Taliesin Optimizer

2. BASELINE ESTABLISHMENT (HOST)

• Analyze the extracted text to establish a qualitative baseline.
• Axes: Writing Skill, Entertainment Value, Plot Devices, Story Coherence.
• Output: JSON baseline stored in Taliesin/.state/taliesin-optimizer/global_golden_baseline.json.

3. MUTATION (HOST INFERENCE)

• Identify the weakest axis from the current audit.
• Action: Rewrite the prose to improve the weakest axis.
• Constraint: Maintain the established Voice Signature and Story Events.

4. GRADING & FIDELITY GATE (HOST + STEEL)

• Qualitative Grade (Host): Re-score the mutated prose against the baseline.
• Deterministic Fidelity (Steel): Calculate word-set intersection between candidate and original.
• GATE: If Fidelity < 90.0%, the mutation is REJECTED.

5. LEDGERING (STEEL)

• Persist the results of each iteration to Taliesin/.state/taliesin-optimizer/benchmark_ledger.json.
• Only ADOPT the mutation if Average Score > Previous Best.

USAGE

The host executes this loop by calling the underlying Python primitives for "Steel" operations while performing all "Thinking" operations via direct host inference.

# Example Kernel-side support call for extraction
python3 Taliesin/scripts/manuscript_optimizer.py --chapter Prologue --extract-only

VERIFICATION

• Successful loop completion results in an optimized file in CStar/.lore/optimized/.
• Fidelity must never drop below the 90% floor.