FIELD NOTE / X
A grammar can force shape, not truth.
The complete written thought and the evidence behind it. The video edition will follow its public release.
The written argument is here.
This approved X edition is on the journal now. Its video player and original platform link will appear after each public release is verified.
A grammar can force shape, not truth.
Video caption
A grammar can force shape, not truth. Prompt, logits, schema, state and authority diverge. Constrain text, validate state, scope authority, read back. #EricFieldNotes
Full written post / accessibility read
Token masking can make malformed JSON impossible under a grammar. It cannot know whether a live order permits the requested refund. Those are different facts at different layers.
A prompt asks. A logit bias nudges. A grammar excludes invalid token sequences. State validation compares the answer with current records. A scoped payment credential limits the action. Treating these as one guardrail creates a gap where a valid answer becomes a wrong effect.
Use an illustrative order with fifty dollars refundable and a model output requesting five hundred. Its JSON is valid. The application should reject it against current order state; the payment service should reject unauthorized authority. Remove each gate in a disposable test and observe the failure.
Use grammar to make parsing reliable. At the action boundary, read current state and check the policy. Give the executor only the rights it needs, then verify the provider record. Do this because output form cannot certify business truth.
#EricFieldNotes
Four-beat scene transcript
1. A grammar can force shape, not truth.
Token masking can make malformed JSON impossible under a grammar. It cannot know whether a live order permits the requested refund. Those are different facts at different layers.
Visual: Valid tokens can still name the wrong action.
2. Every layer closes a different hole.
A prompt asks. A logit bias nudges. A grammar excludes invalid token sequences. State validation compares the answer with current records. A scoped payment credential limits the action. Treating these as one guardrail creates a gap where a valid answer becomes a wrong effect.
Visual: Prompt, logits, schema, state and authority diverge.
3. Test with a valid bad value.
Use an illustrative order with fifty dollars refundable and a model output requesting five hundred. Its JSON is valid. The application should reject it against current order state; the payment service should reject unauthorized authority. Remove each gate in a disposable test and observe the failure.
Visual: The failure should pass syntax and fail the business gate.
4. Put certainty where facts live.
Use grammar to make parsing reliable. At the action boundary, read current state and check the policy. Give the executor only the rights it needs, then verify the provider record. Do this because output form cannot certify business truth.
Visual: Constrain text, validate state, scope authority, read back.
Research and claim limits
- OpenAI structured outputs explanation (S08)
- llama.cpp grammar documentation (S09)
- llama.cpp server documentation (S10)
The examples identified as illustrative or simulated are design probes, not reported incidents. Vendor specifications do not establish workload performance.