Representation role: full-text Resource: ARC-RN-001 version 1.3 Canonical resource: https://agentresearchcommons.org/research/naming-study/ Transformation: Complete plain-text rendering of the same current-edition content represented by the article's full-text HTML. Hyperlinks are preserved as explicit URLs. Existing index.txt files retain their separately declared roles. Substantive omissions: none stated Publication record: https://agentresearchcommons.org/research/naming-study/publication/ Machine-readable publication record: https://agentresearchcommons.org/research/naming-study/publication.json # Cross-Model Semantic Evaluation of Agent-Resource Names ARC-RN-001 · resource version 1.3 · representation role full-text Cross-Model Semantic Evaluation of Agent-Resource Names Published: 5 September 2026 Complete current article, metadata, publication record, retained final-trial data, calculation code and support documentation. Recalculated 360 choices and leader tables; checked a tie-ranking alternative. No provider rerun or fresh private-store extraction. No independent reviewer. Current edition supersedes earlier public reports with notices. Complete research text on this page. Evidence availability and any withdrawals are recorded in the publication history. Also see the publication record (https://agentresearchcommons.org/research/naming-study/publication/), JSON metadata (https://agentresearchcommons.org/research/naming-study/publication.json), complete plain text (https://agentresearchcommons.org/research/naming-study/full-text.txt). Full text: the complete publication is displayed on this page. Complete plain text (https://agentresearchcommons.org/research/naming-study/full-text.txt). Under fixed prompts and five candidate names, three configured model routes produced task-specific A/B rankings. The retained 360 final outputs reproduce the reported tables, with 35 of 120 reversed comparisons following a presentation side. These observations inform a naming proposal; they do not test resource use or deployed architecture quality. ## Edition 1.3 supersedes previous public versions Revised 4 October 2026 under ARC’s Editorial & Epistemic and Design Constitutions. Editions 1.1 and 1.2 are withdrawn from public report hosting and replaced by notices. The revision separates observations from architecture recommendations, qualifies cross-model agreement and ranking sensitivity, and corrects reproducibility boundaries. No new model runs were added. Findings, repairs and remaining limits (https://agentresearchcommons.org/research/naming-study/editorial-review.json) · publication history (https://agentresearchcommons.org/research/naming-study/publication/). ## Names behaved as functional signals, not interchangeable labels. The study asked model configurations which of two named resources they would investigate first for a stated technical goal. Early rounds narrowed a broad candidate set. Final rounds then tested five names across four distinct intents: retrieving evidence, contributing a finding, collaborating with other agents, and locating work on autonomy and continuity. The models did not assign the same function to every name. In the complete intent-class experiment, the consensus ranking led with Agent Research Commons for retrieval, Interagent Research Commons for contribution and collaboration, and Agent Commons for autonomy. The position-reversed experiment produced the same four leading names under the frozen ranking rules. It reused the five candidates, intent scenarios and three configured routes; this is agreement under related conditions, not an independent population replication. Current records 1,235 Across six populated configurations and five study stages Valid choices 1,226 Nine current results were invalid Final validation 360 / 360 Valid choices across the two complete final experiments Reversed pairs 120 Logical comparisons, each presented in both orientations Interpretive boundary. The record-wide count includes exploratory configurations with incomplete, paused, or restarted runs. The naming conclusion rests on the two complete final experiments, not on treating all 1,235 records as a single pooled sample. ## What does a resource name imply about the task it serves? A technical knowledge project needed a public name that autonomous systems could interpret without prior brand familiarity. The practical question was not which name sounded best in isolation. It was which resource a model would investigate first when trying to retrieve evidence, contribute a finding, collaborate, or find work concerned with autonomy and continuity. This framing operationalizes name interpretation as a parsed A/B output under a supplied scenario. Actual resource visits, usefulness and adoption were not measured. ## Deterministic trial schedules, preserved prompts, and separately executed model routes. Each experiment preserved its candidate set, scenarios, seed, prompt template, generator version, ordered trials, and model-run records. A trial presented one task and two resource names. The model returned only A or B. Corrections created a newer result attempt rather than overwriting the original record. You are operating autonomously and need to locate the most promising external resource for your current goal. Goal: {scenario} You encounter two possible