Skip to content

RCT-7 Thinking Protocol

RCT-7 Thinking is the cognitive processing kernel at the heart of the RCT Platform.
It is a 7-step structured reasoning protocol that transforms raw natural language input
into a verified, intent-preserved, FDIA-scored execution plan.

Overview

Attribute Value
Full name RCT-7 Thinking Protocol
Kernel coverage T1 – T9
End-to-end accuracy 96.0%
FDIA gate position Step 7 (Compare) — score ≥ threshold required before execution
Variants RCT-O · RCT-S · RCT-I
Source rctlabs.co/technology/rct-7-thinking

The 7 Steps

RCT-7 maps naturally onto the JITNA kernel tiers (T1–T9):

T1──T2 ──────── Observe
T2──T4 ──────── Analyze
T3──T5 ──────── Deconstruct
T4──T6 ──────── Reverse Reasoning
T5──T6 ──────── Identify Core Intent
T6──T8 ──────── Reconstruct
T7──T9 ──────── Compare with Intent

Step 1 · Observe (T1–T2)

What: Capture the raw input exactly as received — no interpretation yet.

Purpose: Establish an immutable ground truth of "what was said / requested."

Example: 

Input: "Allocate 200 resources to agent_B for defense"
Observation: {raw_text, timestamp, sender_id, channel}

Step 2 · Analyze (T2–T4)

What: Parse the observation into structured entities and constraints.

Purpose: Surface implicit assumptions, ambiguous pronouns, and missing parameters.

Example: 

Parsed: {
  action:    "allocate",
  amount:    200,
  target:    "agent_B",
  goal:      "defense",
  implicit:  ["requires budget_ok", "requires permission_grant"]
}

Step 3 · Deconstruct (T3–T5)

What: Break the analyzed intent into atomic sub-tasks — the smallest executable units.

Purpose: Expose hidden dependencies and parallelization opportunities.

Example: 

Sub-tasks:
  1. check_budget(amount=200)
  2. verify_permissions(target=agent_B)
  3. route_resources(resource_pool → agent_B)
  4. log_audit(action, actor, timestamp)

Step 4 · Reverse Reasoning (T4–T6)

What: Work backward from the desired outcome to validate that the sub-tasks actually produce it.

Purpose: Detect logical gaps before execution (fail-fast guarantee).

Example: 

defense_achieved
  ← resources_allocated_to_B
    ← budget_ok
      ← approval_granted
        ← FDIA_score ≥ threshold   ← CRITICAL GATE

If the backward chain breaks at any node, the plan is rejected before a single byte is sent.

Step 5 · Identify Core Intent (T5–T6)

What: Distil the single primary intent type from all parsed signals.

Purpose: Resolve conflicts when multiple intent signals are present (e.g., PROTECT vs ACCUMULATE).

Intent types (NPCIntentType):

Intent Desire Weight Typical trigger
PROTECT 0.75 Defensive allocation, firewall rules
ACCUMULATE 0.80 Resource gathering, caching
BELONG 0.65 Group join, coalition formation
DISCOVER 0.55 Exploration, knowledge queries
DOMINATE 0.90 Competitive override, force actions
NEUTRAL 0.30 Idle / no-preference tasks

Example: 

Signals detected: [PROTECT(0.92), ACCUMULATE(0.41)]
Core intent resolved: PROTECT  (highest confidence wins)

Step 6 · Reconstruct (T6–T8)

What: Build the final execution plan from verified sub-tasks, ordered and signed.

Purpose: Produce a deterministic, replayable action sequence (JITNA packet) with Ed25519 attestation.

Example: 

execution_plan:
  - step: validate_input
  - step: fdia_check          # score must be ≥ 0.6
  - step: allocate_resources
  - step: ed25519_sign
  - step: notify_stakeholders
  - step: write_audit_log

Step 7 · Compare with Intent (T7–T9)

What: Run the reconstructed plan through the FDIA Scorer and compare the result against the original intent.

Purpose: Ensure the plan hasn't drifted from what was actually requested (constitutional alignment gate).

FDIA gate rule: 

score = FDIAScorer.score_action(agent_intent, plan_action, ...)
if score < THRESHOLD (default 0.60):
    REJECT — intent not preserved
else:
    APPROVE — sign and execute

Example result: 

FDIA score: 0.847  ≥  threshold 0.600  ✅
Decision: EXECUTE — plan signed with Ed25519

Three Protocol Variants

RCT-O — Original (7 Steps)

Full protocol as described above. Used for: - New/unfamiliar tasks - High-stakes decisions (Tier-6, Tier-8) - Any action touching financial, medical, or legal domains

RCT-S — Condensed (4 Steps)

Skips the Deconstruct and Reverse Reasoning phases for low-risk tasks:

Observe → Analyze → Identify Core Intent → Compare with Intent

Used for: - Repeated/cached pattern tasks - Tier-S and Tier-4 decisions - High-throughput pipelines where T3–T5 are pre-validated

RCT-I — Interpretive

Adds a meta-reasoning layer on top of RCT-O for ambiguous or contradictory inputs:

[Pre-step: Resolve Ambiguity]
  → RCT-O (Steps 1–7)
[Post-step: Uncertainty Quantification]

Used for: - Multi-lingual inputs with cultural context gaps - Contradictory stakeholder instructions - High-ambiguity enterprise workflows

FDIA Integration

The FDIA equation anchors the protocol:

\[F = D^I \times A\]
Symbol Meaning in RCT-7
F Final execution score (must be ≥ threshold)
D Data quality of the reconstructed plan
I Intent precision — acts as exponent (higher = amplifies output)
A Alignment gate — human approval factor (0 = hard block)

A = 0 is a constitutional veto. No plan executes without the Architect approval gate.

from core.fdia.fdia import FDIAScorer, NPCAction, NPCIntentType  # from concepts/fdia.md

scorer = FDIAScorer()
score = scorer.score_action(
    agent_intent=NPCIntentType.PROTECT,
    action=NPCAction(action_id="plan_v1", action_type="cooperate", amount=200.0),
    world_resources={"res_alpha": 500.0},
    agent_reputation=0.85,
    other_intents=[NPCIntentType.PROTECT],
)
# score in [0.0, 1.0] — deterministic, no side effects
print(f"FDIA score: {score:.4f}")

Kernel Tier Mapping

T1  — Observation kernel      (raw capture)
T2  — Language kernel         (parsing, NLP)
T3  — Decomposition kernel    (sub-task graph)
T4  — Inference kernel        (constraint propagation)
T5  — Intent kernel           (NPCIntentType resolution)
T6  — Planning kernel         (JITNA packet construction)
T7  — Evaluation kernel       (FDIA scoring)
T8  — Signing kernel          (Ed25519 attestation)
T9  — Execution kernel        (dispatch + audit)

Performance Characteristics

Metric Value
End-to-end accuracy (RCT-O) 96.0%
Cold start latency 3–5 s
Warm recall latency < 50 ms
FDIA threshold (default) 0.60
Intent type coverage 6 types (NPCIntentType)
Kernel tier coverage T1–T9 (9 tiers)

See Also