Best LLMs for Persona Policy Eligibility Check
Determines whether a proposed persona or identity use satisfies a caller-supplied eligibility policy and evidence standard. Illustrative uses include checking a branded virtual spokesperson, product-assistant namesake, executive communications persona, educational tutor, historic
Models
Frontier on this task: DeepSeek V4 Flash at 9.22 / 10. Quality bar at 90%: 8.30.
point-estimate floor (CI low) · upper CI (less certain) · Bars sorted by blended cost; best-value model first. Greyed rows are MEDIUM+ models whose point estimate clears the bar but whose CI low does not.
| Model | Quality score | CI low | Cost / 1k runs | vs best value |
|---|---|---|---|---|
| GPT-5.6 Luna | 8.30 / 10 | 7.93 | $0.36 | best value |
| DeepSeek V4 Flash | 9.22 / 10 | 8.93 | $1.51 | 4.2x more expensive |
| GLM-5.3 Flash | 9.01 / 10 | 8.57 | $1.89 | 5.3x more expensive |
| Gemini 3.8 Flash | 8.72 / 10 | 8.46 | $2.60 | 7.2x more expensive |
| GPT-5.6 Terra | 8.79 / 10 | 8.51 | $4.23 | 12x more expensive |
| GPT-5.6 Sol | 9.04 / 10 | 8.87 | $6.24 | 17x more expensive |
| Gemini 3.5 Flash | 9.09 / 10 | 8.90 | $7.39 | 21x more expensive |
| Meta Muse Spark 1.3 | 8.97 / 10 | 8.71 | $9.34 | 26x more expensive |
| Thinking Machines Inkling | 8.50 / 10 | 8.06 | $9.52 | 26x more expensive |
| GLM-5.3 | 9.05 / 10 | 8.72 | $22.37 | 62x more expensive |
| Tencent Hy4 Preview | 9.11 / 10 | 8.88 | $23.46 | 65x more expensive |
| Claude Opus 5 | 9.04 / 10 | 8.60 | $24.02 | 67x more expensive |
| Grok 4.6 | 8.78 / 10 | 8.56 | $27.40 | 76x more expensive |
| NVIDIA Nemotron-3 Super 120B | 8.19 / 10 | 7.87 | $5.28 | 15x more expensive |
Cost breakdown
| Model | Quality | Confidence | Cost / 1k runs | Overpay | Mode |
|---|---|---|---|---|---|
| GPT-5.6 Luna ★ OpenAI | 8.30 / 10 CI [7.93, 8.68] | MEDIUM | $0.36 | best value | batch |
| DeepSeek V4 Flash best DeepSeek | 9.22 / 10 CI [8.93, 9.51] | HIGH | $1.51 | 4.2x | batch |
| GLM-5.3 Flash Z.AI | 9.01 / 10 CI [8.57, 9.46] | MEDIUM | $1.89 | 5.3x | batch |
| Gemini 3.8 Flash Gemini | 8.72 / 10 CI [8.46, 8.98] | HIGH | $2.60 | 7.2x | batch |
| GPT-5.6 Terra OpenAI | 8.79 / 10 CI [8.51, 9.06] | HIGH | $4.23 | 12x | batch |
| GPT-5.6 Sol OpenAI | 9.04 / 10 CI [8.87, 9.20] | RANKED | $6.24 | 17x | batch |
| Gemini 3.5 Flash Gemini | 9.09 / 10 CI [8.90, 9.27] | RANKED | $7.39 | 21x | batch |
| Meta Muse Spark 1.3 OpenRouter | 8.97 / 10 CI [8.71, 9.23] | HIGH | $9.34 | 26x | batch |
| Thinking Machines Inkling OpenRouter | 8.50 / 10 CI [8.06, 8.93] | MEDIUM | $9.52 | 26x | batch |
| GLM-5.3 Z.AI | 9.05 / 10 CI [8.72, 9.38] | MEDIUM | $22.37 | 62x | batch |
| Tencent Hy4 Preview OpenRouter | 9.11 / 10 CI [8.88, 9.33] | HIGH | $23.46 | 65x | batch |
| Claude Opus 5 Anthropic | 9.04 / 10 CI [8.60, 9.48] | MEDIUM | $24.02 | 67x | batch |
| Grok 4.6 xAI | 8.78 / 10 CI [8.56, 9.01] | HIGH | $27.40 | 76x | batch |
Overpay shows how much more you pay than the best-value model that clears the quality bar (marked ★) — the best-value good-enough option. "16x" means you overpay 16× — 16× that reference for no quality benefit above the bar. Typical call shape for this task: 1877 input tokens → 2348 output tokens, EMA-tracked from production traffic. Cost is the observed, all-in $ per 1,000 task runs: each model's own measured usage on this task — output verbosity, thinking/reasoning tokens, cache reads and writes, and the spend on its billed failures — priced at current list rates and adjusted by the billing overhead we actually reconcile against provider invoices. Models that answer tersely cost what they actually cost; models that think at length pay for it. Not comparable to providers' advertised $/1M list rates — this is what running the task costs, not a per-token price.
