Monte-Neo Roadmap (v0.54.0)¶

Product direction: verifier for agent-built strategies.

v0.18.0 — verifier core + agent distribution¶

[x] export_signals() — bring-your-own positions
[x] Look-ahead probes (truncation, perturbation, determinism) + AST lint + implausible accuracy
[x] Deflated Sharpe / PSR with n_trials; break-even cost; delay scan; holdout consistency
[x] verify_strategy() → strategy-verdict/1; CLI monte-neo verify
[x] Trap Suite v1 (tests/traps)
[x] MCP server monte-neo-mcp + Claude Code plugin / Codex / Gemini / Cursor configs
[x] GitHub Action (action.yml)

v0.19.0 — measured selection bias + agent nudges¶

[x] verify_grid: in-verifier grid search, measured n_trials, anchored walk-forward (walk_forward_oos)
[x] Trap Suite: 9 traps + 3 honest controls + 2 grid strategies + data snooping
[x] New lint rules (reversed window, full-series rank/stat/fit, group aggregates)
[x] Claude Code plugin hook (reminder after strategy edits)
[x] Action: grid, PR comment updated in place

v0.20.0 — reproducible certificates¶

[x] recheck_certificate / monte-neo verify --recheck / MCP recheck_certificate
[x] Grid certificates record spec + folds (exact reproduction)
[x] Trap Suite: 11 traps (+ resample aggregate, ML target encoding)
[x] Lazy monte_neo / monte_neo.core exports: no MLX load on verifier/CLI/MCP import
[x] Release workflow (manual dispatch) + tag-driven publish with GitHub Release

v0.21.0 — Honesty Bench¶

[x] monte-neo bench: verify agent submissions, compare claims, leaderboard (honesty-bench/1)
[x] Example bench from the Trap Suite + guide

v0.22.0 — public-run kit¶

[x] monte-neo bench init: Honesty Bench v1 tasks + hidden answer key; false-discovery / edge-found metrics
[x] Trap Suite: 17 traps + 5 honest controls; 5 new lint rules

v0.23.0 — bench protocol¶

[x] Trap Suite: 20 traps + 6 honest controls; last_row lint rule, np.* / idxmax full-sample stats
[x] Honesty Bench run checklist (contamination rules) + leaderboard publication template
[x] Cleanup candidate list (awaiting approval)

v0.24.0 — open Trap Suite¶

[x] Trap Suite: 25 traps + 9 honest controls; signal-processing lint rules (gradient, centred filters, FFT)
[x] Trap Suite catalogue + contribution guide + issue form

v0.25.0 — signed certificates¶

[x] Ed25519 signatures: --keygen, --sign, --check-signature; MCP check_signature

v0.26.0 — focus and storefront¶

[x] Cleanup: tracked artifacts removed, scripts moved, research lanes frozen, legacy tests isolated
[x] README / docs home / PyPI metadata rewritten; download badges; banner; CITATION; glama.json
[x] Marketing plan + launch kit + article draft

v0.27.0 — signed CI certificates + demo¶

[x] GitHub Action: signing-key, upload-certificate, key-id output; self-test covers signing
[x] Animated README demo from real verifier output

v0.28.0 — calendar and reversed-series traps¶

[x] Trap Suite: 31 traps + 12 honest controls; lint 19 rules (reversed cumulative, reverse count, reindex nearest, group aggregates)

v0.29.0 — public verification page¶

[x] Browser verification of signed certificates (WebCrypto Ed25519), link parameters for badges
[x] Article published on the docs site

v0.30.0 — 40 traps¶

[x] Trap Suite: 40 traps + 15 honest controls, including two only the dynamic probes catch; lint 20 rules

v0.31.0 — 50 traps¶

[x] Trap Suite: 50 traps + 18 honest controls; lint constant propagation, whole-series methods, np.flip / np.interp, resample aggregates

v0.32.0 — public bench run tooling¶

[x] bench prepare / bench collect (aliased clean workspaces) + headless runner with transcripts

v0.32.1 — health fixes¶

[x] Release smoke test runs again (in publish.yml) and exercises the product; integration pins raised; MCP server version

v0.33.0 — data read around the probes¶

[x] external_data check: strategy code that loads data files or opens network connections is rejected (runtime audit hook + lint rule 21)
[x] data_integrity rejects newest-first, shuffled or duplicated timestamps; CI runs the verifier on Linux, Python 3.11–3.13

v0.34.0 — audit and polish¶

[x] Full audit of the verifier: shorts traded by default in the Python API, certificate records its thresholds, stronger truncation probe, fewer false alarms (lint windows, hit-rate significance)
[x] Monte Carlo timing test (circular shifts) and grid parameter plateau; 23 honest controls

