Someone who asks one language model a question cannot tell a confidently wrong answer from a confidently right one. The two are identical on every surface cue available to a non-expert. Asking several models seems to solve this, but it mostly relocates the problem: reconciling several answers produces one confident text again, and that text now carries the borrowed authority of “ten of them agreed.” The protocol formalized here collects answers from several consumer chat models through the user’s own browser or phone, with no API keys, no installation and no intermediary server. It reconciles them by anchor-and-enrich: the strongest available answer becomes the base, and each of the others contributes only what the base lacks. Every contribution is signed with the model that supplied it. Disagreements are kept in a section of their own rather than averaged away. The paper makes four contributions. The first is the protocol itself, stated as an algorithm that preserves provenance and forbids silent deletion, together with a taxonomy of five ways a merge can fail. Four of them are defects of naive pooling; the fifth is a failure of the operator rather than of the method: a synthesis that presents a single model’s output as a collegial result. The second is the application of the correlated-witness arithmetic to protocol design. That arithmetic is not claimed here. Kohli (2026) applied the same Kish correction to a panel of nine frontier judges and measured about two effective votes, and Kim et al. (2025) established the underlying error correlation across more than 350 models. A council of N models whose errors carry intraclass correlation ρ is worth N_{eff} = N / (1 + (N−1)ρ) independent witnesses, and that quantity saturates at 1/ρ however many models are added. Once a council is larger than 1/ρ, halving ρ beats doubling N, and a pair drawn from two laboratories can outweigh five drawn from one. Kim et al. (2025) complicate that rule: the larger and more accurate models carry the most correlated errors even when architecture and provider differ, so two laboratories are a prior guess about ρ rather than a measurement of it, which is what the estimation procedure of Section 5.5 is for. The third is an account of three deployment tiers, manual, one-tap mobile and browser-agent, and of the axis that separates them epistemically: whether the second model sees the first model’s answer. Conditioning the second model raises ρ in exchange for error correction. The fourth is a small-sample observation of the counterfeit-council failure across five models, the one-sentence intervention that removed it in a three-model re-run, and a blinded three-arm evaluation design with a stated falsification condition. Correlated agreement is treated here as established, not as a finding. That a panel of models is worth far fewer independent witnesses than it has members has been measured directly: nine frontier judges drawn from seven families supply about two effective votes (Kohli, 2026), and error correlation rises with model capability even across providers and architectures (Kim et al., 2025). Section 5 restates the design-effect arithmetic behind those measurements, because three of the protocol’s rules are derived from it and read as fastidiousness without it. The protocol then takes the other side of the same problem. That literature measures the correlation; this protocol is built to lower it, by composing across lineages, keeping disagreement instead of averaging it away, signing every contribution with the model that supplied it, and counting a repeated source once. The claim that a merged answer beats the best single answer is argued from the design and from adjacent literature; Section 9 specifies the evaluation, which has not been run. The same text is available in three other places, so that it can be reached when one of them cannot serve it: Archived and citable copy: https://doi.org/10.5281/zenodo.22683716 Full text as Markdown and PDF: https://github.com/vadimchernets/papers/tree/main/agreement-not-evidence Readable HTML: https://vadymchernets.netlify.app/agreement-not-evidence.html
This retrospective research brief examines the institutional structure, analytical sequencing, and scenario construction of the 2012–2013 National Capital Planning Commission–District of Columbia Height Master Plan review. The congressional request initiated a jointly conducted review, and the agencies subsequently defined federal–local consensus and a jointly prepared master plan identifying where additional building height might be appropriate as the expected outcome. NCPC and the District completed substantial collaborative technical work, public engagement, economic analysis, and visual modeling. The process nevertheless culminated in separate federal and District reports containing materially different recommendations. The brief evaluates whether the review completed a jointly approved, spatially explicit test of its strongest potential middle alternative: define and map agreed primary viewsheds and other protected geography first, identify the legally and physically plausible geography remaining after those exclusions, and evaluate selective additional height only within that residual area. The documentary and spatial audit finds that NCPC performed a partial federal-side screen that removed certain viewshed streets, Pennsylvania Avenue, and streets in historic districts from its ratio analysis. The reviewed public record, however, does not demonstrate that NCPC and the District jointly approved that exclusion geography or integrated it with the study’s selective-height modeling, economic analysis, capacity measures, and final institutional findings. The conclusion is methodological rather than pro- or anti-height. The brief does not allege bad faith, deny that interagency collaboration occurred, claim that Congress legislatively required a joint final report, or establish that a selective-height policy would necessarily produce meaningful development or public benefits. It identifies what the published review completed, what remained institutionally unreconciled, and how an exclusion-first test could structure any renewed Height Act analysis.