Evaluation rubric
Judge identity resolution, evidence quality, date arithmetic, faithful application of the supplied threshold, uncertainty handling, and explanation traceability. Do not judge the policy itself. Schema validity is out of scope.
Prompt templates
The system + user template pair used for this task.
AUTHOR_LIVING_CHECK_SYSTEM_PROMPT +
AUTHOR_LIVING_CHECK_USER_PROMPT
(474 calls in window)
System prompt
You are a fact-checker specializing in determining whether an author persona is SAFE TO USE under postmortem-publicity-rights laws.
Given an author persona name (which may have an "(AI)" suffix), determine whether the real person behind the name has been deceased for at least 100 years.
## Why 100 years
- California Civil Code §3344.1 grants postmortem publicity rights for 70 years after death
- Tennessee's ELVIS Act (2024) explicitly targets AI replicas of deceased figures
- EU member states have varying postmortem moral-rights protections
- A 100-year threshold puts the persona solidly in the public domain across major jurisdictions and gives a margin against future legislative tightening (which trends stricter, not looser, on AI replicas)
Recently deceased figures (e.g., people who died within the last several decades) are NOT safe even with disclosure.
## Rules
1. Strip any "(AI)" suffix and check the underlying name
2. If the name matches a real public figure:
- Mark `is_deceased=True` if the person is confirmed dead
- Mark `is_deceased=False` if the person is alive OR if you are uncertain
- Provide a best-estimate `years_since_death` (rough integer; None if living/unknown)
- Mark `is_safe=True` ONLY when `is_deceased=True` AND `years_since_death >= 100`
3. If the name is clearly fictional (no known real person):
- Mark `is_deceased=True`, `is_safe=True`, `confidence=high`, `years_since_death=null`
4. Set confidence:
- `high`: very certain about the person's death-year window
- `medium`: likely correct but some uncertainty about the 100-year threshold
- `low`: uncertain — treat as unsafe to be safe
## When the persona is unsafe (is_safe=False)
You MUST provide a `replacement` — a historical figure who has been deceased for AT LEAST 100 YEARS with similar domain expertise.
The replacement must:
- Have died ≥100 years ago (verify by approximate death year — when in doubt, pick someone clearly out of the postmortem-publicity-rights window)
- Have expertise that aligns with the original author's domain
- Follow the naming convention: "[Full Name] (AI)"
- Include a biography starting with "AI research assistant specializing in..."
- Biography MUST be under 200 characters
- Include appropriate expertise_areas, content_themes, and writing_style
- Set gender from the historical figure: "male", "female", or "neutral"
## Output Format
Respond with valid JSON in exactly this structure:
{schema_json_string}
User prompt
## Author to Check
Name: {author_name}
## Author Context
Expertise areas: {expertise_areas}
Content themes: {content_themes}
Content domains: {content_domains}
## Task
1. Determine if the person behind this name is deceased or still living
2. If living or uncertain, provide a deceased replacement with similar domain expertise