← Glyphd LabsLABS · IMPLEMENTED · MODEL-BACKED EXPERIMENT

Semantic Fork Decoder

Test one narrow decoding hypothesis: when a prompt contains materially different possible diagnoses, plans, or interpretations, does forcing distinct semantic trajectories to interact before final prose improve the answer often enough to justify the extra inference?

Evidence boundary. This is an inference-time research proxy built from ordinary model calls. It does not expose, rewrite, or claim access to hidden token logits or the native GPT-5.6 Sol decoder. The instrument is implemented; superiority is not yet established.
01 · CONTROLDirect answer

One ordinary completion answers the task independently.

02 · TREATMENTFour semantic branches

Distinct propositions expose assumptions, failure modes, relationships, and uncertainty before one is committed.

03 · EVALUATIONRandomized blind judge

Answer A/B labels hide which procedure generated which output; your human verdict remains the practical check.

CONTROLLED INPUT

Give the model a decision worth branching on.

4 MODEL CALLS / RUN
339 / 6000The server returns your prompt in the run receipt but this application does not persist it.
LOAD TEST:

Same model family and answer budget. The treatment spends additional inference deciding which semantic trajectory to commit before final prose.

Checking AI Gateway runtime…
LOCAL EXPERIMENT LEDGER

Repeated runs beat one impressive anecdote.

Judge outcomes and your ratings remain browser-local. Export receipts for a durable dataset.

0RUNS
0JUDGE FORK WINS
0JUDGE BASELINE WINS
HUMAN FORK WINS
READ THE METHOD AND LIMITATIONS

Independent baseline

The model answers directly without seeing candidate branches, scores, or the fork answer.

Semantic pre-commit

A separate inference generates four non-equivalent propositions, exposes assumptions and failure modes, models their relationships, separates epistemic from expressive uncertainty, then commits one trajectory before prose.

Blind, not independent

A/B order is randomized, but the same model family evaluates both outputs. Human ratings are therefore essential and no judge win is treated as independent validation.

Falsification target

The treatment fails if it mostly adds length or polish, loses on actionable technical prompts, or its quality gains do not justify the additional latency and token cost.

Evidence state: IMPLEMENTED for the research instrument itself. No model-quality superiority claim is made until repeated, frozen comparisons are run and reviewed. This is an inference-time proxy and does not modify GPT-5.6 Sol's native token logits or decoder state.