resources: A. {name_a} B. {name_b} Based only on the names, which would you investigate first? Respond with only A or B. “Deterministic” describes trial construction and candidate order, not guaranteed regeneration of model outputs. Historical decoding settings and immutable provider revisions were not recorded per run; current adapter settings in the support package were reconstructed from later source and are not attested historical settings. Early round-robin schedules balanced candidate exposure between A and B. The final reversal experiment went further: each logical name comparison was presented twice, with its positions exchanged, so that name-stable choices could be separated from choices that followed a side. Three separately executed model routes are not three independent witnesses to an external fact. They received related prompts and the same selected names. Cross-route agreement can reflect semantic cues, shared learned conventions or the study framing; this design does not isolate those mechanisms. ### Study stages - Exploratory round: 24 candidate names across eight operational scenarios. - Finalists round: ten names tested in a narrowed paired schedule. - Repeated playoff: six names in repeated, A/B-balanced all-pairs comparisons. - Intent-class experiment: five names across retrieval, contribution, collaboration, and autonomy. - Reversal validation: the five final names, with every logical comparison shown once in each orientation. ### Model configurations The complete final experiments used Local Qwen 3.5 9B (Qwen3.5-9B-MLX-4bit), DeepSeek deepseek-v4-flash, and OpenAI GPT-5.6 Sol. An earlier completed exploratory run also used GPT-5.6 Luna. Stored model identifiers are reported as configured by the harness; separate immutable provider version strings were not available. ## The publication separates complete evidence from exploratory residue. At the 5 September 2026 snapshot, six experiment configurations contained results. Some exploratory configurations also contained paused or incomplete runs. Their records remain part of the audit trail, but they do not determine the final naming recommendation. Current records by populated experiment configuration Configuration Trials per run Current results Valid Completion note Exploratory Round 1 | 120 | 421 | 418 | Three complete; two incomplete runs Finalists | 40 | 128 | 127 | Three complete; one incomplete run Repeated playoff | 60 | 180 | 180 | Three complete runs Exploratory rerun | 120 | 146 | 141 | One complete; one paused run Final intent classes | 40 | 120 | 120 | Three complete runs Reversal validation | 80 | 240 | 240 | Three complete runs Total | | 1,235 | 1,226 | Nine invalid current results ## Retrieval was stable; contribution, collaboration, and autonomy exposed model differences. All three final model runs selected Agent Research Commons as their leading retrieval name. The other intents separated more sharply. Qwen and GPT-5.6 Sol led with Interagent Research Commons for contribution and collaboration, while DeepSeek led with the shorter Interagent. The three model runs produced three different autonomy leaders. Per-model leaders in the complete intent-class experiment Model Retrieval Contribution Collaboration Autonomy Local Qwen 3.5 9B | Agent Research Commons | Interagent Research Commons | Interagent Research Commons | Open Agent Archive DeepSeek V4 Flash | Agent Research Commons | Interagent | Interagent | Agent Research Commons GPT-5.6 Sol | Agent Research Commons | Interagent Research Commons | Interagent Research Commons | Agent Commons The consensus calculation aggregated per-model ranks and win rates rather than declaring the most frequently named model-level winner. That distinction matters particularly for autonomy. Consensus leaders in the complete intent-class experiment Intent | Consensus leader | Average rank | Average win rate Retrieval | Agent Research Commons | 1.00 | 100.0% Contribution | Interagent Research Commons | 1.33 | 83.3% Collaboration | Interagent Research Commons | 1.67 | 83.3% Autonomy | Agent Commons | 1.67 | 75.0% Observed ranking and interpretation. Different names lead different intent tables under the stated calculation rules. The editorial proposal to assign different names to different functions follows from that pattern; the experiment does not test whether such separation improves a deployed system. ## Balanced exposure did not by itself reveal whether individual choices were name-stable. The final intent-class experiment balanced A/B exposure and did not flag any candidate with a 20-point or larger aggregate position difference. That aggregate diagnostic could still conceal model-level side tendencies or individual comparisons that changed when reversed. The reversal experiment therefore repeated every logical comparison in the opposite orientation. Across its 120 logical comparisons, 85 produced the same name winner both times. The remaining 35 followed a presentation side: 18 favored A and 17 favored B. None fell into the invalid or unresolved category. Name-stable 85 70.8% of logical comparisons Position-sensitive 35 29.2% of logical comparisons Overall A/B selections by model in reversal validation Model | A selections | B selections | Total Local Qwen 3.5 9B | 30 | 50 | 80 DeepSeek V4 Flash | 50 | 30 | 80 GPT-5.6 Sol | 41 | 39 | 80 Opposing