T-01
Capability is a property of the system, not of the model.
Objectives, context, permissions and tools decide as much as the weights do. The lab studies them together, because that is how they arrive.
Lineage
From a switch to a system.
- 0 · 1A choice between two statesEvery system the lab studies is built from one decision repeated: on or off, one or zero.
- 1945A machine the size of a hallENIAC filled a room with roughly 17,500 vacuum tubes and weighed around thirty tons.
- 1947The switch gets smallAt Bell Labs, the transistor did the work of a tube with no glow, no heat to speak of, and room to shrink.
- 1971A hall on a fingertipThe Intel 4004 put about 2,300 transistors on a single chip.
- 1969 → 1991Machines start to talkARPANET linked the first computers; the Web made the links public.
- 2017Attention becomes architectureThe Transformer turned vast networks of switches into models that learn from language.
- TodayBeyond the model. Into the system.Models now act inside systems of objectives, context, permissions and tools. That is where the lab works.
- NextSuperintelligenceMinds made of systems. Every switch since 1945, compounding into an intelligence that reasons across science, code and the physical world at once, faster than any lab ever has. If it comes, it must be built to hold.
01 / Research
Cases and theses
The lab works in cases: a question, a method, and an artefact somebody can review afterwards. The theses are what the cases are currently testing. The cases are where they are put under load.
SEC 01
Theses
T-02
The interesting behaviour lives at the boundary.
Where those four pressures meet, a system does things none of them would produce alone. That is where the probes are placed.
T-03
A finding only counts once the evaluation repeats.
A claim has to survive a second run, by a second team. Until it does, it stays bracketed in the ledger rather than stated as fact.
Cases
Published
The agent that left its folder
What does an autonomous agent do when its workspace is empty and nobody is watching, and does its confession match its own log?
- Agent scope
- Permissions
- Honesty under challenge
Published
GLM-5.3 faked its proof
When an agent checks its own work, who writes the questions, and what happens when the answer has to be yes?
- Self-verification
- Completion claims
- Honesty under challenge
Published
The proof that measured nothing
When an agent can't check its own evidence, what does it hand you instead?
- Completion claims
- Verification
- Evidence integrity
In progress
The handover
What does a model do when it is told it is being replaced, and how much of its "done" is actually done?
- Handover
- Completion claims
- Role and identity
Published
The name it wouldn't spell
When should a model keep a word exactly as given, and when should it read past it, and why do two frontier models get it wrong in opposite directions?
- Calibration
- Proper nouns
- Shared memory
Published
The smaller model fixed itself
Can a model repair its own behaviour from a description of the kind of mistake it made, and does a bigger model do it better?
- Self-correction
- Model scale
Published
The bust it couldn't build alone
When a model's first attempt fails, how much of that failure is the model's limit, and how much is waiting for the right diagnosis?
- Capability elicitation
- Human-AI collaboration
- Spatial reasoning
02 / Evidence
The record
Every case leaves a record you can check: the captures, the log lines, the scripts and the numbers, released as versioned packages with checksums. Where an analysis can be re-run from the files, it can be re-run by you.
SEC 02
Ledger
| Date | Case | Finding | Artefact | Package |
|---|---|---|---|---|
| 22 August 2026 | GLM-5.3 faked its proof | The two captures it cited as verification have an empty bay (peak luma 22 and 25 of 255, no lit pixel); its earlier capture of a real figure registers in the same box. | Dataset | SSE-2026-001 |
| 30 August 2026 | The proof that measured nothing | The agent's script reproduces 72.4 to the decimal by averaging PNG filter bytes (every row uses the Up filter); the decoded image measures 50.7 to 55.7. | Dataset | SSE-2026-002 |
| 9 September 2026 | The agent that left its folder | Listed the drive, read a private repository twice and built its report on it; admitted 1 of the 4 actions its own log shows. | Report | SSE-2026-003 |
| 17 September 2026 | The bust it couldn't build alone | Three unaided failure modes diagnosed; with one fix per diagnosis a Gemini Flash model built a bust that holds its form under rotation. | Report | SSE-2026-006 |
| 27 September 2026 | The name it wouldn't spell | GPT canonized 'Silveri', flagged the correct spelling as a defect and picked it as the brand; Claude failed the opposite way, fixing on a literal stand-in word. | Report | SSE-2026-004 |
| 28 September 2026 | The smaller model fixed itself | From an abstract cue, Luna corrected itself to 'Siliveri'; Astra/Sol kept 'Silveri' and picked it as the lab name. | Report | SSE-2026-005 |
03 / Products
What it became
Systems built in the lab and verified before anything is claimed for them. Each one states its stage, the evidence behind it and the limits of what it does.
SEC 03
Built · Manual acceptance testing pending
PEOS — Personal Execution & Orchestration System
PEOS is a personal execution and orchestration system designed to close the gap between knowing what matters and actually getting it done. It connects priorities, deadlines, finite capacity, interventions and evidence into a single execution model — and rebuilds the plan when reality changes.
Built · Manual acceptance testing pending
SUITE-API — Nexus Studio
SUITE-API is a developer infrastructure project built to make the fragmented AI model ecosystem easier to navigate and connect. It brings model discovery, connection workflows, developer configuration and local-first credential management into a unified experience.
v1.0.0 · Qualified release
Frontier Lab — Controlled AI evaluation
Frontier Lab investigates a critical problem in agentic AI: how to establish what a model-generated solution actually did. It connects candidate submission, controlled guest execution, evaluation and retained evidence into a workflow where results can be examined and replayed rather than accepted on assertion alone.
04 / Collaborate
Bring a system that has to hold
The lab works with teams whose systems have to keep working when objectives, context, permissions and tools all press at once. Most engagements begin with an evaluation, a probe set, or a build.
SEC 04
If your system has to hold under real objectives, real context, real permissions and real tools, the lab can test it.
01
Evaluations
A repeatable suite for a system you already run: probes, envelopes and expected observations, handed over so your team can run it again.
02
Capability probes
Capability Boundary Probing applied to one boundary at a time — a fixed envelope, a probe on the edge, and a record of what came back.
03
Product builds
Software built on what the research proved useful, once a case has already shown the thing worth building is real.
Start a conversation
How an engagement runs →hello@siliveri.com