Harden alpha grammar and auth-aware research

Diversify alpha research lanes

Prevent research generator duplicate exhaustion

Improve alpha research selection and submission verification

Implement adaptive seed pool with family-grouped empirical weights and retired dead-ends

Refine submitted seed mutation strategy, add fallback toggles, and introduce generation queue backoff

feat: implement QuantForge v2.5 payout optimizations: AST cloning filter, high-liquidity directives, and safe simulation timeout slot recycling

feat: implement strict date-based Power Pool theme targeting to plug TOP3000 universe leak

Implement lookback parameter mutation on historically submitted seed alphas

Introduce ExpressionFactory grammar generator, store raw result payload in DB, and handle incomplete/timeout simulation statuses

feat: add per-call variation seeds to eliminate LLM expression recycling; expand strategy list 10->25

Execute Phase 16: Wrap queue database session commits with IntegrityError handlers to prevent crash on duplicate expression generation

Execute Phase 13.5: Support High Turnover (HTVR) Campaign with evaluator thresholds, lookback constraints, and boosted USA regional weight

Execute Phase 13 updates: Purge ts_skewness and ts_kurtosis, define explicit safe list of operators, and prevent constant/no-trades expressions

Execute Phase 13: Synchronize all remaining local prompt adjustments, regional weight configurations, and operator diversity instructions

Phase 15: Expose silent NVIDIA errors and deep purge few-shot dataset examples

Phase 14: Align prompt schema for Nvidia NIM, OpenAI, Kimi, Zhipu, Deepseek

Execute Phase 13: Rescue reasoning outputs with fallback parsing, and switch primary NVIDIA model to standard text model meta/llama-3.3-70b-instruct

Restore primary NVIDIA NIM model to llama-3.3-nemotron-super-49b-v1.5 after 70B 404 validation

Execute Phase 12: Purge hallucinated dataset prefixes, restrict system prompt to strict Price/Volume fields, and update NVIDIA NIM model to stable llama-3.1-nemotron-70b-instruct

Execute Phase 10: Implement provider-specific response parsing and robust NoneType validation in universal router

Execute Phase 9: Align OpenAI/NIM headers Content-Type casing, implement PV-safe fallback, and integrate Dynamic Pyramid Multiplier directives

Implement deepseek provider diversion to NVIDIA NIM endpoint for nvapi keys

Switch primary NIM model to llama-3.3-nemotron-super-49b-v1.5 and update model priority list

Increase HTTP timeout to 90s, implement Nvidia key diversion for Zhipu provider, and restrict expression validator to base Price-Volume fields only

Fix universal LLM router headers bug and implement PV-safe deterministic fallback expressions

Update NVIDIA NIM Nemotron model identifier to nvidia/nemotron-3-ultra-550b-a55b in fallback list

Support both ZHIPU_API_KEY and ZHIPY_API_KEY environment variables; add z-ai/glm-5.2 to Nvidia NIM model fallbacks list

Phase 8: Free-tier optimization & fallback math fix - use ts_decay_linear in mock, silent key bypass, limit batch size to 5, 5s sleep, update Nemotron model name

Phase 7: Analytics dashboard overhaul - trash_reason tracking, enriched /api/stats, Semaphore(11), premium dark/light UI, LLM router model tracking, pause/resume endpoints

Prioritize Gemini and Nemotron models to top of list and quieten model fallback logging noise

Implement Phase 6 indestructible multi-provider LLM fallback router with 12 prioritized models

Implement sequential LLM request staggering, infinite WQ concurrent simulation limit retries, and clean None metrics trashing in evaluator

Update Gemini REST model to gemini-3.5-flash and hardcode unitHandling mapping to VERIFY

Implement native REST client for Gemini rotation, map unitHandling to IGNORE, and fix none neutralization fallback expression wrapping

Fix unitHandling value to DO_NOT_VERIFY and default advanced options to None inside generator.py

Map settings values to uppercase ON/OFF/VERIFY strings and restrict generation to WQ_ALLOWED_REGIONS

QuantForge v2.0 Phase 1 & 2: Core Async client, models, configuration, and Agentic Generation Layer