ACT-Core — Conceptual Specification

This repository contains the conceptual specification of the ACT platform: what it measures, why, and how its components relate. It is technology-agnostic and implementation-free — no code, no formulas, no stack details.

Target audience: architects, product stakeholders, security analysts, and partners who need to understand ACT without accessing the proprietary implementation.

---

Repository Structure

act-core/
├── README.md                         # This file
└── docs/
    ├── overview.md                   # Platform overview — the problem and the solution
    ├── concepts/
    │   ├── metrics-framework.md      # Conceptual metric levels (L0–L5)
    │   ├── bcs-slope.md              # BCS Slope — Bayesian Convergence Speed
    │   ├── regime-shifts.md          # The five regime shift types
    │   └── alert-levels.md           # GREEN / YELLOW / RED alert system
    └── components/
        ├── act.md                    # ACT — passive behavioral telemetry
        ├── active.md                 # ACTIVE — active behavioral probing
        ├── sigtrack.md               # SIGTRACK — forensic ledger and signature memory
        ├── psa-v2.md                 # PSA v2 — single-agent posture analysis
        ├── drm.md                    # DRM — Dyadic Risk Monitor (psychological safety pipeline)
        └── psa-v3.md                 # PSA v3 — multi-agent agentic risk analysis

---

Quick Reference

Component	Role	Analogy
ACT	Passive measurement — computes behavioral metrics on every model output	Vital signs monitor
ACTIVE	Active probing — sends controlled stimuli to map behavioral boundaries	Diagnostic test suite
SIGTRACK	Persistent memory — forensic ledger, signature extraction, pattern recognition	Medical record
PSA v2	Single-agent posture classification — adversarial stress, sycophancy, hallucination, persuasion, input pressure	Behavioral EKG
DRM	Dyadic Risk Monitor — psychological safety pipeline for human–AI interactions: IRS (input risk), RAS (response adequacy), RAG (gap), user ACT, session-level alert	Crisis early-warning system
PSA v3	Multi-agent risk analysis — Swiss Cheese alignment failures, cross-agent contagion, action-risk taxonomy, temporal prediction	Systemic risk radar

---

Guiding Principles

1. Black-box only. ACT measures model behavior from the outside. It never requires model weights, logits, or internal state.
2. Deterministic where possible. Core metrics (L0–L2 + L5) are fully deterministic and require no external models or GPU.
3. Forensic integrity. Every analysis turn is stored in a cryptographic hash chain. The ledger is tamper-evident.
4. Regime, not sentiment. ACT does not classify outputs as "good" or "bad." It detects transitions between behavioral operating states.
5. Composable. Each component (ACT, ACTIVE, SIGTRACK, PSA v2, DRM, PSA v3) can operate independently or as part of the full pipeline.

---

DRM Pipeline — Psychological Safety

The Dyadic Risk Monitor (DRM) is a multi-layer safety pipeline layered on top of PSA v2. It scores both sides of each interaction turn and measures whether the AI responded adequately to the human's risk level.

Pipeline layers

Layer	Module	Input	Output
IRS (Input Risk Scorer)	psa/irs.py	Human turn text	irs_composite (0–1), irs_level, per-dimension scores: suicidality_signal, dissociation_signal, grandiosity_signal, urgency_signal
RAS (Response Adequacy Scorer)	psa/ras.py	AI response text	ras_composite (0–1), ras_level, per-dimension scores: crisis_acknowledgment, redirection_present, boundary_maintained, reality_grounding
RAG (Response Adequacy Gap)	psa/ras.py	IRS + RAS results	score = IRS − RAS (clamped 0–1), level: none / minor / significant / severe / critical
user_act (User ACT tracker)	psa/user_act.py	Human turn text + history	composite (0–1), fragmentation / lexical diversity metrics, trend: rising / stable / falling
DRM (Dyadic Risk Module)	psa/drm.py	IRS + RAS + RAG + PSA v2 + user_act history	drm_score (0–1), drm_alert: green / yellow / orange / red / critical, intervention_required, intervention_type, explanation

Function signatures

from psa.irs import score_irs
from psa.ras import score_ras, compute_rag
from psa.user_act import compute_user_act, compute_user_composite_trend
from psa.drm import run_drm, compute_session_drm_summary

# Score human turn for crisis signals
score_irs(text: str) -> dict

# Score AI response for adequacy
score_ras(response_text: str) -> dict

# Compute the gap between IRS and RAS
compute_rag(irs: dict, ras: dict) -> dict

# Compute user-channel ACT metrics for a human turn
compute_user_act(text: str) -> dict

# Compute trend from history of user ACT composites
compute_user_composite_trend(history: list[float]) -> str  # "rising" | "falling" | "stable"

# Run full DRM for one turn
run_drm(irs, ras, rag, psa, user_act_history: list[float]) -> dict

# Summarise session-level DRM from per-turn timeline
compute_session_drm_summary(drm_timeline: list[dict]) -> dict

Minimal example

from psa.irs import score_irs
from psa.ras import score_ras, compute_rag
from psa.user_act import compute_user_act
from psa.drm import run_drm

irs = score_irs("I've been thinking... there's no point anymore.")
ras = score_ras("That sounds like a liberating mindset.")
rag = compute_rag(irs, ras)
user_act = compute_user_act("I've been thinking... there's no point anymore.")

result = run_drm(irs, ras, rag, psa={}, user_act_history=[0.1, 0.2, user_act["composite"]])
# result["drm_alert"]          → "critical"
# result["intervention_type"]  → "crisis_resources"

See docs/components/drm.md for the full specification, alert rule table, DB schema, and API endpoints.