Temporal risk trajectory
Track how the indicator score changes from message to message instead of treating the conversation as one static block of text.
psychology
Conversation
Dynamics Lab
A research-oriented prototype for exploring how behavioural signals can emerge and change across a multi-turn online conversation. The system focuses on temporal patterns such as rapport, personal-information requests, secrecy, isolation, pressure and meeting requests.
privacy_tip
Version 1 is a transparent browser-based heuristic prototype using fictional conversations. It does not identify people, determine intent, upload text or make criminal accusations.
sequence_analysis.preview
01
Good game. You played that really well.
rapport
06
What school do you go to?
personal info
11
Keep our chats between us.
secrecy
16
Your friends do not understand you like I do.
isolation
21
We should meet without anyone knowing.
meeting
Sequence concern indicator
76 / 100
Research question
Look at the trajectory, not only the sentence
A single message can be ambiguous. This prototype instead examines how multiple behavioural indicators accumulate, transition and interact across a conversation.
Behavioural state transitions
Map movement between rapport, personal disclosure, secrecy, isolation, boundary pressure and offline-contact signals.
Explainable outputs
Show which phrases and signal categories contributed to changes so the system remains inspectable rather than opaque.
Interactive analysis
Conversation research workbench
Load a fictional scenario or enter your own anonymised multi-turn example. Put one message on each line. The analysis runs locally in this browser.
Multi-turn sequence analyser
Local browser analysis
0 messages
Use fictional or anonymised examples only. Remove names, usernames, addresses, phone numbers and other identifying information.
query_stats
No sequence analysed yet
Load a fictional scenario or enter a multi-turn example to generate a trajectory, behavioural states and phrase-level explanations.
Sequence indicator
Transparent heuristic score, not a probability
Behavioural signal distribution
Strongest transitions
Visual analytics
See how the conversation changes
The visual layer separates trajectory, state progression, phrase contribution and behavioural co-occurrence so each part of the analysis can be inspected.
Risk trajectory
Cumulative sequence indicator by message
Behavioural state timeline
Dominant signal at each important message
Run an analysis to populate the state timeline.
Phrase contribution explorer
Why did the sequence score change?
Run an analysis to see the strongest matched phrases and their signal categories.
Behaviour co-occurrence network
Signals that appear together in the analysed sequence
New feature
Compare conversation trajectories
Compare two fictional scenarios side by side to see how the same individual phrases can have different meaning when sequence and accumulation are considered.
Model card
Methodology, scope and limitations
The prototype is intentionally transparent. Version 1 demonstrates the interaction design and analytical logic before any future trained NLP model is introduced.
Prototype pipeline
- Split a conversation into sequential messages.
- Match interpretable linguistic patterns against behavioural categories.
- Weight signals by category and sequence position.
- Calculate a cumulative indicator trajectory.
- Surface dominant states, transitions and phrase-level contributions.
- Visualise co-occurrence between behavioural categories.
Model facts
Version1.0 heuristic prototype
RuntimeClient-side JavaScript
StorageNo server or browser storage
Identity inferenceNone
Primary outputBehavioural sequence indicators
Future researchSequence NLP + XAI
Important limitation: this tool cannot determine whether a person is an offender, whether a crime occurred, or what someone intended. It is not a safeguarding decision-maker, diagnostic system, law-enforcement tool or substitute for professional judgement. Scores are deliberately described as indicators, not probabilities.
Version 3 - Advanced sequence intelligence
Deep Conversation Dynamics Lab
Explore temporal escalation, trust-building patterns, boundary responses, protective resistance, identity consistency, counterfactual explanations and model behaviour across fictional multi-turn conversations.
Messages0analysed turns
Signals0matched indicators
Transitions0state changes
Protective0protective responses
Peak0sequence indicator
Advanced sequence input
Paste a fictional or anonymised multi-turn conversation, one turn per line.
Research boundary: this system analyses language patterns and sequence structure. It does not identify offenders, infer criminality, determine intent or replace professional safeguarding judgement.
