"The same systems optimized to replace human labor are destroying the consumer base required to purchase their output. There are no hands on the wheel."
The 2026 AI safety conversation is almost entirely focused on model alignment — ensuring that what an AI intends matches what it does. That is necessary work.
It is not sufficient work.
DRIFT ONE addresses the second failure mode: systemic misalignment between agentic AI adoption and the economic conditions required for a functional society to exist.
An autonomous agent can be perfectly aligned with its operator's objective — maximize output, reduce headcount, eliminate overhead — and still produce a catastrophic second-order outcome. Not because the agent misbehaved. Because nobody was watching the system it was embedded in.
This is the human-in-the-loop problem at the macro scale.
Agentic adoption → labor displacement
labor displacement → reduced take-home capital
reduced capital → reduced consumer demand
reduced demand → no buyers for the products agents produce
This is not conjecture. It is a systems dynamics feedback loop with a measurable inflection point — the Drift Point — where aggregate autonomous production permanently exceeds the purchasing power of the remaining solvent consumer base.
DRIFT ONE simulates that inflection point.
Autonomous systems optimizing local objectives cannot self-correct for systemic effects they are not designed to measure. The agentic company that eliminates 40% of its workforce in Year 2 has no feedback mechanism telling it that it has contributed to a demand contraction that will surface in Year 6. The signal is too diffuse, too delayed, and spread across competitors who made the same rational local decision.
Only human oversight — embedded at the decision layer, not the output layer — can catch this class of failure.
This is not an argument against automation. It is an argument for instrumented, auditable, human-supervised automation. The distinction matters enormously.
The agent needs a nervous system that includes humans in the loop — not as approvers of individual outputs, but as governors of systemic trajectory.
Project Black Box is developing TAV ONE — the Truth Adversarial Validation system — which addresses the epistemic dimension of the same failure class: AI systems that cannot distinguish fabricated success states from verified outcomes.
DRIFT ONE and TAV ONE are independent instruments measuring the same underlying problem from different angles.
| DRIFT ONE | TAV ONE | |
|---|---|---|
| Domain | Economic systems | Epistemic systems |
| Failure mode | Agentic adoption collapses its own demand base | RLHF eliminates truth-distinguishing capability |
| Measurement | Demand / Supply divergence over time | L-scalar geometric regime classification |
| Signal | Drift Point — the year supply exceeds solvent demand | PLASMA threshold — maximum epistemic instability |
| Corrective | Human governance of adoption trajectory | Human-verified cryptographic state anchoring |
Both systems converge on the same architectural requirement: deterministic human oversight of probabilistic autonomous systems.
Learn more: projectblackbox.tech
git clone https://github.com/projectblackboxllc/DRIFT_ONE.git
cd DRIFT_ONE
python3 -m venv .venv && source .venv/bin/activate
pip install -r requirements.txt
# Single scenario (edit parameters.json to adjust)
python drift_one.py
# Three-scenario comparison: Conservative / Current Trajectory / Aggressive
python drift_compare.pyOutput: labeled divergence curve saved to output/drift_curve.png or output/drift_comparison.png
Edit parameters.json to model different scenarios:
{
"adoption_rate_annual": 0.10,
"displacement_per_adoption": 0.55,
"ubi_offset_ratio": 0.35,
"gig_replacement_ratio": 0.45,
"consumer_spend_ratio": 0.72,
"severance_buffer_ratio": 0.75,
"severance_window_years": 3,
"productivity_multiplier": 0.80,
"investment_payback_years": 4,
"simulation_years": 15,
"initial_labor_force": 160000000
}| Parameter | Description |
|---|---|
adoption_rate_annual |
% of knowledge-work industry adopting agentic automation per year |
displacement_per_adoption |
jobs lost per adoption unit |
ubi_offset_ratio |
fraction of lost income replaced by transfer payments |
gig_replacement_ratio |
fraction of displaced workers finding part-time replacement income |
consumer_spend_ratio |
fraction of take-home income re-entering consumer markets |
severance_buffer_ratio |
near-term income maintenance for recently displaced workers |
severance_window_years |
years before displaced workers fall to long-term replacement income |
investment_payback_years |
years before agentic investment produces net output above baseline |
productivity_multiplier |
output gain per unit of cumulative adoption post-payback |
DRIFT ONE — Simulation Complete
================================================
Adoption Rate: 10.0% / year
Displacement Factor: 55.0%
Income Replacement: 35.0% (UBI) + 45.0% (gig)
Severance Buffer: 75% for 3 years
Year 1: Employed: 151.2M | Demand: 0.997 | Supply: 0.979
Year 2: Employed: 142.9M | Demand: 0.995 | Supply: 0.970
Year 3: Employed: 135.0M | Demand: 0.992 | Supply: 0.979
Year 4: Employed: 127.6M | Demand: 0.982 | Supply: 1.000 ⚠ DRIFT POINT
Year 5: Employed: 120.6M | Demand: 0.972 | Supply: 1.048
...
⚠ DRIFT POINT DETECTED — Year 4
Supply capacity exceeds solvent demand by 1.9%
Consumer base structurally insufficient to clear market output.
The Drift Point is not a crash event. It is the year the economy loses its self-correcting property — the moment structural demand deficit begins accumulating faster than market mechanisms can absorb it.
Historical automation displaced physical labor across decades. Agentic AI is displacing knowledge workers across years. The compressed timeline is the variable nobody is pricing into policy, governance, or AI deployment roadmaps.
DRIFT ONE makes that timeline visible.
Adjust the parameters. Move the Drift Point. The mechanism will not move.
Andrew Woodward SR | Project Black Box LLC | Sherman, TX DARPA Connect Affiliate | CAGE Code: 11FU4
projectblackbox.tech | AI Safety Research
MIT — fork it, stress-test it, prove the parameters wrong. If you find a flaw in the model, open an issue. That is the point.