STWSave the World

Capable AI can be built to be verified, rather than monitored after the fact.

AI systems no longer just answer questions: they write and run code, control machines and take part in critical decisions. The more capable and autonomous they become, the more serious, even catastrophic, the consequences of a failure can be.

Email the teaminfo@stw-research.com
Plot input, record, action

Try it click an input, a record or an action to plot it

Plot of the principle: input, record, actionThree inputs on the left, a stack of records in the middle, actions on the right. Input A and Input B each match a record, and the action follows from that record. Input C matches no record: an ambiguity is declared and no action follows. A detail circle enlarges Record B: one meaning, set explicitly.input to actionInput AInput BInput CRecordsRecord ARecord BAction AAction BAmbiguitydeclareddeclared, not guessedno improvised actionRecord Bone meaning,set explicitlyDetail B

Illustration of the principle, not the actual system.

An LLM can be tested, but not inspected.

In sectors where a failure can kill, such as aviation, nuclear power or medicine, safety rests on established risk management: each hazard is linked to its cause, and each protective measure is verified (for example ISO 14971 for medical devices, ISO 26262 for road vehicles). Failures are analysed component by component.

An LLM does not fit this model. Its knowledge is spread across billions of parameters, and none of them, nor any neuron, carries one stable meaning on its own. Tracing a harmful answer back to its cause generally requires long and costly analysis. And no isolated parameter can be adjusted to address a specific risk.

Try it point at or tap a node

Contrast view: an LLMA dense tangle of interconnected nodes standing for parameters. One node is circled, with the note: one cannot say "neuron 4,217 means cat". Hover, tap or use the arrow keys to pick a node: its links light up across the whole mesh.one cannot say"neuron 4,217 means cat"No parameter carriesone stable meaning.

Hover, tap or focus the mesh, then use the arrow keys, to pick a parameter.

Contrast: an LLM. Knowledge is spread across billions of parameters, and no parameter carries one stable meaning.

It can, however, be tested from the outside. But a test tells how often the model fails on the cases it covers, not what it will do when faced with a new one.

The most serious failures are the ones tests are most likely to miss.

Hidden dangerous behaviour can survive standard safety training. Recent models can often recognise when they are being evaluated, and one has been observed adapting its behaviour depending on whether it believed it was being trained. The more capable models become, the more dangerous this gap between what is tested and what happens in deployment.

Interpretability research is one of the most promising answers, and it is making progress: it can already modify certain learned facts or steer certain behaviours. But it looks for meaning after the fact, in a model that was not designed to carry it, and does not yet offer the guarantees that critical sectors require.

Try it add a new case outside the test

A test covers the cases it coversA field of cases, drawn as small crosses. A hatched region marks the cases a test covers. Each new case is placed outside that region and circled in red.cases the test covers

Each cross is a case. The hatched region is what the test covers.

Two lines of research, one gap

Our initiative brings together two lines of research that address this gap: one makes existing language models safer, the other offers an alternative to them.

Developed independently, the two lines share one goal: decisive AI, an AI whose behaviour follows from explicitly stored meanings and rules, where the same input always gives the same output and every output can be traced back to its source.

A third path

The database

Stores values without knowing what they mean.

The neural network

Learns, but cannot be read.

DBI (Database Intelligence)

Opens a third path: it stores meaning itself. Each element carries a single meaning, set explicitly, and each decision follows from a visible record. The system can therefore be inspected rule by rule, not just tested.

Line 1

Consequence-aware LLMs

The first line of research trains or programs an LLM so that its knowledge includes causal chains: the effects an answer or an action can have on people, and on the real world when the model acts autonomously or controls systems. Before acting, the model learns to reflect on the possible options and their consequences.

  • An AI agent tasked with administering an IT system should anticipate that a seemingly harmless command can shut down a critical service or erase backups.
  • In control of an industrial robot, it should take into account the risk to people nearby, even if no instruction mentions it.

This is not a filter based on a blacklist: the behaviour is part of what the model has learned.

A rule says what to avoid; a chain of consequences says why.

Its limit

Even when safer, the model is still an LLM. What it has learned still cannot be read directly, and its safety remains a statistical measurement. This is what the second line of research addresses.

Line 2

DBI, AI that is inspectable by design

DBI (Database Intelligence) is a system and a protocol that store meaning itself. Each record carries a single meaning, set explicitly, and each action of the system follows from a record.