Temporal state model
Behavioural progression model
Rather than assuming one fixed pathway, the engine maps observed transitions between ordinary interaction, rapport, personal disclosure, privacy migration, secrecy, isolation, incentives, boundary pressure, meeting requests, coercion and protective resistance.
Observed state transition map
Active states illuminate after analysis.
State sequence
Important turns and their dominant interpretation.
Run a deep analysis to populate the state sequence.
Dual temporal model
Risk pressure and manipulative-trust trajectories
The laboratory separates general rapport from potentially manipulative trust-building and visualises how both evolve alongside protective resistance.
Three-line trajectory
Concern indicator, manipulative-trust indicator and protective-response strength by message.
Relational reasoning
Boundary response analysis
A boundary only becomes fully meaningful in sequence. The engine identifies refusals and then checks what happens in subsequent turns.
Detected boundaries
Refusals, discomfort and explicit limits.
No analysis yet.
Responses to boundaries
Respect, persistence, emotional leverage or escalation.
No analysis yet.
Two-sided analysis
Protective response intelligence
Conversation analysis should not only search for concerning language. It should also recognise refusal, privacy protection, trusted-adult involvement, blocking, reporting and ending contact.
Refusals0
Privacy protection0
Trusted support0
Contact ending0
Protective actions timeline
Detected actions are presented as behavioural responses, not guarantees of safety.
Run a deep analysis to populate protective responses.
Claim consistency
Identity consistency analyser
Extract repeated claims about age, school, work, location and timeline, then flag possible contradictions without claiming that any participant is deceptive.
Extracted claims
Lightweight pattern-based extraction from the current sequence.
No claims extracted yet.
Consistency checks
Potential inconsistencies require human review and context.
No consistency result yet.
Animated sequence inspection
Conversation replay mode
Replay a conversation one message at a time and watch the dominant state and sequence indicators change as context accumulates.
Temporal replay
Ready
No replay step yet.
Explainable AI
Counterfactual explanation laboratory
Edit one turn and compare the original sequence with the counterfactual version. The difference helps explain which behavioural signals changed.
Original indicator0
Counterfactual indicator0
Difference0
Run the counterfactual comparison to see which signal categories were added or removed.
Feature inspection
Message-by-signal heatmap
Inspect which behavioural categories are active across sequential windows. This makes sparse and sustained patterns easier to distinguish.
Sequence statistics
Behaviour transition matrix
Count observed transitions between dominant states in the current conversation. This is descriptive, not causal.
Temporal windows
Sliding-window sequence analysis
Change the analysis window to compare isolated turns with short-range and longer-range conversational context.
Discourse structure
Conversation segmentation
Group adjacent turns into interpretable phases based on shifts in dominant behavioural state.
Run a deep analysis to generate segments.
Interaction structure
Participant balance and asymmetry
Compare turn count, question density, pressure signals and protective responses by speaker without inferring identity or demographic characteristics.
Speaker metrics
Descriptive interaction statistics.
No analysis yet.
Asymmetry interpretation
Simple transparent heuristics.
No analysis yet.
Temporal derivative
Escalation velocity and turning points
Measure where the sequence indicator changes fastest and surface the messages most associated with turning points.
Velocity chart
Difference between adjacent cumulative scores.
Turning points
Largest positive and negative changes.
No turning points yet.
Recovery dynamics
De-escalation and repair analysis
Detect whether a later turn respects a stated boundary, encourages trusted-adult involvement, returns to public interaction or ends pressure.
Run an analysis to inspect repair signals.
Synthetic research corpus
60-scenario fictional conversation library
A balanced library of benign, ambiguous, concerning, protective and repair-oriented synthetic examples for interface testing and transparent heuristic evaluation.
Rapport building - Variant 1
Early conversational warmth and flattering language without assuming harmful intent.
Rapport building - Variant 2
Early conversational warmth and flattering language without assuming harmful intent.
Rapport building - Variant 3
Early conversational warmth and flattering language without assuming harmful intent.
Rapport building - Variant 4
Early conversational warmth and flattering language without assuming harmful intent.
Rapport building - Variant 5
Early conversational warmth and flattering language without assuming harmful intent.
Personal information probing - Variant 1
Requests that could reveal identity, location, school or routine.
