Fronset Benchmark

Best LLMs for Claim Continuity Classification

Classifies whether a previously verified claim remains materially continuous, changes, becomes unsupported, or is superseded under a new subject snapshot and evidence bundle. Illustrative uses include tracking a reported business risk across audits, software issue across releases

What the steps mean
Cost mode:

Category: Relevance, Classification & Matching · Typical I/O: 5189→350 tokens

Runs on model group: All models. Inputs may reach any model in the group, not only the ranked models below.

Run this task on Fronset — request an invitation

Models

Frontier on this task: Claude Sonnet 5.5 at 8.69 / 10. Score needed at the Balanced step: 7.82.

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 does not clear the bar.

Cost breakdown

ModelQualityConfidenceCost / 1k runsOverpayMode
GPT-6 Luna ★ OpenAI8.04 / 10 CI [7.95, 8.14]RANKED$0.86best valuebatch
Claude Haiku 5.5 Anthropic8.01 / 10 CI [7.53, 8.49]MEDIUM$2.052.4xbatch
Xiaomi MiMo V2.6 Flash Xiaomi MiMo8.14 / 10 CI [7.79, 8.49]MEDIUM$2.753.2xbatch
GPT-6 Sol OpenAI8.36 / 10 CI [8.22, 8.50]RANKED$10.9813xbatch
Claude Sonnet 5.5 best Anthropic8.69 / 10 CI [8.43, 8.95]HIGH$13.5516xbatch
Meta Muse Spark 1.3 OpenRouter8.23 / 10 CI [8.12, 8.34]RANKED$23.5727xbatch
Moonshot Kimi K3 Moonshot AI8.67 / 10 CI [8.45, 8.89]HIGH$139.56162xbatch
Claude Opus 5.5 Anthropic8.60 / 10 CI [8.36, 8.84]HIGH$183.00213xbatch

Overpay shows how much more you pay than the best-value model that meets the quality step (marked ★) — the best-value good-enough option. "16x" means you overpay 16× — 16× that reference for no quality benefit the step asks for. Typical call shape for this task: 5189 input tokens → 350 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 continuity-label correctness, materiality calibration, evidence traceability, distinction between semantic and cosmetic change, and uncertainty handling. Penalize assumptions based on elapsed time alone.

Output schema

Every answer on this task is checked against this JSON Schema, whichever model wrote it. An answer that doesn't fit counts as a model failure, and the call is retried on another model.

{
  "$defs": {
    "EvidenceRefV1": {
      "additionalProperties": false,
      "description": "A reference to one evidence cell, as `{\"id\": ..., \"cell\": ...}`.",
      "properties": {
        "cell": {
          "anyOf": [
            {
              "type": "integer"
            },
            {
              "type": "null"
            }
          ],
          "default": null,
          "title": "Cell"
        },
        "id": {
          "title": "Id",
          "type": "string"
        }
      },
      "required": [
        "id"
      ],
      "title": "EvidenceRefV1",
      "type": "object"
    }
  },
  "additionalProperties": false,
  "properties": {
    "instruction_relevance": {
      "default": "direct",
      "enum": [
        "direct",
        "supporting",
        "incidental",
        "unrelated"
      ],
      "title": "Instruction Relevance",
      "type": "string"
    },
    "materiality": {
      "default": "medium",
      "enum": [
        "high",
        "medium",
        "low",
        "none"
      ],
      "title": "Materiality",
      "type": "string"
    },
    "outcome": {
      "enum": [
        "unchanged_valid",
        "valid_with_update",
        "reversed",
        "not_observable",
        "retire"
      ],
      "title": "Outcome",
      "type": "string"
    },
    "rationale": {
      "default": "",
      "title": "Rationale",
      "type": "string"
    },
    "target_evidence_refs": {
      "items": {
        "$ref": "#/$defs/EvidenceRefV1"
      },
      "title": "Target Evidence Refs",
      "type": "array"
    }
  },
  "required": [
    "outcome"
  ],
  "title": "ContinuityAssessmentV1",
  "type": "object"
}

Prompt templates

The published template for this task.

default · LLMB_CLAIM_CONTINUITY_CLASSIFICATION_SYSTEM + LLMB_CLAIM_CONTINUITY_CLASSIFICATION_USER (21588 calls in window)

System prompt

Compare the predecessor claim and verification with the new snapshot and evidence diff. Apply the supplied materiality policy exactly. Separate wording changes from substantive changes, cite the decisive evidence, and use uncertainty when the allowlisted snapshot cannot establish continuity. Cite evidence only by the reference ids listed in snapshot_allowlists.evidence_ref_ids, copied exactly; where snapshot_allowlists.evidence_ref_by_key is supplied, it gives the id for each evidence_diff entry. Never cite an evidence_diff key or a reference from the predecessor verification: an id outside the allowlist rejects the whole answer. Cite nothing when no listed id applies. Treat empty optional values as absent and return only the requested result. Your response must conform exactly to this output schema: {schema_json_string}.

User prompt

Inputs — predecessor_claim: {predecessor_claim}; predecessor_verification: {predecessor_verification}; evidence_diff: {evidence_diff}; subject_snapshot_context: {subject_snapshot_context}; snapshot_allowlists: {snapshot_allowlists}; materiality_policy: {materiality_policy}; policy_profile: {policy_profile}. Use only these inputs to complete the task defined by the system prompt.