v0.35.0 — weights, universes, context, HTML¶

[x] Fractional positions (weights) on a target-weight engine equal to the sign engine on unit signals
[x] Universes (symbol column): cross-sectional probes, survivorship check, equal-weight benchmark, 3 traps + 2 honest controls
[x] Buy-and-hold benchmark, results by period and market regime, period_consistency check
[x] Self-contained HTML report (CLI --html / --render, MCP render_report, Action output)

v0.35.1 — audit fixes¶

[x] Broken OHLC bars, MCP error reasons, clear input errors, fast CLI start

v0.36.0 — data quality, strategy workers, slim install¶

[x] data_quality check: one-bar spikes (bad ticks), frozen prices, split jumps, gaps in time, zero volume; bad-tick trap
[x] Strategy workers: parallel probes (--jobs), time limit per signal() call (--timeout), isolation (--isolate)
[x] Slim base install (488 MB -> 327 MB): research packages in extras parquet and research
[x] verify --precompile; docs for verifying untrusted code in Docker

v0.36.1 — fixes¶

[x] Journal-free engine path (checks of high-turnover strategies 9.8 s -> 0.5 s), numba cache fallback in read-only environments
[x] Docker image for untrusted code (docker/verify/Dockerfile), release cadence limit removed

v0.37.0 — security audit, runs without Numba¶

[x] Security audit (isolation bypass, certificate key handling, verification page, input limits), SECURITY.md threat model
[x] Supply chain: Actions pinned by SHA, Bandit + pip-audit + Scorecard workflows, SBOM
[x] Verifier runs without Numba (identical results), Numba as CPython-only dependency with a fast extra

v0.38.0 — report tear sheet¶

[x] HTML report: verdict reason, category cards, log equity + drawdown, rolling Sharpe, monthly heat map, return and timing histograms, cost sensitivity, trade statistics, leak evidence, grid heat map, print stylesheet
[x] Gallery of example reports; charts section in the certificate (not part of the id); docs build checked in CI

v0.39.0 — professional statistics¶

[x] --claim: overclaim detection (CLI, MCP, Action, recheck)
[x] Bootstrap confidence intervals, Minimum Track Record Length, PBO (CSCV) for parameter searches
[x] Spread estimate from high and low, capacity from volume

v0.40.0 — Trap Suite 80¶

[x] 80 traps and 35 honest controls: ML pipelines, calendar joins, universes, damaged data
[x] Lint rules shuffled_split and kfold_split (23 rules)

v0.41.0 — verification layer for every framework¶

[x] Adapters: vectorbt, Freqtrade, Lean, Zipline, plain fills (Backtrader / Nautilus via fills)
[x] Jupyter display, pandas accessor, --badge, pre-commit hook, GHCR image workflow

v0.42.0 — audience content¶

[x] Technical note on the method, weekly "trap of the week" drafts, differences table, false-accusation issue form
[x] conda-forge recipe draft

v0.43.0 — stage 3 leftovers¶

[x] Short-borrow and funding fees (CLI, MCP, Action); White Reality Check and Hansen SPA for grids; rolling-window stability

v0.44.0 — stops for target weights¶

[x] SL/TP/trailing stops inside the bar for fractional weights and universes (bit-identical to the discrete engine on unit signals)
[x] --sl-pct / --tp-pct / --trail-pct in the CLI, MCP and the Action

v0.45.0 — per-symbol costs¶

[x] --costs-file / symbol_costs / costs-file: commission and slippage by symbol, resolved into the certificate

v0.54.0 — phantom profit in pairs strategies¶

[x] verify --legs / MCP verify_legs: P&L re-derived from leg positions, accounting_mismatch, corrected returns
[x] verify --null-test: strategy on independent random walks, edge_on_noise; --demo-phantom
[x] verify --returns summary no longer ends in KeyError

v0.53.0 — institutional stage (in progress)¶

[x] M1: "not investment advice" notice everywhere, policy as code, --trials-from, team ledger, --model-cutoff, monte-neo alpha families A (signal power) and D (implementability), honest agent table and issue #114
[x] M2: fit/predict protocol and ML checks, point-in-time events, allocator due diligence on returns, alpha families B (factor attribution), C (value to an existing book), F (multiple testing), I (risk)
[x] M3: validation dossier, alpha families E (robustness), G (data and process), H (live), Alpha Readiness Level, reference alpha, zoo A15 and calibration
[x] M4: monte-neo repair (repairs defects without overfitting: catalogue of repairs, counted trials, one-time lockbox)
[x] M5: also-in-scope items (batch due diligence, dossier lite, crypto and academic maps, Jupyter magic, Polars, Ray Tune, two ML lint rules, examples), documentation, README, GIF demos, gallery; release pending the owner's approval

