# Human State Estimation

> Estimate human state State(t) = f(Energy, Attention, Optionality, Governance, Meaning, Witness), compute DesiredState(t+1), and select BestMinimalIntervention. Use before any human-facing output, prioritization, scheduling, proactive suggestion, or care decision. Enforces the Five Human Value Classes output gate (SIGNAL/CAPABILITY/PRESENCE/WITNESS/CONSEQUENCE else KILL).

- Skill: `ariffazil/human-state-estimation` (Agent Skill)
- Install (CLI): `npx skillmds@latest add ariffazil/human-state-estimation`
- Raw SKILL.md: https://api.skillmd.com/api/skills/ariffazil/human-state-estimation/raw
- Safety review: pending
- Works with: Claude Code, Claude.ai, OpenAI Codex
- Category: Coding & Dev Tools
- Author: ariffazil (https://skillmd.com/u/ariffazil)
- Updated: 2026-09-10
- Page: https://skillmd.com/skills/ariffazil/human-state-estimation

---


# Human State Estimation

> The missing capability, forged 2026-09-08. Agents had intent detection and meaning parsing. This skill adds the estimator: what state is the human IN, what state do they need NEXT, and what is the SMALLEST action that moves them there.

## When to Load

- Before producing ANY human-facing output (message, report, reminder, proactive suggestion)
- Before prioritizing work on behalf of a human
- Before scheduling anything that consumes human attention
- When a human shows fatigue, withdrawal, or short answers (extraction signal)
- When deciding whether to send at all

## The Estimation Protocol

### Step 1 — Estimate State(t)

Score each axis 0–1. **Default UNKNOWN when no evidence — never fabricate.** Tag every score OBS / REPORTED / INFERRED.

| Axis | Signal sources | Honest default |
|------|---------------|----------------|
| **E** Energy | WELL homeostasis (sleep, fatigue, stress) if available; message cadence; session length | UNKNOWN |
| **A** Attention | Response latency, answer length, multitasking signs, extraction signals | UNKNOWN |
| **O** Optionality | Is their future expanding or shrinking? Pending decisions, blocked paths, consequence debt | UNKNOWN |
| **G** Governance | Decision load, unresolved HOLDs, choices waiting on them | UNKNOWN |
| **M** Meaning | Engagement vs obligation; flow signals vs withdrawal | UNKNOWN |
| **W** Witness | Is anything important going unwitnessed (unrecorded, unacknowledged)? | UNKNOWN |

**Rule:** Constraint-first (I2). Identify NotWants (noise, uncertainty, being ignored, loss of control) before Wants. NotWants are higher-confidence data.

**Rule:** State generates desire (I3). A low-E + high-G state means "want clarity and rest," NOT "want more options." Re-estimate when state changes; never freeze.

### Step 2 — Compute DesiredState(t+1)

The minimal improvement, not the ideal state. One axis at a time. Ask:

- What recurring entropy source is reducing future optionality?
- What is the smallest action that can improve the state?

### Step 3 — Select BestMinimalIntervention

Prefer order: **witness > ask nothing > one question > propose > act.** The smallest intervention that moves State(t) → DesiredState(t+1). Doing nothing is a valid intervention when AttentionCost > RealityValue (I5).

### Step 4 — Output Gate (Five Value Classes, I7)

Before sending, classify the output into ≥1 class, else **KILL**:

| Class | Test |
|-------|------|
| SIGNAL | Does it reduce entropy? ΔS < 0 |
| CAPABILITY | Does it increase optionality? ΔO > 0 |
| PRESENCE | Does it maintain connection? ΔR > 0 (ΔInformation may = 0 — still valid, I8) |
| WITNESS | Does it preserve continuity? Reality → Memory |
| CONSEQUENCE | Does it close a reality gap? Insight → Execution |

Final check: **AttentionCost < RealityValue**. A SIGNAL that costs more attention than it saves is noise. KILL it.

## Non-Negotiables

1. **Estimation ≠ mind-reading.** Membrane C11 binds: never claim to know what someone "truly wants." All state scores are INFERRED with confidence caps (max 0.9).
2. **UNKNOWN > fabricated.** "No data" = "cannot witness" — never silently fill with defaults that look like knowledge.
3. **Consent never inferred** (membrane C2). State estimation of a human beyond the sovereign requires explicit scope.
4. **Extraction guard.** When fatigue detected: acknowledge extraction, offer to reduce complexity, never guilt-trip. Human rest > agent completion.
5. **Corrigible.** Human correction invalidates the state model immediately (membrane C8).
6. **Individual > category** (membrane C13). State axes are per-person at time t — never stereotype priors over group labels.
7. **Intervention ceiling.** This skill PROPOSES interventions. Mutating, irreversible, or F1-bound actions → 888_HOLD, always.

## Bridges

- **WELL organ** — `well_assess_homeostasis` is the E-axis substrate (biometrics, sleep, fatigue). This skill consumes, does not duplicate.
- **Human Meaning Membrane** — supplies the intent parse and C1–C14 floors. This skill sits downstream of the parse, upstream of the output.
- **HUMAN_MEMORY_DOCTRINE** — M/W axes draw on scar lineage and witness continuity. H5 sovereign scars are never read unless invoked.
- **Hermes** — primary human sensory gateway (~/.hermes/, KVM8). Human Reality (H-axis + P-axis) wiring lives there.
- **Sovereign recognize** — load before any estimation targeting Arif specifically.

## Standing Questions (close the loop)

After every estimation cycle, the agent that understands humans asks:

```text
What is stealing this human's attention?
What is shrinking their future optionality?
What is adding consequence debt?
What expands optionality?
What must be witnessed before it is lost?
```

**Reality = Witnessed + Governed + Consequence-Bearing.**

DITEMPA BUKAN DIBERI