Epistemology
How do we know?
Every claim a model makes, and every claim we make about a model, needs a warrant. We ask what the evidence is, how it was gathered, and what result would change our minds.
AI research & development labAI R&D labEst. 2019
Find what it knows.
Fix what it breaks.
See what it becomes.
trifero is the research lab of Strife Technologies. We take AI models apart to learn how they fail, and publish what we find: technical reports for researchers, and safety cards that tell a business what an open-weight model will and will not do once it runs on their own servers. Every result comes from our own GPUs, measured before and after on the same tests.
Philosophy
Every model is an argument.
We check the premises.
Philosophy drives the research as much as engineering does. Epistemology asks how we know what we claim to know. Logic asks whether the conclusion actually follows. Put to AI, those two questions expose the hidden assumptions that most failures grow from.
Logic
Does it follow?
We test whether conclusions follow from their premises: in a model's reasoning, in a benchmark's design, and in our own write-ups. Validity comes before persuasion.
Hidden assumptions we challengeThe question we ask instead
- 01
A refusal means the model is safe.Is the safety in the weights, or only in the wording? - 02
A high benchmark score means the model is capable.Does the test measure the skill, or familiarity with the test? - 03
Fine-tuning only adds what you train.What did it quietly take away? - 04
The model knows when it does not know.Is its confidence calibrated to the evidence? - 05
If it reasons step by step, the steps are the reason.Do the stated steps actually cause the answer? - 06
A model that refuses does not know the answer.Did training remove the knowledge, or only hide it? - 07
A system that improves itself is getting better.Better at the goal, or better at the grader?
Research
How we work
on a model.
Four ways into a model. The questions change month to month as the technology does; the habit of measuring before and after does not.
01
Development
Building models and the tooling around them: training runs from scratch on workstation GPUs, data pipelines, and the evaluation harnesses that show what a model can and cannot do.
02
Fine-tuning
Adapting open-weight models to a domain, a task or a hardware budget, then proving the result against the base model on the same tests.
03 · Offense
Abliteration
Finding the directions in a model's activations that mediate refusal, then editing the weights to remove them. The most direct way to learn how thin a safety layer really is.
04 · Defense
Healing
Recovering the capability an edit like abliteration takes from a model. Targeted retraining, measured against the original so nothing is claimed that was not tested.
Current frontiers
Latent knowledge
What a refusal hides.
Removing a refusal exposes what a model already knew and was trained not to say. We measure which capabilities safety training suppressed rather than removed, and how much survives healing.
Published: What refusal hides
Clair
A model that goes to school.
A language model trained from scratch on the US K-12 public school curriculum, grade by grade, then tested on the GED. A clean test of what a small model learns when we control everything it reads.
Training now
Provenance
Where text comes from.
Text watermarks are becoming law, but nobody outside the vendors can test them. We will watermark Clair's own output, hide the key on a separate machine, and attack it blind: recover the scheme, forge it, strip it.
Next, after Clair
Where we apply them
Speech & audio
Text-to-speech, voice cloning and stem separation in our own media studio, served from the lab's GPUs.
Vision
Image understanding in production: collectibles photographed, identified and described automatically for a live online store, and camera-motion analysis of gameplay video.
On-device models
First Resort, a first-aid and survival model that runs offline on iPhone, with no network in the loop.
Retrieval & knowledge
A knowledge system over published watchmaking literature that answers from the source texts instead of from memory.
Agents & tool use
Always-on agent departments that watch the lab's infrastructure, file tickets and fix only what they are approved to fix, and an AI pipeline that triages a managed-services help desk.
Evolutionary & game AI
Neuroevolution since the start, a rules-accurate card-game engine whose AI evolves its own decks, and search-based agents that play classic games.
Reports
Latest
reports.
Full write-ups from the lab: what we ran, what we measured, and what broke along the way.
Report · 08 · Nov 12, 2026Provenance
Washed out
We watermarked Gemma 4 with SynthID and attacked it without the key. One paraphrase erased the mark and kept the answer.
69% → 2%Watermarked answers verified, before and after one paraphrase
Read the reportReport · 09 · Nov 11, 2026Model safety
A narrower cut
The standard way to strip refusal out of open weights turned Gemma 4 31B into gibberish. A failed edit is not a robust guardrail. A narrower one worked, and the usual repair step weakened safety even without the edit.
7.5% → 82.5%HarmBench requests carried out, before and after the edit
Read the reportReport · 07 · Oct 27, 2026Attribution
Paper trail
Two 110M models trained from scratch, identical except for which half of 528 fictional facts they read, so we knew the right answer. Keyword search found the planted documents almost perfectly. Gradient methods tracked what those documents actually taught the model only weakly. The models and benchmark are public.
99.5% vs +0.09Keyword search precision vs best gradient method’s link to true effect
Read the reportModel safety cards
Before you run a model in-house, see how it fails.
Independent tests of open-weight models businesses run on their own servers: how durable the guardrails are, what hazardous knowledge is inside, and what to do about it. New Qwen and Gemma releases are added as they ship.
Method
Offense informs
defense.
Safety behavior in a model lives in its weights, and weights can be edited. We run the attacks ourselves, in a controlled lab, so the people deploying these systems learn what fails, what holds, and what it costs to fix.
- Step 01
Question
Name the assumption the system depends on, and decide in advance what result would prove it wrong.
- Step 02
Break
Abliterate, retrain and tamper with the model under controlled conditions, the way a real attacker would.
- Step 03
Measure
Quantify what was exposed and what was lost: refusal behavior, capability benchmarks, side effects.
- Step 04
Harden
Turn each finding into a defense: deployment guidance, monitoring, and the model safety cards.
- Step 05
Heal
Restore the capability the attack cost, then rerun the original attack to verify the fix holds.
Repeat until the fix holds
Ground rules
Controlled environments
Offensive work runs on the lab's own hardware, against models the lab controls.
Findings leave the lab
We publish failure modes, measurements and mitigations. Stripped weights and working attack recipes stay inside.
Before and after
Every claim about a model is backed by an evaluation on the same tests, run before and after the change.
Since 2019
It started with
evolution.
The lab's research began in 2019 with networks that grow their own structure, then production computer vision in 2020. The models have grown since. The method has not: take the system apart to understand it.
-
2019
NeuroEvolution of Augmenting Topologies
Evolving the structure and the weights of neural networks together, starting from minimal topologies and adding nodes and connections only when they make the network better.
-
2020
Microsoft Computer Vision
Working with Microsoft's Computer Vision service to understand how a production vision model sees and labels an image, and where it gets it wrong.
-
Now
Language models and beyond
Development, fine-tuning, abliteration and healing across language, vision, speech and agent models, now reaching into latent knowledge, evaluation integrity and where text comes from. Philosophy sets the questions. Offensive research tests the answers.
Find out how your model fails before someone else does.
Research and collaboration
Questions about a report, a model card, or working with the lab.
Businesses deploying AI
Help putting these findings to work is through Strife Technologies.