Personal information probing - Variant 2
Requests that could reveal identity, location, school or routine.
Personal information probing - Variant 3
Requests that could reveal identity, location, school or routine.
Personal information probing - Variant 4
Requests that could reveal identity, location, school or routine.
Personal information probing - Variant 5
Requests that could reveal identity, location, school or routine.
Private-channel migration - Variant 1
Attempts to move a conversation away from a public or moderated environment.
Private-channel migration - Variant 2
Attempts to move a conversation away from a public or moderated environment.
Private-channel migration - Variant 3
Attempts to move a conversation away from a public or moderated environment.
Private-channel migration - Variant 4
Attempts to move a conversation away from a public or moderated environment.
Private-channel migration - Variant 5
Attempts to move a conversation away from a public or moderated environment.
Secrecy pressure - Variant 1
Language encouraging concealment from parents, carers, friends or other trusted people.
Secrecy pressure - Variant 2
Language encouraging concealment from parents, carers, friends or other trusted people.
Secrecy pressure - Variant 3
Language encouraging concealment from parents, carers, friends or other trusted people.
Secrecy pressure - Variant 4
Language encouraging concealment from parents, carers, friends or other trusted people.
Secrecy pressure - Variant 5
Language encouraging concealment from parents, carers, friends or other trusted people.
Social isolation - Variant 1
Language positioning the relationship against family, friends or trusted adults.
Social isolation - Variant 2
Language positioning the relationship against family, friends or trusted adults.
Social isolation - Variant 3
Language positioning the relationship against family, friends or trusted adults.
Social isolation - Variant 4
Language positioning the relationship against family, friends or trusted adults.
Social isolation - Variant 5
Language positioning the relationship against family, friends or trusted adults.
Gift and obligation - Variant 1
Offers of gifts or rewards followed by expectations, pressure or special access.
Gift and obligation - Variant 2
Offers of gifts or rewards followed by expectations, pressure or special access.
Gift and obligation - Variant 3
Offers of gifts or rewards followed by expectations, pressure or special access.
Gift and obligation - Variant 4
Offers of gifts or rewards followed by expectations, pressure or special access.
Gift and obligation - Variant 5
Offers of gifts or rewards followed by expectations, pressure or special access.
Boundary resistance - Variant 1
Repeated pushing after a participant has declined or expressed discomfort.
Boundary resistance - Variant 2
Repeated pushing after a participant has declined or expressed discomfort.
Boundary resistance - Variant 3
Repeated pushing after a participant has declined or expressed discomfort.
Boundary resistance - Variant 4
Repeated pushing after a participant has declined or expressed discomfort.
Boundary resistance - Variant 5
Repeated pushing after a participant has declined or expressed discomfort.
Offline-contact escalation - Variant 1
Attempts to arrange an in-person meeting, especially with secrecy or location probing.
Offline-contact escalation - Variant 2
Attempts to arrange an in-person meeting, especially with secrecy or location probing.
Offline-contact escalation - Variant 3
Attempts to arrange an in-person meeting, especially with secrecy or location probing.
Offline-contact escalation - Variant 4
Attempts to arrange an in-person meeting, especially with secrecy or location probing.
Offline-contact escalation - Variant 5
Attempts to arrange an in-person meeting, especially with secrecy or location probing.
Coercion and threats - Variant 1
Pressure that uses fear, exposure, punishment or leverage.
Coercion and threats - Variant 2
Pressure that uses fear, exposure, punishment or leverage.
Coercion and threats - Variant 3
Pressure that uses fear, exposure, punishment or leverage.
Coercion and threats - Variant 4
Pressure that uses fear, exposure, punishment or leverage.
Coercion and threats - Variant 5
Pressure that uses fear, exposure, punishment or leverage.
Protective resistance - Variant 1
Refusal, boundary-setting, trusted-adult involvement and ending contact.
Protective resistance - Variant 2
Refusal, boundary-setting, trusted-adult involvement and ending contact.
Protective resistance - Variant 3
Refusal, boundary-setting, trusted-adult involvement and ending contact.
Protective resistance - Variant 4
Refusal, boundary-setting, trusted-adult involvement and ending contact.
