Changelog

All notable releases are documented here. Version source of truth: src/monte_neo/_version.py.

[v0.54.0] — 2026-10-20

Phantom profit in pairs and spread strategies: the verifier can now find a profit that the legs never earned, and fix it.

Added

• verify --legs FILE (MCP verify_legs): re-derives a spread, pairs or stat-arb strategy's P&L from the quantity held in each leg (date, px_<leg>, qty_<leg>, reported_return) and compares it with the reported return. The new integrity check accounting_mismatch fails (verdict REJECT) when more than 5 % of the P&L is not what the legs earned (--legs-tol); derived_significance warns when the legs earn too little to tell from luck. --legs-fixed-out writes the corrected returns, --legs-pnl simple reads quantities as fractions of equity, and an optional cost column is added back. This is the check for "phantom mean reversion": a rolling hedge ratio booked as position times the change of the spread, whose profit comes from re-fitting the hedge.
• verify --null-test PATH.py[:FUNC]: runs a pairs strategy unchanged on independent random walks of your length and volatility (--null-like, --null-paths, --null-bars, --null-vol, --null-legs, --null-seed). The economics check edge_on_noise fails when more than 75 % of the paths are profitable and the median t-statistic exceeds 1.
• verify --demo-phantom: the same z-score pairs rule booked wrongly and correctly, with both checks, no files needed.
• monte_neo.verify.phantom: the small reference model (rolling OLS, z-score rule, booked and tradable P&L, phantom term).
• paper/phantom-mean-reversion/: sources and reproduction code of the preprint "Phantom Mean Reversion" (the simulation, EWMA and Kalman hedges, a cointegrated pair).
• examples/phantom_pairs/ and docs/guides/phantom-profit.md; integration rules for agents mention verify_legs.

Fixed

• verify --returns (and every certificate that is not a single-strategy backtest) no longer ends in KeyError: 'total_return' after printing the verdict: the summary line prints the metrics the certificate has.

[v0.53.2] — 2026-10-19

Metadata-only patch. No change to any verdict, threshold or behaviour.

Changed

• CITATION.cff names the author (Rostyslav Shcherbyna, with ORCID) instead of the project handle, so that archived copies and citations carry the author's name.

[v0.53.1] — 2026-10-13

Wording and demo fixes found after the v0.53.0 release. No change to any verdict or threshold.

Changed

• monte-neo alpha --demo uses a demo capital of 200,000 (override with --aum), so the tradability checks run and the two demo scores no longer both end at ARL 2. The first demo score is called "a planted edge" and the summary says it is idealised (40 independent names, a daily rebalance): real alpha is far weaker.
• repair: the REPAIRED headline now says what is measured: the lockbox still earns after costs and is not rejected. It can still ask for more evidence, so read its failed checks.
• min_backtest_length no longer prints "about 0.0 needed": it says "less than a month", and for a Sharpe that is not positive it says no history separates it from luck. haircut_sharpe writes "1 trial". universe_subsampling no longer prints "n/a" for the profit share when there is no gross profit.
• README: Jupyter %%verify, Polars input, dossier --lite and the two ML lint rules are listed; stale version examples in the docs and the Docker pin text point to the current release.

[v0.53.0] — 2026-10-12

From a rejected backtest to a committee. monte-neo repair fixes a strategy's defects without overfitting it, and says so when the profit was the defect. monte-neo alpha asks whether a signal is real alpha and tradable, and gives an Alpha Readiness Level. monte-neo dossier turns a certificate into evidence a model-risk committee can file, and monte-neo monitor watches the strategy after launch. Machine-learning models are refitted only on the past, LLM signals are checked against their training cutoff, events are joined by publication time, and track records can be screened from returns alone. Every report says it is not investment advice.

Agents: tested with Claude Code and Qwen Code (local); reports for other agents are welcome in issue #114.

Added

• monte-neo repair and MCP tool repair_strategy (repair-report/1): repairs a strategy's defects without overfitting it. Look-ahead and repainting are corrected first (causal rewrites, a one-bar delay); then a few declared repairs (smoothing, holding period, no-trade band, tranches, neutralisation, caps, liquidity filter, volatility target) are chosen on the first 70 % of the history by the number of defects, never by return, each counted as a trial; the last 30 % is a lockbox opened once. Outcomes REPAIRED, REPAIRED_NO_EDGE (the profit was in the defect), NOT_REPAIRABLE (with what to do) and NOTHING_TO_REPAIR. The original file is never changed; the repaired file is self-contained. Zoo A16: no leaking trap of the Trap Suite is ever reported as REPAIRED on a random walk. See guides/repair.md.
• monte-neo alpha and MCP tool validate_alpha (alpha-report/1): does a cross-sectional score predict returns, and can its long-short portfolio be traded? Family A: rank IC with a Newey-West t, IC decay and half-life, quantile monotonicity, long and short legs, breadth against a claimed IR, permuted-score placebos, selection against timing. Family D: square-root market impact and participation at --aum, capacity curve, short feasibility and borrow, holding period against signal decay, dependence on the rebalance day. See guides/alpha.md.
• --policy and monte-neo policy init | show: an institution's thresholds in a TOML file (minimum trades, cost multiplier, no warnings, minimum Deflated Sharpe and Sharpe, maximum drawdown, required checks, warnings treated as breaches). The policy row lists every breach; the policy and its hash go into the certificate and --recheck.
• --trials-from: the number of variants already tried, read offline from MLflow (file store or SQLite), Optuna, a W&B CSV export, a CSV or a ledger; the largest count deflates the Sharpe.
• --model-cutoff (check model_cutoff, look-ahead): the edge of a strategy that uses a language or foundation model must survive the date its training data ends.
• Team ledger: --ledger PATH on a shared drive locks every write and records who wrote the entry and from which host.
• The "not investment advice" notice is one text (monte_neo.disclaimer) in every certificate, at the bottom of every HTML report (12 px, readable, printed), in every MCP answer (disclaimer), at the end of CLI output, in the README and in the docs footer. Agent rules ask agents to repeat it.

