Mindtro Research

Explore what comes next.

Research, experimentation and collaboration in agentic AI.

Each study explains what we tested, what we learned and what remains unresolved. Results apply to the tested conditions and are not a guarantee of product performance.

01 · Pyro · temporal intelligence

Pyro Time-Aware LLM

Pyro explores whether a language model can use the time that actually passes between events. If the text stays the same but the timing changes, can the model change its answer correctly? We study interval selection and duration comparison in separate experiments on a small model.

Provide elapsed time directly as numerical input
Compare correct timing with randomly mixed and zeroed timing
Measure performance on time intervals that differ from training conditions

Explore the programme →

02 · Pyrodit RM · risk intelligence

Pyrodit RM — detecting risk in text

Pyrodit Risk Model studies harmful-content detection while preserving legitimate expression. The latest English study examines training continuity, data coverage and contextual precision under explicit acceptance criteria. The report separates measured English results from future language evaluation.

Evaluate detection sensitivity and false alarms together
Compare full-model continuation with a fixed word-pattern control
Check context and language-specific behaviour before accepting a model

Explore the programme →

03 · Pyrodit Security · security research

Make the basis for security decisions visible

A security check should explain what it examined and which evidence supports its finding. We study how text scanning, sensitive-data masking and permission decisions can use these findings with clear responsibilities.

Find risk signals using rules, name recognition and analysis of meaning
Scan reports findings; Guard applies organisational rules to decide what happens next
Masking hides sensitive text while keeping a record of the transformation

Explore the programme →

04 · Pyrodit Research · emerging systems

Autonomous Enterprise OS

We are researching how specialist AI assistants can participate in business processes under human oversight. The Enterprise Agentic Runtime approach calls for explicit tasks, permissions and resource limits, with consequential actions open to review and reversal. This is an active research direction.

Record workflow steps and rules in definitions called FlowDefs
Set explicit limits on each assistant’s tools, budget and decision authority
Design human approval, action records and recovery for consequential operations

Explore the programme →

05 · Open source · Vector compression

Semafold

Semafold is Mindtro’s open-source Python library for compressing the numerical representations used by AI systems. It helps teams reduce embedding storage and explore smaller attention-cache tensors, with the saved bytes and reconstruction error visible together.

Embedding and vector storage
Attention-cache tensor experiments
Measured size and quality trade-offs

Explore the programme →

Research record

Experiments by research track

These are technical research reports, not peer-reviewed publications. Each report identifies its source, measurement conditions and limitations. The evidence section explains what can be downloaded or reproduced.

Understanding time · controlled experiment

The model uses timing information to select the right interval

Does the model use actual timing information to select the correct interval between events, or does it guess from clues in the text?

Experimental criteria met · three training runs

The experiment tests selection of a time interval from given options. It does not establish a ready-to-use language model that can answer any time-related question. An earlier experiment also left reliable matching of repeated intervals to events unresolved.

Read the report →

Comparing durations · training experiment

Duration reasoning: from comparison to numerical answers

Does also teaching the model ratios between durations help it identify which duration is longer and estimate its value?

Partial progress · numerical-answer target not met

This is not a completed system for answering duration questions. Further work must align predictions with real durations and limit answers when the model is uncertain.

Read the report →

Pyrodit RM · R7 recovery study

Risk detection: training continuity and contextual precision

Can retaining the full model and restoring broader training data improve detection without increasing false alarms or losing existing capabilities?

R7 evaluation complete · R5 retained as the serving baseline

These are English development results, with a separately selected threshold for each evaluation group, not one validated production threshold. They do not establish performance in Turkish, German or other risk categories. The context set had been examined before and is not an untouched independent test. No final context test was run for R7 because no candidate qualified; the reported context figures belong to the separate word-pattern control.

Read the report →

Open source · Vector compression

Semafold: smaller vectors, measurable trade-offs

How much space can AI vectors save when the complete compressed representation, including metadata, is counted?

Synthetic benchmark · source report 0.1.0, 31 March 2026 · library reviewed at 0.2.0

Compression is lossy. Real retrieval quality and end-to-end model behaviour need testing on the intended workload. Very small cache blocks can grow relative to 16-bit storage because of metadata overhead. The compression and cache APIs remain preview interfaces; ready-made serving integrations are outside this report.

Read the report →

Research reports

These reports explain completed experiments and research plans. They have not yet undergone independent academic peer review.

Temporal intelligence

The model uses timing information to select the right interval

Correct timing substantially improved interval selection. The improvement also held when the text stayed identical, supporting the conclusion that the model uses timing rather than relying only on textual clues.

Read the report →

Objective design

Duration reasoning: from comparison to numerical answers

Adding a task that teaches duration ratios improved how the model compares durations. That improvement was not sufficient to produce accurate numerical answers for the user.

Read the report →

Risk modelling

Risk detection: training continuity and contextual precision

The R7 study evaluated full-model training continuity and broader data coverage against predefined acceptance criteria. It clarified the trade-off between risk detection and contextual false alarms, while retaining R5 as the serving baseline. The findings inform the next controlled comparisons of language pretraining and label consistency.

Read the report →

AI security

Make the basis for security decisions visible

Read the report →

Open source · Vector compression

Semafold: smaller vectors, measurable trade-offs

Read the report →

Pyrodit Scan Lab · Reasoning evaluation

Does the answer hold up under scrutiny?

A fluent answer can still miss the goal or abandon a correct conclusion under pressure. Scan Lab’s reasoning engine evaluates the answer you can observe, the task it must solve and its response to feedback.

Explore the programme →

Know what a security check actually examined

A risk signal is not a decision on its own. The examined input and supporting evidence must be visible; organisational rules determine the action.

Explore the programme →

Our direction for enterprise AI

We are researching how AI can carry out organisational processes while preserving authority limits, human approval and accountability. These are research plans, not completed product capabilities.

Explore the programme →

Build with Mindtro

Let’s find the right approach
for your organisation.

Start a conversation ↗

Explore what comes next.

Pyro Time-Aware LLM

Pyrodit RM — detecting risk in text

Make the basis for security decisions visible

Autonomous Enterprise OS

Semafold

Experiments by research track

The model uses timing information to select the right interval

Duration reasoning: from comparison to numerical answers

Risk detection: training continuity and contextual precision

Semafold: smaller vectors, measurable trade-offs

Research reports

The model uses timing information to select the right interval

Duration reasoning: from comparison to numerical answers

Risk detection: training continuity and contextual precision

Make the basis for security decisions visible

Semafold: smaller vectors, measurable trade-offs

Does the answer hold up under scrutiny?

Know what a security check actually examined

Our direction for enterprise AI

Let’s find the right approachfor your organisation.

Let’s find the right approach
for your organisation.