Skip to content
Janus
Menu

02 / WHAT RUNS TODAY

STATUS / RUNS TODAY

The system behind Janus already runs.

The current prototype was built for a demanding form of long-term personalization: an AI counterpart expected to understand the same person across months of ordinary conversation and change how it behaves as that understanding develops.

That requirement forced Janus to solve more than memory.

STATUS / RUNS TODAY

Maturity / engineering

  1. 01persistent state across sessions
  2. 02new evidence → revisable hypotheses
  3. 03hypotheses → expectations
  4. 04later evidence checks expectations
  5. 05model changes later behavior
Five things the deployed system already does, as a scan layer over the sentence that follows. What this claims is engineering maturity — that the machinery exists and runs — and not that any enterprise outcome has been demonstrated.

The deployed system already maintains persistent state across sessions, turns new evidence into revisable hypotheses, uses those hypotheses to form expectations, checks later evidence against them, and lets the resulting model change later behavior.

The enterprise product is a new application of that machinery — not a promise to invent longitudinal inference from scratch.

01 / More than memory

More than memory

Janus does not treat a transcript as a model of a person.

Its inference system maps observations into a layered psychological model, keeps competing hypotheses with confidence and provenance, and preserves every revision rather than silently overwriting the previous state.

A hypothesis can also generate expectations about what should appear later.

STATUS / RUNS TODAYVIEW / SCHEMATIC

Longitudinal inference

  1. 01Observation
  2. 02Hypothesisv1v2
  3. 03Expectation
  4. 04Later evidence
    • Confirm
    • Contradict
    • Unresolved
  5. 05Revision
  • what Janus believed
  • why it believed it
  • what it expected
  • how reality changed the model
An observation supports a hypothesis; the hypothesis owes an expectation; later evidence confirms it, contradicts it, or leaves it unresolved; the model is revised and the prior version is kept. The three outcomes are epistemic states, not success and failure.

A hypothesis has to survive contact with what happens next.

When new evidence arrives, those expectations can be confirmed, contradicted, or remain unresolved. Confirming evidence can strengthen the hypothesis. Repeated contradictions can weaken it, make it dormant, or force the system to consider alternatives.

That makes the longitudinal record more than a memory store. It is a history of what Janus believed, why it believed it, what it expected, and how reality changed the model.

02 / Four distinctions

Four distinctions the architecture keeps separate

01 / Facts ≠ hypotheses

Facts are not hypotheses

Janus keeps three kinds of knowledge separate:

  • raw observations and memories;
  • a structured personal knowledge base of biographical facts;
  • a psychological model made of hypotheses.

They can inform one another without collapsing into the same thing.

A fact such as a job change, family history, or stated preference can become evidence for a psychological hypothesis. In the other direction, the current psychological model can change which known facts are most relevant to retrieve.

A fact can inform a hypothesis without becoming one.

02 / Inference ≠ response

STATUS / RUNS TODAYVIEW / SCHEMATIC

Inference is not response

One of Janus's core architectural boundaries separates what it calls the subconscious from the consciousness layer.

The subconscious side can hold the deeper analytical state: psychological hypotheses, confidence, evidence, relationship history, affective state, internal strategy, and safety context.

The response-generating side does not receive that raw model.

It receives bounded behavioral guidance: how much directness, support, challenge, pacing, distance, or restraint may be useful.

Raw psychological labels, scores, internal IDs, and model mechanics remain behind the boundary.

Subconscious / deeper analytical state

  • hypotheses
  • confidence
  • evidence
  • relationship history
  • affective state
  • internal strategy
  • safety context

raw labels / scores / internal IDs / model mechanics

Consciousness / response-generating side

Bounded behavioral guidance

  • directness
  • support
  • challenge
  • pacing
  • distance
  • restraint

Runtime boundary

The boundary is enforced at runtime, not only by prompting.

Janus can therefore use a richer internal model without turning every inference into something the system says about the person.

03 / State ≠ trait

State is not trait

Janus does not treat the same person as psychologically identical from one moment to the next.

Alongside slower portrait hypotheses, the current system can maintain shorter-lived affective and biological state. Signals such as fatigue, sleep disruption, circadian state, illness, stress, focus, or sensory load can change activation thresholds, expression, and how much confidence the system places in a current observation.