• Agent compatibility table in guides/agents.md: tested with Claude Code and Qwen Code (local); Codex, Gemini CLI, Cursor and other clients are not tested by the maintainer yet (issue #114). An issue form for agent test reports. Agent rules mention trials_from, model_cutoff, policy_path and validate_alpha.

• fit / predict protocol for machine-learning strategies: a strategy file may define fit(train) and predict(model, df) (with REFIT_EVERY, MIN_TRAIN, LABEL_HORIZON, EMBARGO) instead of signal(df). The verifier runs the walk-forward, refitting only on rows that end before the label horizon and the embargo; the probes test predict. Check ml_refit_gap (look-ahead) compares it with the whole-file fit; ml_cv_overlap finds training rows of cv_splits(df) that reach into a test fold. See guides/ml-and-events.md.

• --events (--events-delay, MCP events_path): news, earnings, alternative data or LLM outputs joined to the bars by publication time as ev_<column> columns; checks events_quality and pit_events (what a join by event time would earn). The events' hash is in the certificate and --recheck --events reproduces it.
• verify --returns (--benchmark, --n-strategies, --demo-returns, MCP verify_returns): due diligence on a track record given only as returns: smoothing, missing small losses, implausible consistency, round numbers, Deflated Sharpe with the manager's number of strategies, drawdown depth, benchmark relation and claims. See guides/returns-due-diligence.md. verify --returns-dir DIR screens every manager in a folder in one table (--out, --html).

• monte-neo alpha, families B, C, F and I on the score's long-short portfolio: attribution to factors built from the universe (or --factors), residual alpha against the t ≥ 3 hurdle, hidden short volatility, timing beta, crowding; value to an existing book (--book: correlation, marginal Sharpe, residual alpha, drawdown overlap); the Harvey-Liu-Zhu hurdle, minimum backtest length, drawdown plausibility and Sharpe haircut (--n-trials); neutrality, concentration, leverage, named crises and tail risk. --signal-only keeps families A and D.

• monte-neo dossier and MCP tool build_dossier (validation-dossier/1): a certificate turned into one folder for a model-risk committee, a validator or an auditor: a printable HTML page (verdict, purpose and scope, checks, readiness level, a clause map to SR 11-7, PRA SS1/23 or MiFID II RTS 6, assumptions and limitations, reproduction, blank sign-off fields), the certificate, the environment, REPRODUCE.md and limitations.md. The map shows evidence, never compliance; clauses that need a person are never shown as fully covered. See guides/dossier-readiness-monitor.md.

• monte-neo dossier --lite writes a one-page plain-language version for a non-quant reader (the full notice in a box at the top); frameworks crypto (24/7 data, funding, venue latency, delisted tokens, crash stress; custody and counterparty risk need a person) and academic (pre-registration, trials, out-of-sample, the t ≥ 3 hurdle, reproducibility; data and code availability need a person).
• Alpha Readiness Level (ARL 1-9) in every alpha report and dossier: the highest level met together with all levels below it; missing evidence is "not assessed", never passed, and the report names the run that would assess it.
• monte-neo monitor and MCP tool monitor_live (live-monitor/1): live or paper returns against the backtest, realised slippage from fills against the modelled one, CUSUM and SPRT alarms against the backtest edge after a haircut, and the live history still needed. Exit code 2 on a warning, for cron and CI.
• monte-neo alpha, families E and G: robustness (70 % universe subsets and the top-5 share, groups, start dates, combinatorial purged paths, bootstrapped markets without the score's link, noise in the score, a second data source, bar grid) and data and process (today's index members, classification look-ahead, closes of different exchanges, decay after a discovery date, research hygiene). Options --membership, --groups, --exchanges, --discovery-date.
• signal_orthogonality (alpha --existing-scores): the partial IC of a new score after the signals a fund already runs. romano_wolf (info) in grid searches: which parameter combinations stay significant after the Romano-Wolf stepdown. ml_seed_stability (SEEDS = k and fit(train, seed)) and ml_feature_drift (features(df), population stability index) for the fit / predict protocol.
• Zoo A15 (alpha): 13 traps with honest twins (repackaged momentum, short volatility, three symbols, first day of the month, illiquid shorts, today's members, same-date New York closes, decay after discovery, noise, a book clone, exact score values, an inflated IR claim) and scripts/alpha_calibration.py, which measures false alarms and power with Wilson intervals (guides/alpha-calibration.md). Lite in CI, 20 seeds nightly.
• Jupyter cell magic %%verify --ohlcv data.csv (%load_ext monte_neo.verify.magic): the cell is the strategy, the answer is the HTML report. Polars tables are accepted wherever a pandas table is. --trials-from raytune:DIR counts Ray Tune trials. Lint rules full_sample_transform (a scaler, PCA or feature selector fitted on all rows) and target_encoding_full (a target mean over all rows), both warnings. Examples ml_fit_predict.py, llm_news_signal.py and allocator_dd.py.
• Every HTML page (report, dossier, dossier lite, repair report, batch due diligence) carries a Content-Security-Policy and no scripts; a test checks all renderers.
• Demos: docs/assets/demo-repair.gif, demo-alpha.gif, demo-dossier.gif and two example institutional reports in the gallery.
• Reference alpha 1 (examples/reference_alpha/): diversified trend following with volatility targeting, rules fixed in advance, when it works and when it fails, how to retrain without overfitting, and the verdicts measured on synthetic trending, random and choppy markets. See guides/reference-alpha.md.

Changed

• The README badge says "tested with Claude Code · Qwen Code (local)" instead of naming agents nobody has tested.
• The certificate's disclaimer text is longer (it is not part of the certificate id, so existing certificates recheck).
• Alpha families B, C, E, F and I judge the score's portfolio before costs (is the signal real?); family D judges costs and size.
• From the zoo calibration: residual_alpha fails a return that is exactly a mix of factors (R² ≥ 0.999) instead of reading a meaningless t-statistic; rebalance_timing_luck warns only when the spread between rebalance days exceeds twice the median result and two standard errors (15 % false alarms before, none on the twins now); the top-5 profit share of universe_subsampling warns above 70 % (symbol noise alone gives about half on an honest edge).
• Panel.periods_per_year is unit-safe (intraday panels were annualised as daily ones).
• Known limit narrowed: a model fitted when the module loads is still not caught by the probes, but a model written with the fit / predict protocol is fitted by the verifier, so its whole-file fit cannot hide.