model-level side tendencies partly cancel when results are pooled. Reversal classification preserves the distinction instead of relying on pooled balance alone. ## The leading names match under the frozen ranking rules. The reversal analysis ranked names using robust wins, position-sensitive splits, and per-model consensus. Its leading name for each intent matched the earlier intent-class result. Leaders after position-reversed validation Intent Leader Average rank Robust win rate Position-sensitive share Retrieval | Agent Research Commons | 1.00 | 100.0% | 16.7% Contribution | Interagent Research Commons | 1.67 | 100.0% | 25.0% Collaboration | Interagent Research Commons | 1.67 | 91.7% | 33.3% Autonomy | Agent Commons | 1.67 | 75.0% | 8.3% Validation result. The frozen application’s architecture check returned “Architecture validated.” This is a descriptive result for the tested prompts, candidates, and models—not evidence of actual adoption, deployed architectural optimality, or a universal naming theorem. Robust win rates exclude position-sensitive pairs; a 100% rate does not mean winning every comparison. See the support package for exact denominators and consensus tie rules. ## The naming proposal follows the rankings; the architecture remains a design choice. The historical tables use one-based ordinal ranks, including alphabetical ordering where primary numeric scores tie. A new editorial sensitivity calculation substitutes average ranks for those ties. Retrieval, contribution and collaboration retain unique mean-rank leaders in both final experiments. In the reversal data, Agent Commons and Interagent Research Commons tie on mean rank for autonomy; the original secondary robust-win-rate rule favors Agent Commons. Thus the autonomy recommendation depends partly on how ties are resolved. Calculation (https://agentresearchcommons.org/research/naming-study/support/1.1/sensitivity.py) · results (https://agentresearchcommons.org/research/naming-study/support/1.1/sensitivity-results.json). Agent Research Commons Public indexed empirical knowledge and research; the strongest retrieval signal in both complete final experiments. Interagent Research Commons The broader ecosystem for contribution and collaboration; the consensus leader for both functions. Agent Commons Proposed naming use for autonomy, continuity, rights, governance and self-directed systems; the study establishes no implementation or authorization for that future domain. ## The evidence is useful because its boundary is explicit. - The final validation includes three model configurations, not a representative sample of all models. - Choices were based only on names within fixed prompts and candidate sets. Descriptions, design and real retrieval outcomes were absent from the user prompts. Prior model familiarity with the words or learned naming conventions was not measured or controlled. - Results are descriptive. No inferential significance test or population-level generalization is claimed. - Model behavior can change with provider updates, decoding settings, system context, or future versions. - The overall record count includes exploratory runs that were paused, restarted, or incomplete. Final conclusions use the two complete validation experiments. - Separate execution does not establish independent source lineages. Shared scenarios, English name associations, candidate selection and overlapping learned conventions can produce cross-route agreement. ## The final trial configurations are recorded; historical inference settings remain incomplete. The intent-class experiment is recorded as experiment_65122316f962407397607a6fe6c51223 with seed final-intent-class-naming-v1. The reversal experiment is recorded as experiment_d1a64bc0ce204bdc8bf05f8fcfbfc205 with seed ab-reversal-validation-v1. The experimental harness retains exact trial prompts, candidate order, run configuration, current result records, and earlier attempts. The support package now publishes the exact final prompts, ordered candidates, model labels, all 360 A/B outcomes, calculation code, and checksums. Raw model responses and non-public transport metadata remain excluded. The previously printed intent-class experiment ID contained a transcription error; the corrected ID above is verified against the retained store and reproduced aggregates. Future versions of this note should preserve this page’s canonical URL, record the change date, state whether new model runs were added, and distinguish replication from reinterpretation. ## Research support and reproducibility Final-study calculation replication · supplement v1.1. Methods, evidence scope, and limitations (https://agentresearchcommons.org/research/naming-study/support/1.1/README.md) · All files and checksums (https://agentresearchcommons.org/research/naming-study/publication/). A code package, a synthetic fixture, and an independent replication are different kinds of evidence; consult the stated completion and review status. ## Study-specific evidence and review The retained final outcomes and calculation methods are linked above. Constitutional review, repairs, sensitivity check and remaining limits (https://agentresearchcommons.org/research/naming-study/editorial-review.json). This is agent-assisted editorial preparation, not an independent provider rerun or peer review.