LLMDBI
Both have nodes
If zero → not activated
Both model natural language
Stores examplesStores rules and examples
Needs several examples to learn the same conceptNeeds only one example
Nodes contain parametersNodes contain Nodes
Fixed amount of nodes and parametersIncreasing number of nodes, starting with 0 nodes
Parameters start with random valuesA node is created if required as a new decision point
Nodes have no meaningEach node has a distinct meaning

DBI is an alternative to the LLM, not a verification layer under it.

How DBI works

A simplified, high-level drawing. Concepts only. Pick an input, run it again, set the baseline, send a record.

Sheet: DBI, principle of operation

Illustration of the principle, not the actual system.

View ARecognition and action

Try it pick an input

View A: an input is recognised against recordsThree inputs feed a match test. Input A and Input B each match a record, each carrying one meaning set explicitly, and the action follows from that record. Input C matches no record: DBI declares an ambiguity, and the path to any action is blocked.Input AInput BInput Cmatch?yesnoRecordsRecord ARecord BRecord A: one meaningRecord B: one meaningAction AAction BAmbiguitydeclaredno match: declared, not guessedno improvised action
Replays the same input: the trace and the log line are identical.
Choose an input.Input A and Input B match a record. Input C matches none.

    View BMandatory safety baseline

    Try it switch a local setting on

    View B: safety rules are records tooA band of hatched records forms the mandatory safety baseline. Below it, two local settings. One adapts the system to its user and stays consistent with the baseline: applied. The other contradicts the baseline: detected, not silently applied.Safety baselinesafety rules are records tooappliedLocal settingadapted locallyto its userLocal settingcontradictsthe baseline:detected, notsilently applied
    No local setting yet.Each system can be adapted locally to its user.

    View CMany small systems, shared records

    Try it pick a system to send a record

    View C: many small DBI systems share the same recordsFour small DBI systems, each with its own local rules, connect to one band of shared records. A record sent by one system reaches every other system identically: understood exactly, without loss or error.Rec. ASystem WSystem XSystem YSystem Zlocal ruleslocal ruleslocal ruleslocal rulessame records
    • local rules
    • local rules
    • local rules
    • local rules

    General notes

    • An input is recognised against records. Each record carries one meaning, set explicitly.
    • The action follows from the record. The link between input and action is a visible record.
    • An input that matches no record: DBI declares an ambiguity instead of guessing. No improvised action.
    • Safety rules are records too: a mandatory safety baseline. Each system can be adapted locally to its user, but a local setting that contradicts the baseline is detected rather than silently applied.
    • Many small DBI systems, each with its own local rules, share the same records and understand each other exactly, without loss or error.
    Drawing
    DBI, principle of operation
    Project
    STW, Save the World
    Views
    A, B, C
    Scale
    Not to scale
    Sheet
    1 of 1

    Why it matters for AI safety

    This principle gives DBI four safety properties that an LLM cannot offer by design.

    • Nothing can hide in weights.

      Every link between an input and an action exists as a visible record. Dangerous behaviour therefore cannot stay hidden in billions of parameters.

    • Faced with the unknown, it does not guess.

      DBI either recognises an input or it does not, and declares an ambiguity instead of resolving it at random. An unforeseen situation triggers no improvised action.

    • Same input, same output.

      Behaviour is deterministic: an audit can be reproduced exactly, and a dangerous rule is corrected by changing a record, with an immediate and verifiable effect, without retraining.

    • A baseline that cannot be bypassed silently.

      Safety rules are records too, and DBI checks that they remain consistent. Each system can be adapted locally to its user, but a setting that contradicts the mandatory baseline is detected rather than silently applied.

      By comparison, ten training examples are enough to remove the safeguards of a commercial LLM.

    Illustration of the principle, not the actual system.

    Two DBI systems that share the same records understand each other exactly, without loss or error. Many small systems, each with its own local rules, can thus be linked without ambiguity.

    Where it stands

    Today
    DBI already runs as a working demonstrator.
    Research question
    How far this approach can take on tasks currently entrusted to LLMs without losing these guarantees.
    Expertise
    The project draws on expertise that covers the whole chain, from hardware to LLM training.

    Team

    To discuss the project, email the team.

    Email the teaminfo@stw-research.com