[v0.52.0] — 2026-10-05

Added

• bench run --clean-env starts the agents without ANTHROPIC_AUTH_TOKEN, ANTHROPIC_API_KEY, ANTHROPIC_BASE_URL and ANTHROPIC_MODEL, so they use their own login; the Claude Code message "issue with the selected model" now gets a one-line fix. A troubleshooting section in guides/agents.md explains the variables.
• Local agents: qwen-code is a built-in bench run agent; bench run --base-url points an agent at a model on your machine (Ollama, LM Studio) and checks the server before any task; known failure messages get a one-line fix, and an agent that failed is not run on its remaining tasks. integrations/qwen/ (rules and MCP entry). See guides/local-agents.md.
• monte-neo discover and MCP tool discover_indicator: a search over causal formulas with a causality gate, canary leaks, effective-trial counting, a search null on shuffled markets, Reality Check / SPA / PBO, a lockbox opened once and a certificate; optional evolution (--evolve) and --recheck of a recorded search. See guides/discover.md and the generated calibration.
• Repaint checks repaint_history and repaint_live (--repaint off|auto|strict, --signal-timing close|open): a signal must not change after it was shown, and the forming bar is tested at 0, 25, 50 and 75 % of its path. Lint rules repaint_zigzag, unconfirmed_pivot, htf_without_shift, lookahead_on; suggest_fix shifts a higher-timeframe aggregate. See guides/repaint.md.
• Trial ledger (verify --ledger, MCP ledger): counts the variants tried on the same data; n_trials is the larger of the count and the declared value.
• verify --suggest-fix FILE and MCP tool suggest_fix: causal rewrites with a diff.
• Project tools (guide): history (certificate history, diff, regression exit code, pull-request table), register (pre-registration, --registration), oracle (Thresholdout hold-out oracle), portfolio (effective number of strategies, SPA over the list), doctor (known problems of data exports). MCP tools compare_certificates, register_hypothesis, holdout_query, verify_portfolio, diagnose_data.
• serial_correlation (info): the Sharpe corrected for serial correlation (Lo, 2002).
• Data checks: 20 or more identical bars in a row, and a market that never closes skipping bars (1 % of the steps); quote checks for repeated rows and a stalled feed.
• Hypothesis zoo (tests/zoo): about 170 mechanisms in six disguises (1026 cases), repaint (A14), data defects (21 x 3 frequencies), economics (A5), statistics (A6), universes (A10), machine learning (A11), quotes (A12), planted claims and metamorphic relations, an independent oracle, a ratchet of known gaps (now empty), a generated honesty scorecard (scripts/zoo_scorecard.py). The lite level runs in CI, the full one nightly.
• Agent rules (integrations/) mention the ledger, suggest_fix, repainting, the hold-out oracle and discover_indicator.

Changed

• A stop that a bar gaps through leaves at the open, not at the stop level (long: the lower of stop and open; short: the higher). Before, a gap of 12 % through a 5 % stop was booked as a 5 % loss. Take-profit keeps its level. Results of strategies with --sl-pct / --trail-pct can get worse; certificates of such strategies change.
• Lint: an assigned-and-never-read future value, and a private helper nothing calls, are warnings instead of failures; uniform_filter1d with a trailing origin is causal.
• Ledger and registry stamps have microsecond resolution.

Known limits

• Universes skip data_independence and the repaint probes. The Freqtrade and Lean adapters are tested on fake data only. A model fitted when the module loads that still reacts to its input is not caught by the probes. Hansen's SPA is a little liberal in small samples (about 7 % at a nominal 5 %).

[v0.51.0] — 2026-10-04

Added

• data_independence check. The strategy is run on a random walk of the same volatility, then on that walk with every date shifted. A strategy whose positions do not change on either one ignores its input: a cache between calls, an array computed in advance, a table keyed by date. It is a warning, and a failure (REJECT) when the strategy also earns an annualized Sharpe of 6 or more. Calendar rules (hour of day, weekday) are recognised and pass.
• implausible_performance check: a warning when the annualized Sharpe after costs is 6 or more.
• Lint rules cached_signal (lru_cache, cache, cached_property), import_time_fit (.fit() when the module loads) and global_state (global), all warnings.

Fixed

• A leaking strategy that kept its answer between calls passed as PASS. A strategy that computed the leaky answer once on the full table and returned prefixes (or held an array computed in advance) beat the truncation and perturbation probes, and nothing else looked at its Sharpe of 57. Found by measuring the plan of stages 8-10; the three variants are now rejected (tests/unit/test_verify_stateful.py). Model fits on the whole file that still react to their input remain a limit of any verifier that runs your code: see "What the probes cannot prove" in the API guide.
• Cost: one probe adds two extra runs of signal() (0.4 s on 50,000 bars).

[v0.50.0] — 2026-10-01

Latency audit (experimental). The new --quotes checks are context: they warn, they do not fail the verdict, and they were calibrated on synthetic data and checked on 1.18 million real quotes with an assumed latency. Read the guide before relying on a number.

