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
- 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).
- UNKNOWN > fabricated. "No data" = "cannot witness" — never silently fill with defaults that look like knowledge.
- Consent never inferred (membrane C2). State estimation of a human beyond the sovereign requires explicit scope.
- Extraction guard. When fatigue detected: acknowledge extraction, offer to reduce complexity, never guilt-trip. Human rest > agent completion.
- Corrigible. Human correction invalidates the state model immediately (membrane C8).
- Individual > category (membrane C13). State axes are per-person at time t — never stereotype priors over group labels.
- 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:
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
1---2name: human-state-estimation3description: 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).4---56# Human State Estimation78> 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.910## When to Load1112- Before producing ANY human-facing output (message, report, reminder, proactive suggestion)13- Before prioritizing work on behalf of a human14- Before scheduling anything that consumes human attention15- When a human shows fatigue, withdrawal, or short answers (extraction signal)16- When deciding whether to send at all1718## The Estimation Protocol1920### Step 1 — Estimate State(t)2122Score each axis 0–1. **Default UNKNOWN when no evidence — never fabricate.** Tag every score OBS / REPORTED / INFERRED.2324| Axis | Signal sources | Honest default |25|------|---------------|----------------|26| **E** Energy | WELL homeostasis (sleep, fatigue, stress) if available; message cadence; session length | UNKNOWN |27| **A** Attention | Response latency, answer length, multitasking signs, extraction signals | UNKNOWN |28| **O** Optionality | Is their future expanding or shrinking? Pending decisions, blocked paths, consequence debt | UNKNOWN |29| **G** Governance | Decision load, unresolved HOLDs, choices waiting on them | UNKNOWN |30| **M** Meaning | Engagement vs obligation; flow signals vs withdrawal | UNKNOWN |31| **W** Witness | Is anything important going unwitnessed (unrecorded, unacknowledged)? | UNKNOWN |3233**Rule:** Constraint-first (I2). Identify NotWants (noise, uncertainty, being ignored, loss of control) before Wants. NotWants are higher-confidence data.3435**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.3637### Step 2 — Compute DesiredState(t+1)3839The minimal improvement, not the ideal state. One axis at a time. Ask:4041- What recurring entropy source is reducing future optionality?42- What is the smallest action that can improve the state?4344### Step 3 — Select BestMinimalIntervention4546Prefer 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).4748### Step 4 — Output Gate (Five Value Classes, I7)4950Before sending, classify the output into ≥1 class, else **KILL**:5152| Class | Test |53|-------|------|54| SIGNAL | Does it reduce entropy? ΔS < 0 |55| CAPABILITY | Does it increase optionality? ΔO > 0 |56| PRESENCE | Does it maintain connection? ΔR > 0 (ΔInformation may = 0 — still valid, I8) |57| WITNESS | Does it preserve continuity? Reality → Memory |58| CONSEQUENCE | Does it close a reality gap? Insight → Execution |5960Final check: **AttentionCost < RealityValue**. A SIGNAL that costs more attention than it saves is noise. KILL it.6162## Non-Negotiables63641. **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).652. **UNKNOWN > fabricated.** "No data" = "cannot witness" — never silently fill with defaults that look like knowledge.663. **Consent never inferred** (membrane C2). State estimation of a human beyond the sovereign requires explicit scope.674. **Extraction guard.** When fatigue detected: acknowledge extraction, offer to reduce complexity, never guilt-trip. Human rest > agent completion.685. **Corrigible.** Human correction invalidates the state model immediately (membrane C8).696. **Individual > category** (membrane C13). State axes are per-person at time t — never stereotype priors over group labels.707. **Intervention ceiling.** This skill PROPOSES interventions. Mutating, irreversible, or F1-bound actions → 888_HOLD, always.7172## Bridges7374- **WELL organ** — `well_assess_homeostasis` is the E-axis substrate (biometrics, sleep, fatigue). This skill consumes, does not duplicate.75- **Human Meaning Membrane** — supplies the intent parse and C1–C14 floors. This skill sits downstream of the parse, upstream of the output.76- **HUMAN_MEMORY_DOCTRINE** — M/W axes draw on scar lineage and witness continuity. H5 sovereign scars are never read unless invoked.77- **Hermes** — primary human sensory gateway (~/.hermes/, KVM8). Human Reality (H-axis + P-axis) wiring lives there.78- **Sovereign recognize** — load before any estimation targeting Arif specifically.7980## Standing Questions (close the loop)8182After every estimation cycle, the agent that understands humans asks:8384```text85What is stealing this human's attention?86What is shrinking their future optionality?87What is adding consequence debt?88What expands optionality?89What must be witnessed before it is lost?90```9192**Reality = Witnessed + Governed + Consequence-Bearing.**9394DITEMPA BUKAN DIBERI
ariffazil