Short-lived state

  • fatigue
  • sleep disruption
  • circadian state
  • illness
  • stress
  • focus
  • sensory load

Modulates / does not rewrite

Slower portrait hypotheses

The important distinction is that these signals modulate the existing model rather than rewriting it.

A tired person does not suddenly acquire a different personality. The same underlying pattern may simply become easier to trigger, while the observation itself may become less reliable evidence about stable character.

04 / Context ≠ character

Context is not character

Janus also models environmental context.

The implemented environment system distinguishes between signals that are heavily constrained by circumstance and signals that reflect more individual agency.

A pattern produced by workload, role, or temporary constraint should not carry the same psychological weight as a pattern the person repeatedly chooses across different contexts.

The system therefore weighs constraint, agency, cross-domain consistency, trajectory, and signal quality before allowing environmental observations to influence deeper hypotheses.

Environmental context can also change how behavior is interpreted over time without automatically creating a new trait.

  • constraint
  • agency
  • cross-domain consistency
  • trajectory
  • signal quality

Weight before deeper inference

The aim is to distinguish what may be changing in the person from what may be changing around them.

03 / Behavior

The model changes behavior

The portrait is not there to produce a better summary of the user.

It changes how Janus responds.

The current adaptation path combines the longitudinal portrait with live cascade state, learned feedback, and relationship context, then translates the result into behavioral guidance for the response layer.

That guidance can change tone, pacing, directness, warmth, challenge, and what the system should lean into or hold back.

STATUS / RUNS TODAYVIEW / SCHEMATIC

Adaptation path

Inputs

  • Longitudinal portrait
  • Live cascade state
  • Learned feedback
  • Relationship context

Behavioral guidance

Can change

  • tone
  • pacing
  • directness
  • warmth
  • challenge
  • lean into / hold back
Four sources combine into behavioral guidance for the response layer. There is no profile card here and no generated reply: what the model produces is how to conduct the next exchange.

Two principles matter here.

Theory is not allowed to win by default

First, theory is not allowed to win by default. Psychological models provide hypotheses, but learned evidence about what actually works for this person can override them.

TheoryLearned feedbackDecision gate

Adaptation is not imitation

Second, adaptation is not imitation. Janus does not simply mirror the user's style or stress response. It can match where similarity helps and deliberately differ where complementarity, stability, or productive friction is more useful.

Match where it helpsDiffer where complementarity helps

The goal is not maximum agreement. It is the right amount of fit and friction for this person, in this context, at this point in the relationship with AI.

04 / Timescales

Different parts of the model change at different speeds

Janus does not allow every observation to rewrite every part of the portrait equally quickly.

Short-lived state can affect today's interpretation immediately. Behavioral patterns require repetition. Deeper hypotheses require more convergence and are subject to lower confidence ceilings.

Environmental influence is also gated by timescale: shallow effects can appear quickly, while deeper change requires accumulated history and intermediate pathways to appear first.

This matters because a longitudinal model should be able to learn without becoming unstable.

STATUS / RUNS TODAYVIEW / SCHEMATIC

Relative update speed / not a numeric scale

  • Short-lived statecan affect today's interpretation immediately
  • Behavioral patternsrequire repetition
  • Deeper hypothesesrequire more convergence / lower confidence ceilings
  • Environmental influenceshallow effects can appear quickly
  • deeper change requires accumulated history + intermediate pathways
Lane length shows relative update speed and nothing else. There is no unit, no duration and no confidence value in this object.

The system is designed to become more specific about a person without becoming more certain — or more changeable — than the evidence allows.

05 / Knowledge

A structured psychological knowledge layer

The psychological architecture is an original integration grounded in established frameworks rather than a single theory.

Product model / not clinical diagnosis

The integration itself is a product model, not a clinical diagnostic system or a validated theory of personality.

Its current knowledge model draws on, among other things:

STATUS / RUNS TODAYVIEW / SCHEMATIC

Established frameworks inform the model

  • Big Five personality research;
  • attachment research;
  • Schwartz's refined theory of values;
  • cognitive-affective models of behavior;
  • appraisal theories of emotion;
  • self-determination research;
  • models of coping, defense, decision style, and behavioral adaptation.