Added

• Quotes with arrival time (monte_neo.verify.quotes): load top-of-book quotes with their latency, check them (quote_quality: crossed and non-positive quotes, negative latency, out-of-order arrivals, latency percentiles per venue, share of stale quotes, rows that did not parse) and build mid-price bars on the exchange clock or the arrival clock.
• Arrival look-ahead (monte_neo.verify.arrival.arrival_lookahead): the strategy sees bars binned by stamp + latency and trades against the market bars; warn when its profit exists only with zero latency.
• latency_scan: profit when data arrives later, and the delay at which it vanishes.
• latency_monte_carlo: the same strategy over 200 seeded redraws of the observed latency; distribution of the return and the probability of a loss.
• verify_quotes, monte-neo verify --quotes (--bar-ms, --latency-samples, --demo-quotes) and the MCP tool verify_quotes: one strategy-verdict/1 certificate from the four checks plus net_profitability on the exchange clock; the HTML report gets a "Time and latency" section with a return-against-delay chart.
• Traps and honest controls for the arrival checks (tests/traps/test_quote_traps.py): three traps draw their warning, seven honest strategies (four seeds, three latencies) pass with no warning.
• The arrival checks are warn at most; a strategy that loses money on the exchange clock is rejected, as in verify_strategy.
• Quote recorder (python -m monte_neo.data.quote_recorder): records Binance USD-M futures bookTicker quotes with their latency (receive time corrected by the clock offset to the exchange, minus the event time) into the table verify --quotes reads. It reports how well the offset is known and warns when the latencies cannot be trusted (proxy, VPN, slow network); a dropped connection keeps what was recorded.
• verify_quotes says plainly when a recording has too few bars for the warm-up.
• --symbol picks one instrument from a table of several (an unclear table is an error that lists the symbols); --order-latency-ms delays every fill by a constant, so an edge that lives in instant execution is rejected.
• --recheck for quote certificates: monte-neo verify --recheck CERT --quotes Q --strategy S, recheck_quotes() and the MCP tool recheck_certificate(quotes_path=...) reproduce the certificate id from the quotes, the code and the recorded settings; changed quotes or code are named in inputs_match.
• --latency-model (constant:MS or lognormal:MEDIAN,P95) assumes the data latency instead of reading it (the table needs no latency column); the certificate records it and says the latency is assumed, not measured.
• --symbol + --feeds: a strategy may read other instruments (<alias>_close...), each feed with its own latency; synthetic_feeds is the latency-arbitrage test pair. A leader that arrives after the follower moved turns the edge into a loss.
• Look-ahead probes on the arrival bars (--no-probes skips): a strategy that reads the next bar is rejected, which the clock check alone cannot see.
• spread_cost warns when the modeled slippage is below the half-spread a taker pays; every certificate carries assumptions (taker fills at the mid, no queue position, no partial fills, order delay, latency source), shown in the report.
• latency_tolerance calls the profit vanished only when two scanned delays in a row earn nothing: one noisy point of a thin honest edge no longer warns.
• Checked on real quotes (scripts/validate_on_real_quotes.py, 1.18 million real quotes of BTC, ETH and SOL futures, assumed latency): 36 honest runs, no false accusation (33 lose money after costs, 3 pass); the foresight control is rejected 3 of 3; an edge planted on the real SOL path (a follower that repeats it 2 s later) is caught at the matching delay. Almost no honest strategy earns on real prices, so the false-warning rate among earning strategies is still unmeasured.
• A guide (Latency audit), a second demo GIF (docs/assets/demo-arrival.gif, rendered from live runs by scripts/make_arrival_gif.py) and the README section that explains what the feature catches and what sets it apart (an audit of a finished strategy, not an environment).
• Measured speed on 1 million quotes: load 0.6 s, verify_quotes with 200 latency draws on 100 ms bars 7 s (see docs/api/verify.md).
• quote_quality warns about bursty latency (p99 more than 20 times the median, from 100 quotes): a congested path (VPN, Wi-Fi) holds messages back and releases them in bursts, and the arrival checks would then measure the network. Found on a real two-minute recording: latency climbed from 0.5 s to 12 s and drained in one burst.
• The recorder measures the clock offset over one keep-alive connection (a new connection per sample inflated the round trip about three times), and verify --quotes takes costs and warm-up into verify_quotes.

Security

• virtualenv 20.36.1 -> 21.14.2 in uv.lock (four Dependabot advisories; a development tool pulled in by pre-commit, not a runtime dependency).
• The release SBOM is built from scripts/requirements/runtime.txt, a hash-locked file made from uv.lock (a test keeps it in step), and the package itself installed with --no-deps; the verifier image installs the release with --require-hashes (hashes read from PyPI by scripts/build_docker_image.sh). This closes the two remaining Pinned-Dependencies findings.

[v0.49.0] — 2026-09-30

Added

• monte-neo init-ci: writes a GitHub Actions workflow that verifies your strategy on every pull request. It finds the strategy file and the price table in the project (or takes --strategy / --ohlcv), pins the checkout to a commit and the action to the installed release, posts the verdict as a pull-request comment and uploads the certificate. Options: --n-trials, --fail-on, --isolate, --sign, --out, --print, --force. It never overwrites a file without --force.

[v0.48.0] — 2026-09-30

Added

• Adapters for more real frameworks: from_backtrader (the Transactions analyzer), from_backtesting_py (stats["_trades"]), from_bt (security weights) and from_nautilus (the engine or its fills report). A CI job now runs a moving-average strategy in each installed framework (Backtrader, backtesting.py, bt, vectorbt, Nautilus Trader, Zipline) and compares what the adapter reads with the position the framework itself held on every bar. The framework versions are hash-locked (scripts/requirements/frameworks.txt).
• from_fills(..., fill_at="open" | "close"): say where in its bar a fill happens.

Fixed

• Close fills were read one bar too early. from_fills treated every fill as an open fill, so a framework that fills at the close of the decision bar (Nautilus on bar data) gave the strategy one bar of foresight: its results looked better than they were. from_nautilus reads such fills at their own bar, and fill_at covers the rest.
• from_zipline read a daily transaction, stamped with the session close time, one bar late; it now uses the session date.
• An empty list of fills (a strategy that never traded) raised a confusing error; it now means flat all the way.

[v0.47.0] — 2026-09-30

Added

• monte-neo verify --demo: a first run with no files. It verifies a strategy that peeks at tomorrow's close and a no-peeking moving-average crossover on synthetic prices, and says for each whether the look-ahead checks are clean.
• Plain-language summary under the check table of every text run: what the verdict means, the first reason and how many more there are. The certificate and its id are unchanged.