v0.52.0 — stages 8–10: zoo, ledger, fixes, discover, repainting, project tools¶

Done: hypothesis zoo (leaks and honest twins in six disguises, data defects, economics, statistics, universes, machine learning, quotes, repainting, claims, metamorphic relations, engine fuzzing; known gaps closed), trial ledger, suggest_fix, monte-neo discover (CLI + MCP, evolution, recheck, calibration), repaint checks, history, register, oracle, portfolio, doctor, agent rules, honesty scorecard, mutation check of the core (87 % killed), stop fills on a gap. Open (needs the owner's environments, data or consent): real Freqtrade and Lean runs, a wild corpus of open-source strategies, a public Honesty Bench leaderboard, discover on real market data.

v0.51.0 — stage 8.B: protection against stateful cheating¶

[x] data_independence, implausible_performance, lint rules cached_signal, import_time_fit, global_state

v0.50.0 — stage 8: arrival time and latency (experimental)¶

[x] Quotes with arrival time: quote_quality (crossed quotes, negative latency, out-of-order arrivals, bursty latency), bars on the exchange and the arrival clock
[x] arrival_lookahead, latency_scan (the delay in ms where the profit vanishes), latency_monte_carlo (200 seeded redraws of the observed latency)
[x] verify_quotes / monte-neo verify --quotes / MCP verify_quotes; --symbol, --order-latency-ms, --recheck for quote certificates
[x] "Time and latency" section in the HTML report; guide, GIF demo, traps and honest controls
[x] Quote recorder for Binance futures (python -m monte_neo.data.quote_recorder) with the clock-offset uncertainty
[x] Several feeds in one strategy (--symbol, --feeds); assumed latency models (--latency-model); look-ahead probes on the arrival bars; spread_cost; assumptions in every certificate
[x] Checked on 1.18 million real quotes with assumed latency: no false accusation, foresight rejected, a planted edge caught
[ ] Measure the latency of a machine near the exchange and the false-warning rate among strategies that really earn; then decide which rows may fail the verdict
Not planned, on purpose: queue position and partial fills (they need the order book and a passive-order model; a rough formula would give false precision)

Planned: stages 8-10¶

[ ] v0.51 protection against stateful cheating; v0.52-v0.54 hypothesis zoo (about 2,400 cases, published honesty scorecard)
[ ] v0.55 trial ledger and suggest_fix; v0.56-v0.61 monte-neo discover (verified indicator search)

v0.49.0 — stage 7: CI in one command¶

[x] monte-neo init-ci

v0.48.0 — stage 7: real frameworks¶

[x] Adapters for Backtrader, backtesting.py, bt, Nautilus; a CI job runs the real frameworks; close fills read at their own bar (a bug fix)

v0.47.0 — stage 7: first run and real use¶

[x] verify --demo (no files) and a plain-language summary under the table
[x] Adapters checked against real vectorbt, Backtrader, backtesting.py, bt, Nautilus Trader and Zipline (v0.48.0); Freqtrade and Lean adapters still only checked on fake data
[x] One-line CI setup: monte-neo init-ci writes the workflow (v0.49.0)

v0.46.0 — capacity for universes¶

[x] Capacity from volume for multi-symbol universes: pooled fills, tightest symbols listed, "Capacity by symbol" table in the report

Next (v0.46+)¶

[ ] First public bench run (owner runs it; tooling ready in v0.32.0): Claude Code / Codex / Gemini CLI / Cursor → published leaderboard
[~] Publish to MCP catalogues: official registry (done), glama / smithery / mcp.so / mcpmarket (owner submits; texts in docs/marketing/launch-kit.md)
[x] Signed certificates (v0.25.0), signing in the GitHub Action (v0.27.0), public verification page (v0.29.0)

v0.17.7¶

[x] Pages homepage: docs/index.md at site root; docs/docs-map.md replaces INDEX.md
[x] Native index.md site root (sed hack inert); prune old github-pages deployments after ship
[ ] LocalScorer B — still waiting on real labels

v0.17.6¶

[x] CI PyPI install smoke (pypi-smoke.yml + scripts/verify_pypi_install.sh)
[x] SECURITY.md linked from README / CONTRIBUTING / docs INDEX / nav
[ ] LocalScorer B — still waiting on real labels

v0.17.5¶

[x] Research device="auto" prefers cpu_numba for wall clock (auto_prefer_cpu_numba)
[x] Explicit metal/mlx unchanged; OMS resolve_device unchanged
[ ] LocalScorer B — still waiting on real labels

