Live System  ·  H3.0 Active

OiMy Architecture
Human 3.0 Update

How a family companion now understands not just what you're struggling with, but how you engage with change — and surfaces the gap between intention and action.

v8.4 post-H3.0
NPS baseline 8.4 / 10
Companion Grok 4.3
4 files modified
89 skills active
38 family containers

What Changed

Four targeted changes. All additive — null profile fields fall through to existing logic unchanged.

New File
archetypes/h30.py
360 lines. H30Engine class with 4 SQLite tables. Quadrant + level inference via Grok 4.3 every 10 messages. Commitment extraction, false transform detection, 14-day probe rate limiter.
Modified
archetypes/inference.py
H30Engine wired into InferenceEngine.__init__. Guarded import — if H3.0 fails, self.h30 = None and everything else is unchanged.
Modified
api.py
tick() called on chat_light AND coach paths. [H3.0] context tier injected into _build_profile_context. False transform probe wired into both _call_openrouter_companion and _call_deepseek_v3.
Modified
intent_router.py
Quadrant tiebreaker added after routing decision. When confidence < 0.70 and quadrant is set, preferred scaffold is logged as metadata. Additive only — never overrides unambiguous routes.
Bug Caught During Integration Testing
The original H3.0 tick() was wired only into coach/expert/planner workflows. The chat_light path — which handles ~80% of all conversations — returned early before the tick fired, meaning H3.0 would have collected zero data in production. Fixed before any sessions ran: tick() now fires in the chat_light path before the early return. Verified via direct DB inspection (msg_count increments correctly).

Architecture Flow

Every message path — with H3.0 steps highlighted in green

📱
Telegram / WhatsApp
Inbound message · JWT auth
🛡
Pre-Filter
Deterministic rules · <1ms · No LLM
⚡
Intent Router
gpt-5.4-mini · ~180ms · scaffold classification
📡
Signal Collector
Stress markers · barriers · urgency cues
🧬
H3.0 Engine
tick() · commitment extraction · false transform check · Grok 4.3 inference every 10 msgs
New
🧠
Deep Profile Pipeline
Maslow · motivator · social needs · cadence-gated
📋
Profile Context Builder
[Family] + [Psych] + [H3.0] tiers assembled into system prompt
Modified
🪞
False Transform Probe
2+ signals on topic + cooldown clear → gentle probe injected into system prompt
New
🎯
BCT Engine
Technique selection now gated by h30_level when set · null = prior logic unchanged
Modified
💬
Companion Model
grok-4.3 · full context · fallback: haiku-4-5
🧵
Memory Write-Back
Honcho · Vector · EntityExtractor (async)
✅
Response Delivered
Governance checked · BCT-gated tone · probe woven in naturally

System Layers

Eight distinct layers — each purpose-built. H3.0 adds two new ones and modifies two existing ones.

🛡
Pre-Filter
Zero LLM · deterministic
Zero LLM
  • Latency<1ms
  • ChecksEmergency keywords, injection guard, onboarding gate
⚡
Intent Router
Scaffold classification
gpt-5.4-mini
  • Latency~180ms
  • Outputskill_id + workflow + confidence score
  • H3.0Quadrant tiebreaker when confidence < 0.70
🧬
H3.0 Engine New
Human 3.0 framework
archetypes/h30.py
  • InferenceGrok 4.3 every 10 messages
  • Quadrantmind / body / spirit / vocation
  • Levelconformist / individualist / synthesist
  • False TXCommitment extraction + say/do gap detection
  • CooldownProbe fires max once per topic / 14 days
Dan Koe / HUMAN 3.0 4 SQLite tables Null-safe
🧠
Deep Profile Pipeline
Psychological inference
Cadence-gated
  • MaslowHierarchy level classifier
  • MotivatorDecision motivator classifier
  • SocialSocial needs classifier
  • DeepOpus 4.8 · 11-layer inferences · once per user
🎯
BCT Engine Modified
Behavior Change Techniques
H3.0-gated

When h30_level is set, technique scoring is adjusted for the user's engagement style. Null level = prior logic unchanged.