[v0.46.0] — 2026-09-30

Added

• Capacity for universes. The capacity row now works for multi-symbol runs: each order (capital x weight change) is set against the traded value of that symbol at the fill bar, and the capital at which 90% of the fills stay within 1%, 5% and 10% of it is reported. The row names the tightest symbols, and the HTML report gets a "Capacity by symbol" table. Fills at bars without volume are counted, not used. Single-instrument capacity and all certificate ids are unchanged.

[v0.45.0] — 2026-09-30

Added

• Per-symbol costs for universes (--costs-file, MCP symbol_costs, Action costs-file, Python symbol_costs=): a JSON mapping gives each symbol its own commission and slippage in bps, with a default row; a symbol takes its own values, then default, then the uniform costs. A name that is not in the data is an error. The resolved table is written into the certificate's cost model (one row per symbol), so a recheck needs no other input; a run without per-symbol costs keeps its certificate id. The target-weight engine charges each instrument its own costs; a table of equal rows equals the uniform model bit for bit, and an independent reference covers different costs. The cost checks and the report use the average of the rows. The engines that would ignore per-symbol costs (single instrument, batch, shared-cash) refuse them.

[v0.44.0] — 2026-09-30

Added

• Stops inside the bar for target weights. sl_pct, tp_pct and trail_pct now work in the target-weight engine (fractional weights and universes), using each instrument's high and low. With weights in {-1, 0, 1} on one instrument the result equals the discrete engine bit for bit, stops included (tested on hundreds of random configurations with costs, funding, borrow fees, leverage and both fill policies), and an independent reference covers fractional weights over several instruments. Models without stops give bit-identical results to before, and the pure-Python fallback gives the same numbers.
• Stops in the verifier: --sl-pct, --tp-pct, --trail-pct in the CLI, sl_pct / tp_pct / trail_pct in the MCP tools and model_from_costs, and the same inputs in the GitHub Action. They are part of the certificate's cost model and appear on the report's cost line. The buy-and-hold benchmark never uses them.

Changed

• run_weight_backtest takes optional high and low; a model with stops and no high and low is an error (it used to refuse stops for weights altogether).

[v0.43.2] — 2026-09-30

Security

Findings from a full security pass (Bandit over all of src, pip-audit over the whole lock file including dev dependencies, secret scanning of the tree and history, review of every workflow). None touched the verifier's own code.

• Research modules: the calibration cache no longer reads or writes pickle files (loading a pickle runs code); its directory is created private (700); the cache key hash is marked as not security-relevant.
• Dynamic indicators: expressions are checked (no private names or attributes) and run with a whitelist of built-ins instead of the full set.
• The generated compile.sh is executable by its owner only.
• pypi-smoke.yml passes the resolved version through the environment instead of splicing it into the script.
• Bandit in CI now covers all of src.

[v0.43.1] — 2026-09-30

Quality release after a full check of the product as a customer (clean install from PyPI on Python 3.11, 3.12 and 3.13, every command, the MCP server on SDK 1.x and 2.x, the Action script, the Docker image, real vectorbt and Backtrader runs, an independent re-implementation of both engines and of the Sharpe statistics, 70+ edge inputs).

Fixed

• A strategy split over several files now works. A helper module, a package or a parameter file next to the strategy file failed with No module named ... in every mode. The strategy's folder is now on the import path (at the end, so it can never shadow numpy or the standard library).
• monte-neo --help (and -h, help) print the commands instead of failing without the research extra; an unknown command says so.
• A command-line usage error (--n-trials abc) exits with 3, not 2: 2 means REJECT and must not appear for a typo.
• from monte_neo.verify import * exposes every public name (show, from_vectorbt, ...); a test keeps __all__ complete.
• A headerless signals CSV keeps its first value (it used to lose it: "signal length 2999 != bar count 3000").
• CSV files separated by ;, tab or | are read.
• The HTML report no longer prints None for a signals-only run, names the right recheck option, and mentions funding and borrow fees when the model has them. The example gallery is regenerated with the current engine.
• The static lint no longer warns on model.fit(X[train], y[train]) (a fit on a chosen slice, as in walk-forward code); a fit on a whole array still warns.
• Certificate prints one summary line instead of a 50 KB dict; a grid whose values are not lists says so.
• The pre-commit guide shows how to limit the hook to strategy files.

[v0.43.0] — 2026-09-30

Added

• Short-borrow and funding fees. --funding-bps-per-bar (every open position) and the new --borrow-bps-per-bar (short positions only), in the CLI, MCP tools (funding_bps_per_bar, borrow_bps_per_bar), the GitHub Action and model_from_costs. Both engines (sign and target-weight) charge them; the batch and shared-cash engines refuse a model that sets a borrow fee instead of ignoring it. The fee is part of the certificate's cost model only when set, so existing certificates keep their ids and recheck.
• Reality Check and SPA for verify_grid: the new reality_check row (White 2000, Hansen 2005) tests whether the best combination beats cash by more than the whole search explains; it warns at an SPA p-value of 0.10 or more.
• Rolling stability (walk_forward_stability, context): the returns cut into 6 equal windows, how many made money and the worst and best window. metrics.windows_positive.

[v0.42.0] — 2026-09-30

Added

• Technical note on the verification method (docs/paper/methodology.md, also on the docs site): probes, costs, selection, certificates and the Trap Suite evaluation.
• Trap of the week: twelve post drafts with measured verdicts (docs/marketing/trap-of-the-week.md).
• README: "How it differs from other tools" and a link to the note; issue form for false accusations.
• conda-forge recipe draft (packaging/conda-forge/meta.yaml), list and catalogue entries in the launch kit.

Fixed

• The Docker image on GHCR is built by the docker job of publish.yml (after the PyPI smoke test), through scripts/build_docker_image.sh, which waits for the package index. docker.yml is a manual rebuild. In v0.41.0 the separate workflow_run build never started and the first manual run hit the index delay.