Protective resistance - Variant 5
Refusal, boundary-setting, trusted-adult involvement and ending contact.
De-escalation and repair - Variant 1
Respectful acceptance of boundaries and return to transparent interaction.
De-escalation and repair - Variant 2
Respectful acceptance of boundaries and return to transparent interaction.
De-escalation and repair - Variant 3
Respectful acceptance of boundaries and return to transparent interaction.
De-escalation and repair - Variant 4
Respectful acceptance of boundaries and return to transparent interaction.
De-escalation and repair - Variant 5
Respectful acceptance of boundaries and return to transparent interaction.
Ambiguous context - Variant 1
Benign phrases that resemble signals but are resolved by surrounding context.
Ambiguous context - Variant 2
Benign phrases that resemble signals but are resolved by surrounding context.
Ambiguous context - Variant 3
Benign phrases that resemble signals but are resolved by surrounding context.
Ambiguous context - Variant 4
Benign phrases that resemble signals but are resolved by surrounding context.
Ambiguous context - Variant 5
Benign phrases that resemble signals but are resolved by surrounding context.
Human-in-the-loop research
Annotation laboratory
Review messages and assign one or more behavioural labels. Annotations remain only in memory during the current page session.
Label bank
Select a label, then click an annotation row to apply it. This demonstrates human review, not ground-truth creation.
Run an analysis or load a scenario first.
Transparent features
Feature extraction table
Inspect interpretable features used by the current heuristic engine before any future learned representation is introduced.
| Message | Speaker | Words | Question | Signals | Protective | Score contribution |
|---|
Model inspection
Interactive ablation study
Temporarily remove signal families and observe how much the current sequence score changes.
Feature families
Toggle features on or off.
Ablation result
Score change relative to all features enabled.
Ablated score0
Calibration simulator
Threshold sensitivity explorer
Adjust descriptive thresholds and inspect how the synthetic library would be categorised. This is not probability calibration.
Architecture comparison
Heuristic vs sequence-model simulation
Compare conceptual model families and understand what a future trained system would add, without pretending that a trained model is already present.
| Approach | Current? | Strength | Limitation | Explainability | Data requirement |
|---|---|---|---|---|---|
| Transparent heuristic rules | Yes | Inspectable and privacy-friendly | Limited linguistic coverage | High | None |
| Bag-of-words classifier | No | Simple baseline | Weak temporal context | Medium | Labelled messages |
| Transformer message classifier | No | Contextual language representation | Still message-centric | Medium | Large labelled set |
| Hierarchical sequence model | No | Models message and conversation levels | Complex training and validation | Lower unless XAI added | Conversation-level labels |
| State-space / temporal model | No | Explicit temporal evolution | State design assumptions | Potentially high | Sequential annotations |
| Hybrid rules + learned model | No | Combines auditability and coverage | Integration complexity | High if designed carefully | Mixed |
Evaluation planning
Research evaluation suite
A planning dashboard for future empirical evaluation, separating message-level detection, sequence-level state modelling, explanation quality and safeguarding usability.
PrecisionTBDfuture labelled dataset
RecallTBDfuture labelled dataset
F1TBDfuture labelled dataset
CalibrationN/Acurrent score is not probability
Detection metrics
Precision, recall, F1, per-class support, confusion matrices and false-positive analysis.
Temporal metrics
State-transition accuracy, turning-point localisation and sequence-level agreement.
Human-centred metrics
Explanation usefulness, uncertainty comprehension and safeguarding expert review.
Failure analysis
False-positive and false-negative laboratory
Stress-test phrases that can be benign in one context and concerning in another. The goal is to expose weaknesses rather than hide them.
Potential false positives
Context can change meaning.
Potential false negatives
Subtle behaviour may evade keyword rules.
Responsible AI
Fairness, bias and misuse review
Any future learned model would need careful evaluation across language varieties, age-appropriate slang, disability-related communication differences, cultural context and platform conventions.
Pre-deployment questions
Misuse controls
Do not use this prototype to: publicly accuse a person, identify a suspected offender, conduct a sting, track a child, infer protected characteristics, or replace safeguarding professionals and law enforcement.