  • ConformistBoost: action plans, implementation intentions
  • IndividualistBoost: values clarification, identity-based goals. Suppress: prescriptive directives
  • SynthesistBoost: systems mapping, meta-level reflection
💬
Companion Model
Primary conversation engine
grok-4.3
  • RouteOpenRouter · x-ai/grok-4.3
  • Fallbackanthropic/claude-haiku-4-5
  • Context[Family] + [Psych] + [H3.0] + BCT hint + probe
  • Latency2–8s typical

Illustrative Example

One message traced through every layer — including the new H3.0 steps

"My husband and I keep fighting about screen time every night. I told myself I'd set proper limits last week but I never actually did it."
user_id: family_sarah_4721  ·  channel: telegram  ·  message #8 in current session
  • 1
    0ms
    Pre-Filter & Auth
    JWT validated. No emergency keywords detected. Not in onboarding state. Passes through immediately.
  • 2
    ~180ms
    Intent Router gpt-5.4-mini
    Classified as screen-time scaffold. High confidence. Routes to chat_light workflow.
    Outputskill_id: "screen-time"  |  workflow: "chat_light"  |  confidence: 0.91
  • 3
    parallel
    Signal Collector
    Relationship stress detected. Barrier type: inconsistency. Topics tagged: screen_time, marriage, parenting.
  • 4
    sync · every message
    H3.0 Engine — tick() New
    tick() fires. Message #8 registered. Commitment language detected ("I told myself I'd"). Checks prior commitments on screen_time topic — finds one from 6 days ago that was never resolved. False transform signal #2 logged. Probe now eligible.
    commitment_extracted: "set proper screen time limits" topic: "screen_time" // prior commitment found 6 days ago — no progress mention since false_transform_signal: logged (#2 on this topic) probe_eligible: true // quadrant + level from prior inference at msg #0 quadrant: "spirit"  ·  level: "individualist" msg_count: 8  ·  next_inference: msg #10
  • 5
    cadence-gated
    Deep Profile Pipeline
    Existing profile loaded from SQLite. Maslow: love and belonging. Decision motivator: avoid conflict. Social need: partnership.
  • 6
    context assembly
    Profile Context Builder Modified
    Three tiers assembled. The new [H3.0] tier surfaces engagement level, primary life domain, and a BCT coaching hint directly into the system prompt.
    [Psych] Maslow: love/belonging  ·  Motivation: avoid_conflict [H3.0] Engagement: individualist  ·  Domain: spirit BCT hint: Offer frameworks, not directives. Avoid "you should/must". Identity-based framing.
  • 7
    probe check
    False Transform Probe New
    2 signals on screen_time. Never probed before. 14-day cooldown clear. Probe fires — injected as the opening of the system prompt instruction. Will be suppressed for next 14 days on this topic.
    Probe injected into system prompt
    "We've talked about screen time a few times — I'm curious how it's actually been going day to day, not just intentions?"
  • 8
    BCT selection
    BCT Engine Modified
    Level is individualist. Boosts: values clarification, identity reframing, autonomous goal setting. Suppresses: prescriptive action plans, authority directives. Selected: identity_reframing.
  • 9
    2–6s
    Companion Model grok-4.3
    Full system prompt: scaffold + [Psych] + [H3.0] + BCT hint + false transform probe. Model generates response shaped by all three layers simultaneously.
  • 10
    delivered
    Response
    Probe woven in as natural opening. BCT framing: identity-based, no "you should". Tone: collaborative, not directive.
    "We've talked about screen time a few times — I'm curious how it's actually been going day to day, not just the intentions.

    The nightly fight usually isn't about screens. It's about you and your husband not holding the same line. What does a rule that actually works for both of you look like — not the ideal version, the one you'd both actually follow?"
    Why this works: probe surfaces the gap without lecturing. Identity reframing shifts ownership from rules to a shared agreement. No "you should" anywhere. Autonomy preserved for an individualist user.

Model Stack

Every model in the system — what it does, how fast, and why

Model Role Latency API Cost
gpt-5.4-mini Intent router + entity extraction ~180ms OpenRouter Low
grok-4.3 Companion Primary companion + coach — full context, BCT-gated, H3.0-aware 2–8s OpenRouter Medium
claude-haiku-4-5 Companion fallback when circuit breaker opens 1–4s OpenRouter Low
grok-4.3 H3.0 Quadrant + level inference — fires every 10 messages per user 3–8s OpenRouter Low (cadence)
claude-opus-4.8 Deep inference (11-layer profile, once per user) + NPS eval user simulator 60–120s OpenRouter High (once)
kimi-k2.6 Tall model: complex coaching, BCT, medical, multi-constraint planning 15–30s OpenRouter Medium
deepseek-chat DeepSeek V3 — alternate companion path (direct API) 1–4s api.deepseek.com Very Low
gpt-5.5 QA evaluator for NPS eval suite (scores per-turn on 5 dimensions) 3–10s OpenAI Medium (eval only)

NPS Progression

From early beta to post-H3.0 — each step tied to a specific architecture change

v3–v4
2–3
v5–v6
4–5
v7 (Kimi)
6.0
Jun 1 2026
8.4
H3.0 eval
Running…
Jun 1 jump: scaffold enrichment with Becky Kennedy, Gottman, Huberman expert knowledge. H3.0 eval running now — 30 synthetic users across Opus 4.8, Grok 4.3, GLM 5.2 evaluators.