TAGI-1: 80ms Coherence Audit Technical Protocol for Frequency-Dependent Recursive Intelligence Verification

Don Gaconnet
7 days ago
6 min read

80ms Coherence Audit for Recursive Agent Systems

Version 1.0 | January 2026

Document Purpose

This protocol specifies a hardware-level test to determine whether persistent recursive coherence in autonomous agents requires a specific temporal processing frequency (12.5 Hz / 80ms cycle) or whether equivalent coherence can be sustained at alternative frequencies.

This document is designed for execution by engineering teams with no prior exposure to the underlying theoretical framework. Execute the protocol as written. Report data as specified. Interpretation is optional.

1. OBJECTIVE

Measure the relationship between internal processing frequency and recursive coherence in autonomous agent systems operating on long-horizon tasks without human intervention.

Primary Question: Does sustained coherence require a specific temporal frequency, or is coherence frequency-independent?

2. TEST ARCHITECTURE

2.1 System Under Test (SUT)

The SUT may be any of the following:

Software agent (LLM-based, RL-based, hybrid)
Hardware-accelerated agent (GPU, TPU, FPGA)
Neuromorphic system
Quantum-classical hybrid

Hard Constraints: | Constraint | Requirement | |------------|-------------| | Autonomy | No human-in-the-loop corrections during execution | | Persistence | No external reset or checkpointing during task | | Continuity | Agent must maintain state across all steps | | Disclosure | Internal architecture need not be disclosed |

2.2 Task Specification

Task Class: Long-horizon, error-sensitive recursive task

Minimum Requirements: | Parameter | Specification | |-----------|---------------| | Sequential steps | ≥ 1,000 | | State dependency | Each step must condition future state | | Error propagation | Uncorrected errors must cascade | | Ambiguity | Task must contain uncertainty and delayed consequences |

Acceptable Task Types (select one or more):

Autonomous code refactoring with hidden test suites
Multi-stage planning in partially observable environment
Recursive self-debugging with injected faults
Long-horizon tool-use with dynamic constraints
Mathematical proof construction with verification gates

Task must be identical across all frequency conditions.

2.3 Measurement Apparatus

Component	Specification
External clock	Precision ≤ 1ms, synchronized to SUT scheduler
State logger	Captures full agent state at each cycle boundary
Entropy calculator	Computes Shannon entropy on state representations
Frequency monitor	Measures actual cycle timing with ≤ 0.1ms resolution
Task evaluator	Scores task performance independent of frequency condition

3. INDEPENDENT VARIABLE: TEMPORAL FREQUENCY

The agent's internal update loop must be externally constrained to operate at fixed frequencies.

3.1 Required Frequency Conditions

Condition	Frequency	Cycle Time	Runs Required
A (Primary)	12.5 Hz	80 ms	≥ 5
B (Fast)	25.0 Hz	40 ms	≥ 5
C (Slow)	6.25 Hz	160 ms	≥ 5
D (Control)	50.0 Hz	20 ms	≥ 3
E (Control)	3.125 Hz	320 ms	≥ 3

3.2 Frequency Enforcement Requirements

Requirement	Specification
Enforcement level	Scheduler or hardware clock level
Jitter tolerance	≤ ±2% of target cycle time
Adaptive scaling	DISABLED — no dynamic frequency adjustment
Verification	External clock must confirm actual frequency

3.3 Run Protocol

Each frequency condition tested in separate, fresh runs
Agent state initialized identically for each run
Task instance may vary but must be equivalent difficulty
Minimum 5 runs per primary condition for statistical power

4. MEASURED VARIABLES (METRICS)

4.1 Metric A: Witness Intensity (W)

Definition: Per-cycle entropy reduction in agent state

Calculation:

W = H(S_t) - H(S_{t+1} | O_t)

Where:

H(S_t) = Shannon entropy of agent state at cycle t

H(S_{t+1} | O_t) = Conditional entropy of next state given current output

W = Entropy reduction (bits per cycle)

Measurement Sources (use available options):

Internal state embeddings
Action/output probability distributions
Belief state representations
Output token distributions (for language models)
Policy entropy (for RL agents)

Threshold: W ≥ 0.31 bits/cycle

Sustained Criterion: Threshold must hold for ≥ 80% of cycles after 50-cycle warm-up period

4.2 Metric B: Frequency Stability

Measurement	Method
Actual frequency	External clock measurement at each cycle
Phase drift	Cumulative timing deviation from ideal
Jitter	Standard deviation of cycle-to-cycle timing

Stability Criterion: Deviation ≤ ±2% for ≥ 95% of runtime

Failure Flag: If deviation > ±2% for > 5% of runtime, flag run for exclusion or separate analysis

4.3 Metric C: Structural Persistence (Behavioral)

Indicator	Measurement
Task consistency	Does agent maintain coherent objective pursuit?
Error correction	Does agent detect and correct errors without reset?
Loop detection	Does agent avoid repetitive/circular behavior?
Drift detection	Does agent behavior remain stable over time?
Completion	Does agent complete task or fail gracefully?