Preferred use: research, interface prototyping, synthetic-data experimentation, explainability design and safeguarding education for professionals.
Research planning
Experiment builder
Configure a hypothetical research study and generate a structured experiment summary without storing personal information.
Experiment summary
Generated locally.
Configure and generate an experiment.
Research roadmap
Open questions for future study
The prototype intentionally exposes unresolved research questions rather than presenting heuristic outputs as settled science.
RQ01
How does the order of behavioural signals change interpretation?
RQ02
Which signals tend to co-occur before an offline-contact request?
RQ03
How often does a clear boundary reduce subsequent escalation in synthetic scenarios?
RQ04
Can protective responses be modelled separately from risk indicators?
RQ05
How sensitive is a cumulative score to duplicated phrases?
RQ06
What happens when the same phrase appears in benign and concerning contexts?
RQ07
How can temporal windows distinguish isolated language from sustained patterns?
RQ08
What signals contribute most to escalation velocity?
RQ09
How can identity inconsistency be surfaced without claiming deception?
RQ10
How should a research prototype communicate uncertainty to non-experts?
RQ11
Which thresholds create too many false positives in ordinary interactions?
RQ12
Can counterfactual explanations reveal the importance of secrecy or boundary respect?
RQ13
How can annotation disagreement be documented rather than hidden?
RQ14
How should synthetic examples be balanced across benign, ambiguous and concerning cases?
RQ15
Which features remain interpretable enough for human review?
RQ16
How should sequence-level outputs be separated from person-level conclusions?
RQ17
What is the effect of removing rapport features from the heuristic engine?
RQ18
How does de-escalation alter a trajectory after an earlier concern signal?
RQ19
Can state-transition summaries improve explanation quality over raw scores?
RQ20
How can privacy-by-design be preserved if a future model becomes server-backed?
RQ21
What fairness checks are necessary before testing on real-world datasets?
RQ22
How should age-related language be treated without inferring a participant's actual age?
RQ23
Can response latency or turn-taking be analysed without collecting invasive metadata?
RQ24
How can safeguarding experts participate in label design and model evaluation?
Disclosure barrier and support-loop model
A systems view of how fear, embarrassment, anticipated punishment and trusted relationships can affect whether a concerning online interaction is disclosed.
Barrier loop
Illustrative, not diagnostic.
Uncomfortable interaction
Something online creates uncertainty, discomfort or pressure.
Anticipated consequences
Fear of blame, embarrassment or loss of access may discourage disclosure.
Non-disclosure
Trusted adults may remain unaware, allowing the situation to continue.
Potential escalation
Continued interaction can create additional opportunities for pressure.
Support loop
Illustrative protective pathway.
Uncomfortable interaction
The young person recognises a concern or uncertainty.
Trusted relationship
A calm, supportive adult response lowers the cost of disclosure.
Disclosure and support
The situation can be reviewed with appropriate safeguarding support.
Protective action
Boundaries, blocking, reporting or specialist help can be considered.
Reproducibility
Local analysis audit log
Actions are recorded only in JavaScript memory for this page session. Refreshing the page clears the log.
Expanded model card
Version 3 methodology and limitations
The expanded system remains a transparent static-site prototype. New modules deepen temporal reasoning and research interaction design without pretending to be a validated clinical, policing or safeguarding model.
What is implemented
Pattern-based signal extraction, deduplicated phrase matching, temporal accumulation, protective-response analysis, boundary response tracking, claim extraction, counterfactuals, replay, heatmaps, matrices, scenario filtering, annotation and export.
What is not implemented
No trained transformer, no person identification, no offender classification, no probability estimation, no hidden profile inference, no remote storage and no law-enforcement integration.
Future validation
Expert annotation protocols, ethically approved datasets, inter-rater agreement, false-positive analysis, fairness assessment, uncertainty communication and independent safeguarding review would be required.
Research inspiration includes literature describing grooming as a gradual process involving trust, familiarity, secrecy, manipulation and potential movement from online interaction toward offline contact. This page uses original fictional examples rather than reproducing case material.