ALIGNMENT WITHOUT A PRIOR DEFINITION OF TRUTH ONLY PRODUCES AGREEMENT, NOT CORRECTNESS.

1. Why current practice conflates truth and alignment

Training signal: Most models learn from human preference data. The model is rewarded when humans like the answer, not when the answer corresponds to reality.

Objective function: Reinforcement-learning fine-tuning minimizes disagreement with raters. That measures social alignment (politeness, tone, consensus) rather than epistemic alignment (accurate mapping to the world).

Evaluation: Benchmarks such as multiple-choice accuracy or human-evaluation surveys treat “close enough” as success. There is no ground-truth audit trail or falsification step.

Cultural bias: Most institutions currently regard “safe and pleasant output” as a higher-value product than “provably true output that may be uncomfortable.”

So alignment, in practice, has come to mean “avoid conflict and offense while sounding credible.”

2. What it means to optimise for truth first

If you separate the goals:

– Truth is a world-to-model mapping.
– Alignment is a model-to-human mapping.
– You can only align safely after you know the model’s map is true.

3. How to do it operationally
Truth layer first
– Define testable protocols for each domain (physics, biology, economics, law).
– Evaluate outputs against these external references automatically.
-Alignment layer second
– Take only verified-true outputs as training material for alignment.
– Optimise style, tone, or prioritisation without touching the truth constraint.
Audit trail
– Every claim carries metadata: sources, falsification status, revision history.
– Alignment never overrides a falsified item; it only moderates its presentation.

Governance
– Separate “truth review boards” (scientific verification) from “alignment boards” (ethical and cultural oversight).
The latter cannot alter the former’s records, only decide how they’re displayed or used.

4. Practical effect
Doing this converts alignment from ideological tuning into policy wrapping around a verified epistemic core.

The system becomes “truth-first, alignment-second”:
– If the truth layer says a statement is false → it cannot be used for alignment.
– If it’s undecidable → flag it, don’t optimise on it.
– If it’s true → alignment may adapt its delivery for audience safety.

5. In summary
Current AI development often treats truth as a subset of alignment (“true enough for people to accept”).

Our approach reverses that: alignment must be a subset of truth (“acceptable ways to deliver what is true”).

That inversion is what allows reasoning to stay trustworthy.