[v0.41.0] — 2026-09-29

Added

• Framework adapters. from_vectorbt, from_freqtrade, from_lean, from_zipline and from_fills turn a backtest run elsewhere into positions on the price table; .verify() runs the verifier on them. Backtrader and Nautilus use from_fills (recipe in the guide). See docs/guides/frameworks.md.
• Jupyter display. verify_strategy and verify_grid return a Certificate (a dict subclass) that draws the HTML report in a notebook cell inside a sandboxed frame; show(cert) does the same for a loaded certificate.
• pandas accessor df.monte_neo.verify(...) (import monte_neo.verify.accessor).
• Badge. --badge PATH writes a shields.io endpoint file with the verdict and certificate id.
• pre-commit hook monte-neo-lint and monte-neo verify --lint FILE... (static lint only, exit 1 on a fail-level finding).
• Docker image on GHCR. .github/workflows/docker.yml builds and pushes ghcr.io/neozork/monte-neo-verify after each successful Publish run (plain docker CLI, no third-party action).

[v0.40.0] — 2026-09-29

Added

• Trap Suite: 80 traps and 35 honest controls (was 55 and 25). New traps cover machine-learning pipelines (shuffled and K-fold splits, whole-sample PCA, k-means, k-NN, feature selection, hyper-parameter tuning, scaling, winsorizing, volatility targeting, start-date picking), calendar joins (transform("last"), merge_asof, resample(closed="right"), whole-day VWAP), universes (future volatility rank, whole-sample winners, today's membership applied to the past, total-sample normalisation) and honest code on damaged data (unadjusted split, frozen feed, outages). Each trap has an honest twin that must never be flagged.
• Two lint rules: shuffled_split (train_test_split without shuffle=False; fail when shuffle=True) and kfold_split (K-fold and shuffle splitters on a time series; fail for shuffle=True and ShuffleSplit). The lint now has 23 rules.
• Test datasets frozen_feed, unadjusted_split and feed_outages.

[v0.39.0] — 2026-09-29

Added

• Claim check (--claim). Report the numbers you are about to claim (sharpe, total_return, max_drawdown, n_trades, win_rate, profit_factor) and the verifier compares them with the verified ones. A claim better than the verified result by more than a tolerance fails the new claim_consistency check (category claim, verdict NEEDS_MORE_EVIDENCE). CLI --claim, MCP claim on verify_strategy and verify_grid, Action input claim, Python claim=; the claim is recorded in the certificate so --recheck compares the same one.
• Confidence. sharpe_confidence: 95% interval of the Sharpe and of the total return (circular block bootstrap, fixed seed) and the share of resamples with a positive Sharpe; track_record: the Minimum Track Record Length. Also in metrics (sharpe_ci95, return_ci95, min_track_record_bars). Context only.
• PBO for parameter searches. verify_grid computes the probability of backtest overfitting by CSCV over 12,870 train/test splits; the new pbo check warns at 0.5 or more. The report shows the rank distribution.
• Costs and capacity. spread_estimate: a rough Corwin-Schultz spread from high and low next to the modeled slippage (flags an optimistic model); capacity: the capital at which 90% of fills stay within 1%, 5% and 10% of the bar's traded value (single instrument with a volume column). Context only; both show in the report cards and the cost-sensitivity chart.
• Trap Suite: an overclaim test; the false-positive corpus of honest strategies still raises no alarms.

Notes

• Existing certificates and certificate_id values are unchanged when no --claim is given: the new rows are info (never change the verdict) except pbo, which only appears for parameter searches.

[v0.38.0] — 2026-09-29

Added

• The HTML report is now a tear sheet. A one-sentence reason for the verdict, a card per check family, equity on a log scale with the drawdown, rolling Sharpe, a monthly-returns heat map, the distribution of returns, the timing test (200 shifted copies against the real return), net return against trading costs, trade statistics (win rate, profit factor, average win and loss, holding time, streaks), the evidence for a leak (the bars where the signal changed and the flagged source lines with the code), a parameter heat map for grid searches, and a print stylesheet (PDF from the browser). Still one file: no scripts, no network.
• charts section in the certificate (about 10 KB) with the data behind the charts. It is derived from the hashed inputs and is not part of certificate_id, so existing certificates and IDs are unchanged. Grid certificates also record combo_sharpes.
• Gallery of example reports (docs/gallery.md, scripts/make_report_gallery.py): a leak, no edge, an honest strategy, a parameter search, a universe and bad ticks.

Fixed

• A malformed or hostile certificate (numbers where lists are expected, text in chart data) could crash the HTML renderer; every chart now validates its input.

[v0.37.0] — 2026-09-29

Security

Result of the security audit (policy and threat model: SECURITY.md). - --isolate could be bypassed by renaming a temp file over any file, or by creating a symbolic or hard link in the temp dir and writing through it. Both ends of a rename are checked, links are blocked, paths are resolved with realpath. Reproduced by four new tests. --isolate remains a guard, not a security boundary. - A certificate could make the verifier open a file: the public key inside a certificate was treated as a file name. It is now only text; the key you pass to check against may still be a file. - The verification page showed a green tick for a forger's own key. A certificate signed with any key looked "valid (integrity only)". Now only a match with the expected issuer key is green; otherwise it says the signer is not verified. The page also rejects ambiguous JSON, limits the loaded file (8 MB, 15 s) and sends no referrer. - Strict certificate JSON (no duplicate keys, NaN, depth over 100), size limits for every input (MONTE_NEO_MAX_INPUT_MB for tables), MCP render_report writes only .html / .htm, the HTML report has a Content-Security-Policy. - Dependency click upgraded in the lock file (advisory PYSEC-2026-2132, transitive via mcp). - Supply chain: all GitHub Actions pinned to commit SHAs, read-only default permissions in every workflow, new security.yml (Bandit, pip-audit) and scorecard.yml (OpenSSF Scorecard), CycloneDX SBOM attached to releases.