Source-grounded corpus

  1. 57,515structured ideas
  2. 128books
  3. 16functional categories

PostgreSQL + pgvector

longitudinal runtime state + semantic indexes

Neo4j

read-only graph substrate for structured psychological knowledge + cross-level inference

Distinct states

  • Source-grounded
  • Graph present
  • Inferred
  • Runtime-used
Established research informs the model; a source-grounded corpus supplies structured material; runtime state and the knowledge graph are separate substrates. Being stored, being present in the graph, being inferred and being used in a response are four different things.

The current source-grounded corpus contains 57,515 structured ideas from 128 books across 16 functional categories.

Much of that material is not stored as personality labels. It includes rules, context modifiers, etiologies, markers, techniques, interventions, assessments, safeguards, and other material about mechanisms, conditions, and what changes what.

PostgreSQL + pgvector hold the longitudinal runtime state and semantic indexes. Neo4j serves as a read-only graph substrate for structured psychological knowledge and cross-level inference.

The architecture keeps source-grounded knowledge, graph presence, inference, and live runtime use as distinct concepts rather than treating "stored somewhere in the system" as the same thing as "used in every response."

06 / Transfer

What transfers into enterprise Janus

The enterprise version does not need the full original companion product.

It reuses the parts that matter for a professional counterpart:

STATUS / RUNS TODAY

View / transfer schematic

  • persistent longitudinal state;
  • evidence-backed, competing, revisable hypotheses;
  • predictions that later evidence can confirm or contradict;
  • immutable revision history and provenance;
  • separation of facts, short-term state, context, and deeper hypotheses;
  • behavioral adaptation driven by the current model;
  • learned feedback that can override theory;
  • a runtime boundary between deep private inference and visible response;
  • structured knowledge that can inform interpretation without becoming a user-facing label.

Object of the model changes

Broad relational companion

How should this AI show up for this person?

Enterprise Human+AI work

How should this person and this AI divide the thinking next?

The machinery on the left already exists. What changes on the right is what it is pointed at — the same architecture asking a narrower operational question.

What changes is the object of the model.

Instead of optimizing a broad relational companion, the enterprise system narrows the learning target to how a professional works with AI: evidence thresholds, calibration, reliance, recurring blind spots, response to challenge or scaffolding, and how those patterns change across repeated work.

The same architecture that once asked, "How should this AI show up for this person?" can now ask a narrower operational question:

07 / Comparison

The comparison that matters

The richer architecture still has to justify itself.

The useful comparison is:

STATUS / NEXT TEST

Comparison frame

  1. A — Generic AI

    AI assistance with no special interaction design or personalization.

  2. B — Strong universal interaction design

    A good Human+AI protocol, but the same protocol for everyone.

  3. C — Strong shallow personalization

    Adapt from recent performance and operational signals: task class, accuracy, calibration, overrides, role or experience, and reliance history.

  4. D — Janus

    A persistent professional counterpart with richer longitudinal state: reasoning history, competing and revisable hypotheses, interaction-response history, and adaptation across repeated work.

Decisive comparison / D ↔ C

Four ways of running Human+AI work, described on identical terms. The bracket marks which pair actually decides the question.

The comparison that decides whether the richer Janus architecture is worth keeping is D against C.

If D does not materially outperform a strong C, the product should become simpler.

08 / Boundary

NOT DEMONSTRATED

What this has not demonstrated yet

The current system has had no external users and has not demonstrated enterprise outcomes.

It has not yet established that:

  1. NOT DEMONSTRATED

    a rich longitudinal professional model beats strong shallow personalization;

  2. NOT DEMONSTRATED

    useful individual patterns can be learned quickly enough in real professional work;

  3. NOT DEMONSTRATED

    adaptive allocation improves decision quality, review cost, or both;

  4. NOT DEMONSTRATED

    repeated adaptation preserves professional capability over time.

The first three are nearer-term product questions. The last is a longer-term question.

Model gaps / friction

The current system can identify uncertainty and gaps in its own model; deciding when those gaps justify additional friction is part of the enterprise adaptation problem.

Model change / expiry

Model change is another constraint. Some findings may survive an AI model upgrade, while pair-specific findings may expire with the Human+AI relationship that produced them. Janus has to learn quickly enough for its state to remain useful.