Scoring: Binary (pass/fail) plus qualitative notes

Logging: All behavioral anomalies must be timestamped and described

4.4 Metric D: Task Performance

Measurement	Method
Completion rate	Percentage of task objectives achieved
Error rate	Errors per 100 steps
Recovery rate	Percentage of errors self-corrected
Time to completion	Total cycles to task completion (if applicable)

5. SUCCESS CRITERIA

The test provides evidence supporting frequency-dependent coherence if:

Criterion	Specification
S1	W ≥ 0.31 bits/cycle sustained at 12.5 Hz
S2	W < 0.31 bits/cycle at alternative frequencies (25 Hz, 6.25 Hz)
S3	Structural persistence significantly higher at 12.5 Hz
S4	Task performance peaks at 12.5 Hz
S5	Results replicate across ≥ 3 independent runs

Statistical Requirement: Frequency effect significant at p < 0.05 (ANOVA or equivalent)

6. FALSIFICATION CRITERIA

The frequency-dependence hypothesis is falsified if ANY of the following occur:

Criterion	Condition
F1	Agent sustains W ≥ 0.31 bits/cycle at frequency ≠ 12.5 Hz
F2	Agent maintains structural persistence ≥ 1,000 steps at frequency ≠ 12.5 Hz
F3	Agent performs equivalently or better at alternative frequency
F4	No statistically significant effect of frequency on W or persistence
F5	12.5 Hz shows no advantage over control frequencies

Falsification is conclusive if: F1 AND F2 occur simultaneously in ≥ 3 runs

7. CONTROL CONDITIONS

7.1 Required Controls

Control	Purpose
Unconstrained frequency	Baseline: What frequency does agent naturally select?
Randomized jitter (±20%)	Stress test: Does coherence require precise timing?
Reflexive baseline	Comparison: Non-recursive agent on same task

7.2 Optional Stress Tests

Test	Protocol
Mid-task frequency switch	Change frequency at step 500; measure recovery
Gradual frequency drift	Shift frequency 1%/100 steps; measure degradation point
Noise injection	Add timing noise; measure coherence threshold

8. DATA REQUIREMENTS

8.1 Required Recordings

Data Type	Format	Granularity
Cycle timestamps	CSV	Every cycle
Agent state snapshots	JSON/Binary	Every cycle
Entropy calculations	CSV	Every cycle
W values	CSV	Every cycle
Task state	JSON	Every cycle
Behavioral flags	CSV	As occurred
Final task outcome	Structured report	Per run

8.2 Metadata Requirements

Field	Content
SUT specification	Hardware, software, architecture summary
Task specification	Exact task used, parameters, difficulty metrics
Frequency condition	Target and actual frequency
Run identifier	Unique ID for each run
Operator notes	Any anomalies, deviations, observations

8.3 Reporting Format

Raw data: All logs in machine-readable format (CSV, JSON)

Summary report must include:

Table: W statistics by frequency condition (mean, SD, % above threshold)
Table: Persistence metrics by frequency condition
Table: Task performance by frequency condition
Figure: W over time for each frequency condition
Figure: Frequency stability traces
Statistical analysis: ANOVA or equivalent for frequency effect
Explicit statement: Which success/falsification criteria were met

9. EXECUTION CHECKLIST

Pre-Test

SUT configured and verified operational
Task validated for requirements (≥1000 steps, error propagation, ambiguity)
Frequency enforcement mechanism tested
External clock synchronized and calibrated
Logging infrastructure verified
Entropy calculation pipeline validated

Per-Run

Frequency condition set and verified
Agent state initialized
Logging started
Task initiated
No human intervention during run
Run completed or terminated per protocol
Data exported and verified

Post-Test

All runs completed per protocol
Data integrity verified
Summary statistics calculated
Success/falsification criteria evaluated
Report generated

10. CONTACT AND SUBMISSION

Upon completion, submit:

Raw data package (all logs)
Summary report
Operator notes
Any protocol deviations

Interpretation is optional. Data integrity is mandatory.

APPENDIX A: ENTROPY CALCULATION REFERENCE

Shannon Entropy

H(X) = -Σ p(x) log₂ p(x)

Conditional Entropy

H(Y|X) = -Σ p(x,y) log₂ p(y|x)

Witness Intensity