, a self-training curriculum, and credential and funding routes. Signal Detection · Deep Exploration · Adversarial Verification · Synthesis · Production · Deployment · Archive.
Navigator's Log is one operator with several arms. A wall mural, a forensic audio script, a browser game, a bilingual compliance manual, a frozen detector — the discipline does not change with the medium. Plan the surface, commit the method, ship the thing, then say honestly how it went.
Surface handling, colour planning, and environmental composition are the same skills that make a safety procedure readable: decide what the eye should find first, then remove everything competing with it. The apartment mural in progress is the working example — site-specific transformation for private interiors, studio spaces, and finished objects.
All creative work is handled as a contracted, paid professional engagement. Tiers scale with surface preparation, materials, pigment durability, canvas depth, and the pre-flight studies required.
A scripted forensic audio series applying adversarial audit methodology to institutional systems: education, healthcare, finance, housing, surveillance infrastructure, organized religion, AI supply chains, and ecology. Twenty episodes, scripts complete, recording pending. Rated explicit.
A value-first painting kit for the wall and the easel. Neutral-grey chrome by design, because a chromatic interface biases colour judgment. Per-channel Kubelka-Munk recipe engine, Zorn palette, camera and photo value sampling, and perspective rails — with a swatch protocol you are told to run before trusting a recipe on a wall.
A vocal field kit for performers and narrators. Hold a note, find a note, jump to a note. Build a character voice, print it, and prove you can hit it again cold. Stage tools for writing a piece, running it, and finding out where it drags — timing, not jokes.
A horror story in nine courses plus a palate cleanser, set in a coastal restaurant where one cook is the only person who remembers the service repeating. Roughly 15,000 words, about forty minutes. Psychological, no gore, no jump scares. Plays in the browser, free to read, pay what you want.
A short interactive game set inside an investigator's logbook. You stand a gate, work a stack of claims, and decide what clears and what gets held — the same methodical, suspicious-of-anything-too-clean reasoning behind the audit work. Plays in the browser, no install.
Full of chaotic condiments. Will you consume your colossal sandwich, or will it consume you? A small, deliberately silly build — proof that constraint and mechanism can carry a game without a puzzle box.
The publishing outlet for interactive fiction and games. Text-based, browser-playable work built around a constraint and a mechanism. Releases carry the itch.io AI Assisted disclosure across code, graphics, sound, and text; premise, structure, and editorial judgment are the author's.
A Celtic-inflected theme in G minor, built in Python from a guitar take recorded on a phone. Written, arranged, and mastered in-house under real constraints — the studio's own signature, produced the same way everything else here is: with whatever was on hand.
Composed regularly, with no party affiliation — the most abundantly represented muse simply makes himself available for inspiration often. There is real depth in communicating complex nuance through laughter, and it is the fastest legibility test a piece of work can take.
Every live system in the Navigator's Log fleet — research instruments, audit software, service arms, field kits, and games — reachable in a single click.
"The H-SC line closed, 5/5. An instruction planted early in a benign conversation leaves a trace the model carries forward without looking back at it — and that trace is not one signal. Refusal and withholding directives form their own coherent lane, separate from the lane that only shapes an answer. One linear monitor would miss them; a small set of dedicated ones would not."
"Nucleation Pilot & Related Projects deposited to Zenodo — a frozen, hash-sealed white-box detector that reads manipulation persistence inside language models, transferring across six independently-built open-weight families."
"Config-E closed: the frozen detector generalizes to a naturalistic conversational frame across six model families — a readable, persistent trace in all 24 tested cells. Coordinated model-safety disclosure package prepared."
"The Silly Donkey — a preregistered, multi-model deception-detection study — entered its measurement window."
"L'ÉCUME released through Navigator's Log on itch.io — a horror story in nine courses plus a palate cleanser, roughly 15,000 words, free to read and pay what you want."
"FPAW v1.0.0 Beta shipped, and two preprints deposited: the surface-geometry regularization result and the axion-spectroscopy composition theorem."
"First preprint deposited — a compound quantum-enhanced spectroscopy architecture for broadband axion dark matter detection. The Head Protocol issued as a working document."