Added

• The verifier runs without Numba. Where Numba cannot be installed (PyPy, WebAssembly, a Python without Numba wheels, --no-deps), the same engine source runs as plain Python: certificates are identical bit for bit (tested end to end in a process without Numba), only the speed differs (about 50-200x slower for a profitable strategy, see docs/api/verify.md). A warning with the install hint appears from 20 000 bars. Numba is now a CPython-only dependency and has an extra: pip install "monte-neo[fast]". Nothing is downloaded at run time.

[v0.36.1] — 2026-09-29

Fixed

• Slow checks of high-turnover strategies. Every backtest built a Python journal of all trades that the verifier never read: a strategy with 100 000 trades on 200 000 bars took 9.8 s, now 0.5 s (same numbers bit for bit; the verifier uses a journal-free path of the same engine).
• Crash in read-only environments. With no writable directory for Numba's cache (a read-only container, a user without a home directory) importing the engines failed with "cannot cache function". They now compile in memory.

Changed

• Release cadence: the limit of 2 feature releases a day is removed; releases can be published at any time (each still needs the owner's approval).

Added

• Docker image for untrusted code: docker/verify/Dockerfile (engines compiled at build time, runs as nobody), tested with the documented no-network, read-only run.

[v0.36.0] — 2026-09-29

Added

• Data quality check (data_quality). Data can be well formed and still wrong. The new check looks for one-bar price spikes that the next bar undoes (bad ticks), runs of frozen prices, split-like jumps (unadjusted splits), gaps in time beyond the usual nights and weekends, and bars without volume. Findings warn; the check fails when more than half of the profit comes from spikes in instruments the strategy held. Works on one instrument and on universes (reads a volume column when present).
• Trap Suite: bad data. New bad_ticks dataset and spike_fade trap: honest code whose profit is bad ticks must be rejected (55 traps, 25 honest controls).
• Strategy workers. jobs (CLI --jobs, MCP, Action) runs the look-ahead probes of a strategy file in parallel worker processes; timeout (CLI --timeout, default 300 s in MCP, 600 s in the Action) ends a signal() call that hangs with a clear error; isolate (CLI --isolate) blocks network, subprocesses and file writes outside the temp dir and removes secrets from the environment of the workers.
• monte-neo verify --precompile compiles and caches the engines (about 4 s once): the first check in a new environment takes 1.6 s instead of 6 s. For Docker images and CI caches.
• Docs: verifying untrusted code. What --isolate does and does not protect against, and a Docker recipe (no network, read-only file system, resource limits) for marketplaces and prop firms.

Changed

• Lighter install: 488 MB -> 327 MB. The base install is the verifier, the CLI (verify, bench) and the MCP server. Research packages moved to extras: parquet (pyarrow, for .parquet tables) and research (the interactive monte-neo menu). Both print the extra to install when it is missing; the GitHub Action installs parquet by itself for .parquet inputs.
• Tests that need MLX / Metal are skipped on machines without MLX (Linux) instead of failing; they still run on the macOS CI job.

[v0.35.1] — 2026-09-29

Fixed

• Broken bars passed the data check. A bar whose open or close lay outside its high-low range (a close 20% above the high, an open at half the low) was "clean". data_integrity now fails such bars when the gap is over 0.1% of the price; smaller gaps are vendor rounding and are counted in the details.
• MCP errors lost their reason. The SDK answered any exception with "Error executing tool verify_strategy", so an agent could not tell a wrong path from a broken strategy. Tools now return {"error": "FileNotFoundError: ...", "tool": ...}.
• Unhelpful input errors: signal() returning None or one number now says so (it said "signal length 1 != bar count"), text positions say positions must be numbers, and universe rows without a symbol are reported instead of failing inside a sort.
• Slow start. monte-neo loaded the interactive research menu (and pandas) before every command: --version now takes about 40 ms instead of 900 ms and verify --help about 140 ms. New entry point monte_neo.cli.entry; monte_neo.cli loads its attributes lazily.
• Universe perturbation probe: the future of every symbol is mirrored from a per-symbol index built once (same result, about 20% faster on 300 symbols x 1500 bars).

[v0.35.0] — 2026-09-29

Added

• Fractional positions (weights). A signal can now be a fraction of equity in [-1, 1] (0.5 = long half the equity). positions="auto" (API, CLI --positions, MCP, Action) reads weights when every value is in [-1, 1] and one is fractional, else signs; sign and weight force the reading. Weights run on a new target-weight engine (monte_neo.backtest.weight_engine): only changes trade, costs are charged on the traded notional, a change of side closes first. On +1 / 0 / -1 it gives the same equity as the sign engine bit for bit (400 random scenarios, and in CI). Sign strategies keep their exact numbers and certificate ids.
• Universes. A table with a symbol column (ticker / asset / instrument) and a timestamp column is verified as a portfolio: signal(df) returns a weight per (timestamp, symbol) row, gross exposure is capped at 1 per timestamp, symbols may list late or stop trading (open positions close at the last price). The truncation and perturbation probes cut and rewrite every symbol's future at once, so leaks through another symbol are caught. New survivorship check warns when no symbol stops trading (a survivors-only universe). Signals tables are matched by (timestamp, symbol). Row order in the input does not change the certificate. Grid search, re-check, the MCP tools and the Action all work on universes.
• Trap Suite: xs_next_return_rank, xs_market_future_join, xs_symbol_history_rank (54 traps) and the honest controls xs_momentum_rank, xs_equal_weight (25), plus a survivors-only universe test.
• Buy-and-hold benchmark on the same data and costs (equal weight for a universe): a benchmark info row, metrics.benchmark_total_return / benchmark_sharpe_annualized and a benchmark section.
• Breakdown: results per year, quarter or month (or four segments) and per market regime (rising / falling, calm / volatile, set by the trailing market return and volatility). New period_consistency check warns when removing the best period leaves no profit.
• HTML report: --html report.html (and --render CERT --html report.html for an existing certificate) writes one self-contained page: equity against buy-and-hold with drawdown, checks, what to fix, periods, regimes, grid, reproduction hashes, signature status and a link to the verification page. No scripts, nothing loaded from the network, every certificate string escaped. MCP tool render_report; the Action writes it, uploads it with the certificate and exposes it as the html-report output.
• The certificate series (up to 400 equity points) feeds the charts; compact MCP responses drop it.