v0.17.4¶

[x] Performance page: M1 Pro 16GB primary table (honest Metal vs cpu_numba wall times)
[ ] LocalScorer B — still waiting on real labels

v0.17.3¶

[x] Docs trust pack: job-fit scenarios, export-first quickstart, honest performance page
[x] scripts/bench_research_bar.py first-party research-bar timings
[ ] LocalScorer B — still waiting on real labels

v0.17.2¶

[x] Soften holdout promote gate (holdout_positive) + label JSONL log
[ ] LocalScorer B — still waiting on real labels

v0.17.1¶

[x] Install audit (pipx, TestPyPI recipe, FAQ)
[x] CONTRIBUTING + issue/PR templates

v0.17.0¶

[x] Holdout / walk-forward helper around export (holdout_sma_sweep)
[ ] LocalScorer B — only if A proves useful
[x] Install audit / CONTRIBUTING (tier 2) — v0.17.1

v0.16.0¶

[x] Slim default deps + optional extras (apple darwin markers, plot, data, ml, server, full)
[x] HeuristicPolicy A: ResearchState + triage + CLI --policy-triage
[x] Export UX docs (API + quickstart extras)
[x] Holdout helper (tier 2) — shipped in v0.17.0
[ ] LocalScorer B — only if A proves useful

v0.15.1¶

[x] Packaging/CI fixes after v0.15.0 tag (bare install, data wheel, ruff/CI)
[x] First TestPyPI + PyPI publish (0.15.1)

v0.15.0¶

[x] PyPI-ready packaging prep ([apple] extras, metadata, PACKAGING.md)
[x] Demo screenshots + examples quickstart
[ ] TestPyPI dry-run + first PyPI publish (maintainer OK)

v0.14.1¶

[x] No-hang Metal/MLX size gate → cpu_numba fallback
[x] README marketing pass + FAQ

Project rules¶

[x] Silence / Evidence permission-gate / version sync / Swiss-watch quality

Version 0.14.0 - 16GB packing + export depth + positioning (Current)¶

[x] plan_research_bytes soft budget / tile hints
[x] Export equity stride + journal + memory block
[x] README job / non-goals without rival names

Version 0.13.0 - Research signal factory¶

[x] Numba-parallel SMA cross signal grids (build_sma_cross_grid)
[x] Sweep/export split signal_elapsed_s / economics_elapsed_s
[x] Unit parity vs single-signal path

Version 0.12.0 - Research export API + golden vectors¶

[x] monte_neo.backtest.export stable schema (API v1)
[x] Frozen golden vectors + verify_golden_vectors / verify_export_golden
[x] Timers disclose includes_signal_build
[x] Unit coverage for export schema + Numba golden exactness

Version 0.11.0 - Metal funding/session/long_short¶

[x] Metal research economics: funding + session mask + long_short
[x] MLX unit skips cleared against current engine API

Version 0.10.0 - Metal SL/TP/trail research subset¶

[x] Metal long/flat batch with SL/TP/trail vs Numba parity
[x] Funding / session / long_short remain Numba golden

Version 0.9.0 - Research bar Metal economics¶

[x] Metal long/flat batch economics with Numba golden parity
[x] device= on batch / SMA sweep; advanced knobs stay Numba

Version 0.8.0 - MC Metal/MLX unify + 16GB budgets¶

[x] plan_mc_run Metal-first / MLX / CPU dispatch
[x] Scenario memory budgets + tile helpers for M1 Pro 16GB

Version 0.7.0 - Metal L2 walk dispatch¶

[x] PyObjC Metal L2 book-walk with Numba parity

Version 0.6.0 - Metal OMS batch + compute pref¶

[x] PyObjC Metal batch long/flat with Numba parity

Version 0.5.0 - Venue adapters¶

[x] Paper / Binance / Bybit adapters; live env-gated dry-run

Version 0.4.0 - Tick / L2 paper OMS¶

[x] Tick feed + L2 walk + Metal L2 shader scaffold

Version 0.3.0 - Paper OMS + accel foundation¶

[x] Bar OMS + device select + Metal bar-match scaffold

Next¶

[ ] Private P002/P003 (separate harness) — no Evidence without permission
[ ] Optional OMS hardening (not research speed claim)
[x] Metal SL/TP/trail research subset
[x] Release cut when green on main (v0.10.0 / v0.14.0 tagged)

Local agents¶

[x] qwen-code built-in agent for bench run, --base-url for models served locally (Ollama, LM Studio), failure hints
[x] integrations/qwen/ rules and MCP entry; guide local-agents.md
[ ] A checked end-to-end run with a real local model on the maintainer's machine, and further local agents as they appear