A Truthfulness Audit of DIALOGIC's Public Content
Making epistemic transparency apply to ourselves.
Public transparency
In August 2026 we ran a full AI agent truthfulness audit of our public content — comparing what we claim against what we have actually built. This is a summary of what it found, and what we changed as a result.
Why publish an audit?
DIALOGIC describes itself as a platform concerned with agency, uncertainty and epistemic integrity. That principle has to apply to DIALOGIC's own claims. A system should not ask people to distinguish evidence from interpretation while presenting its own aspirations as proof.
This public version shares the conclusions most relevant to that question. It leaves operational debugging and deployment notes in the development record rather than treating them as part of the intellectual argument.
Executive summary
DIALOGIC's public writing is strongest when it describes design principles, boundaries and research questions. The platform genuinely implements a substantial part of its stated epistemic and agency architecture.
- Context is divided into distinct layers.
- Important inferences are represented as provisional rather than factual.
- Session direction and relational state are bounded.
- Current participant corrections outrank remembered or inferred context.
- Sensitive profile use has consent and mode boundaries.
- Public trials are separated from private account memory.
- Longitudinal research has separate consent and aggregate-reporting safeguards.
The largest truthfulness risk is a category error between architecture, aspiration, anecdote and demonstrated outcome. The architecture is substantially real. The proposed developmental effects remain questions to be measured.
Overall judgement
Green means the claim is substantially supported by what we've built. Amber means it's directionally true but not yet fully realised. Red means the claim outran the evidence and we revised it.
| Area | Judgement | What that means |
|---|---|---|
| Epistemic and agency architecture | Substantially supported | Context is separated into layers; inferences remain provisional; participant correction and agency are built into the system's design. |
| ARIS continuity and contextual assembly | Supported, but conditional | Continuity exists, but depends on mode, account state, consent, enabled context, available history and feature boundaries. |
| Cross-feature understanding | Partially supported | Shared account context and Living Profile create real continuity in eligible contexts, not universally across every product or session. |
| Human-development and capacity outcomes | Aspirational / unverified | The platform is designed toward these possibilities, but does not yet have enough evidence to present them as established outcomes. |
| Public/private memory boundaries | Supported | Public ARIS and ECI experiences are separated from private profiles, history and uploaded knowledge. |
| Research consent and aggregate reporting | Supported | Participation has separate consent and aggregate reporting boundaries rather than feeding directly into personalisation. |
| Public product availability | Mixed | The audit found a gap between some items presented in public catalogues and what is currently substantively available. |
| Essay editorial quality | Conceptually strong; publication apparatus incomplete | The ideas are conceptually strong, while publication states, availability signals and some editorial wording still need to be made clearer. |
What we are correcting
The audit is useful only if it changes how the work is described. These are the highest priority corrections we are carrying forward:
01
Outcome language
DIALOGIC sometimes moves from a real architectural capability—such as retaining context or modelling a provisional direction—to language that sounds like a demonstrated human outcome: increased discernment, creative range, agency or capacity. We are revising that language so design aims and research questions are not presented as established effects.
02
Conditional continuity
Descriptions of ARIS can make continuity sound universal and always-on. We are making the conditions visible: relevant mode, account state, consent, enabled profile or memory context, available history and feature-specific boundaries all matter.
Final assessment
DIALOGIC's public thesis is not fundamentally unsupported. There is a real and unusually thoughtful architecture underneath the language: consent-aware context, provisional interpretation, session-scoped direction, relational-state boundaries, profile correction, isolated public trials, and separate aggregate research.
The credibility gap appears when the site speaks as though this architecture has already demonstrated the human outcomes it was built to investigate.
The most truthful public posture is also the strongest one:
DIALOGIC has built a serious experimental environment for relational, context-aware, agency-preserving interaction with AI. It has implemented meaningful safeguards and developed testable hypotheses about continuity, recognition, creativity, and human capacity. The architecture exists. The proposed developmental outcomes remain questions to be measured.