Model provenance — every model used in this research

Every AI model used in this research is named, with the artefact it produced and its billing class. Strategic synthesis used paid Anthropic Claude Opus and Sonnet. Deep web research used paid Perplexity Sonar (pro, deep research, and reasoning tiers) plus a single free Alibaba Tongyi DeepResearch arm as a cross-check. Structured extraction used Claude Haiku. No model output was accepted without a verifiable source URL.

Every number in the pack is only as honest as the model that produced it. This page names every model, its billing class, and the artefact it touched.

Every claim downstream of this page is only as honest as the model that produced it. This page enumerates every model invoked across the research pipeline, its billing class, and the artefacts it produced. Verification method: grep -oE '"model":\s*"[^"]+"' across the published evidence, cross-referenced with the orchestration contract in padel-research-os/CLAUDE.md.

Why this page exists

The public portfolio at avaluev.github.io/padel-market-analysis has, until this audit, named only the Anthropic Claude family in its author footer ("Built with Claude Opus 4.7 + Claude Haiku 4.5 + Claude Sonnet 4.6"). That phrasing was incomplete. The deep-search workhorse of the pipeline was the paid Perplexity Sonar tier; one research arm ran on a free Alibaba Tongyi model. Both facts deserve the same visibility as the Claude lineup.

Radical transparency here is not a virtue signal — it is a hiring signal. Any reviewer who asks "which numbers came from a free model?" deserves an answer that points to a file, not a paragraph.

Models actually invoked

Model IDVendorRoleProduced research arms
perplexity/sonar-proPerplexity (via OpenRouter)Primary deep search and peer discoveryB_sonar_pro.json, E2_pricing_pro.json, G_gtm_sonar.json, H_naval_sonar.json, H2_naval_pro.json
perplexity/sonar-deep-researchPerplexity (via OpenRouter)Long-form research with multi-step web traversalA_sonar_deep.json, E_pricing_sonar.json, F_geo_sonar.json, J_market_size_big4.json
perplexity/sonar-reasoning-proPerplexity (via OpenRouter)Reasoning over peer set and adversarial adversarial reviewI_peers_reasoning.json, Z_redteam_final.json
anthropic/claude-opus-4.xAnthropic (via Claude Code)Plan, synthesis, adversarial review direction, final write-upOrchestration of the published evidence through the published evidence
anthropic/claude-haiku-4.5Anthropic (via Claude Code)Structured-extraction sub-agents and render templatesNormalization steps of the published evidence
anthropic/claude-sonnet-4.6Anthropic (via Claude Code)Mid-tier coding where Opus would be over-specscripts/_build_*.py generation

Free model — single arm

Model IDVendorRoleProduced research arms
alibaba/tongyi-deepresearch-30b-a3bAlibaba (via OpenRouter free tier)Redundancy arm only — output cross-validated by paid Perplexity arms before any claim entered final reportsC_tongyi.json

The Tongyi arm fed into the published evidence as one of three sources for peer discovery; every Tongyi-discovered peer was independently corroborated by B_sonar_pro.json (paid) before being added to the verified set.

Free fallback chain — declared but NOT invoked

Three free models appear in the operating contract (padel-research-os/CLAUDE.md) as a fallback chain when paid Anthropic credits are tight. Verification of the published evidence returns zero invocations of any of them in this run.

Model IDStatus
nvidia/nemotron-3-super-120b-a12b:freeDeclared in fallback chain · NOT invoked
nvidia/nemotron-3-nano-omni-30b-a3b-reasoning:freeDeclared in fallback chain · NOT invoked
inclusionai/ling-2.6-1t:freeDeclared in fallback chain · NOT invoked

This is documented to prevent confusion between declared model availability and measured model usage — two different claims, often conflated in vendor-led marketing.

How model identity is verified per claim

Every numeric claim in a downstream report cites a source_url and a verbatim quote ≤15 words. Those quotes are the gold standard. The model identity behind any given claim is recoverable in two steps:

  1. The claim cites an evidence file (e.g., the published evidence).
  2. The evidence file lists which research arm it derives from; the arm header carries the "model" field.

This page exists so a reviewer can do the trace without reading the source code.

data gaps

Transparency statement for public pages

Research pipeline ran primarily on paid Perplexity Sonar tiers (sonar-pro, sonar-deep-research, sonar-reasoning-pro) for web traversal, with one free Alibaba Tongyi arm used for redundancy. Synthesis, planning, and adversarial review ran on paid Anthropic Claude (Opus 4.x, Sonnet 4.6, Haiku 4.5) via the Claude Code orchestrator. The fallback free-model chain declared in the operating contract (Nvidia Nemotron, InclusionAI Ling) was not invoked in this run. Token-cost splits per model are a documented gap for future runs.

Source files