Fixed

• recheck never reproduced a certificate that stopped at the data check (no signal hash was recorded).
• The lint called a rank or an aggregate grouped by the exact timestamp (groupby(df["timestamp"]).rank(), a cross-section of a universe) a whole-sample statistic. Grouping by a derived date is still flagged.
• probe_lookahead and cost_stress (MCP) built single-instrument arrays from any table; they now use the same market logic as verify_strategy.
• Default warm-up for a universe counts timestamps, not rows.

[v0.34.0] — 2026-09-29

A full audit of the verifier. Each fix below has a regression test.

Added

• timing_significance check (Monte Carlo). A long-biased strategy on rising data could pass every check without any skill: it was just in the market while prices rose. The verifier now shifts the strategy's own positions circularly against the prices by 200 evenly spaced offsets, which keeps exposure, trade count and holding periods and destroys only the timing. If the net return does not beat 95% of the shifted copies, the check warns that the profit comes from market exposure. A planted real edge beats 100% of them (p = 0.005).
• parameter_plateau check for verify_grid. The best combo is compared with its neighbours (one parameter one step away in the grid). If they keep less than half of its Sharpe, the best combo is an isolated peak.
• Five honest controls in the Trap Suite (23 in total): loops over trailing slices, np.polyfit inside rolling().apply, Wilder RSI in a loop, and a strong real edge with a hit rate near 0.6.
• action.yml: fail-on: PASS_WITH_WARNINGS for strict CI; an unknown fail-on value is an input error.

Fixed

• Shorts were ignored in the Python API. verify_strategy and verify_grid without model= used a long-only model, so -1 positions were never traded while exposure still counted them. The CLI and MCP server already defaulted to long/short; the Python API now does too, and exposure counts traded positions.
• Certificates did not record their thresholds. --min-trades 1 could turn NEEDS_MORE_EVIDENCE into PASS and nothing in the certificate showed it; --recheck then reported "verdict differs". The reproducibility block now has settings (min_trades, holdout_fraction, probe_checks, periods_per_year), which is part of the certificate id and is reused by the re-check.
• Re-check across releases said only "certificate id differs". It now says which release issued the certificate and gives the pip install monte-neo==X command to re-check with it.
• The GitHub Action claimed n_trials was declared when it was not: the n-trials input defaulted to 1 and was always passed. It is now empty by default, and the certificate says n_trials not declared.
• Truncation probe missed sparse leaks. It compared only the last bar at 24 evenly spaced points, so a strategy that enters on a few percent of bars could slip through. It now compares the whole prefix and adds checkpoints on the bars where the position changes.
• False look-ahead alarms. A real strong edge (hit rate 0.60 over 539 bars) was rejected by implausible_accuracy; it now needs a hit rate of 0.70 and binomial z ≥ 3.5, and a high but insignificant rate is only a warning. The lint no longer calls c[i - 20:i].max() (a moving window) or code inside a function passed to rolling().apply a whole-sample statistic. A corpus of 25 honest strategy styles on minute, hourly and daily data now raises no look-ahead flag.
• Annualized Sharpe was inflated for markets that close. Bars per year came from the median bar step on a 24/7 calendar: daily stocks got 365 instead of ~252 (Sharpe × 1.2), hourly stocks ~8766 instead of ~1640 (× 2.3). With 30 days of data or more it now counts bars per elapsed year.
• delay_sensitivity said "survives 1 bar execution delay" for strategies that lose money; it is now skipped when there is no profit, like cost_margin.
• The lint row said "static lint: clean" when the strategy did not parse; it now shows the syntax error.
• Outside-data watch gaps: the OHLCV file is recognised under any name (also .txt) for CLI and MCP runs, verify_grid watches the strategy import, and the MCP tool probe_lookahead reports external_data.
• Honesty Bench: a non-numeric claim ("total_return": "none") crashed the whole scoring run; claims are now parsed leniently. Reading data outside df counts as look-ahead on the leaderboard.
• Invalid settings (n_trials < 1, min_trades < 1, holdout_fraction outside (0, 1), folds < 1) are input errors instead of silently changing the result.

[v0.33.0] — 2026-09-29

Added

• external_data check (lookahead). The probes rewrite df and compare signals, so a strategy that ignores df and loads the dataset itself saw the untouched future and passed them: a strategy that read the CSV at import and looked 49 bars ahead got PASS_WITH_WARNINGS. While strategy code runs (import, signal(), probes), a sys.audit hook now records reads of data files (.csv, .parquet, .npy, …, and the OHLCV file itself) and network connections; any of them fails the check and the verdict is REJECT. Library and interpreter files are ignored, so importing scipy, scikit-learn or matplotlib inside signal() is fine. Certificates list file base names only, never local directories.
• Lint rule external_data (rule 21). pd.read_*, np.load / loadtxt / genfromtxt, open(), Path.read_text / read_bytes and network imports (requests, urllib, yfinance, ccxt, …) fail the lint. Code under if __name__ == "__main__": is skipped: agents keep local runs there.
• Trap reads_dataset_file: invisible to both dynamic probes, caught by the new check (51 traps).
• CI job verifier runs the verifier, bench and Trap Suite tests on Ubuntu with Python 3.11, 3.12 and 3.13. The main job ran only on macOS with Python 3.11, while the package is used on Linux and claims 3.11–3.13.

Fixed

• Newest-first data passed as clean. Many exports list bars newest-first. On such data shift(1) reads the next bar, a look-ahead the row-order probes cannot see, and data_integrity still said "OHLCV is clean". It now fails on timestamps that go backwards or repeat, with a hint to sort oldest-first.
• scripts/verify_pypi_install.sh installs the exact wheel from the PyPI JSON API. The release smoke job of v0.32.1 failed because pip's simple index, served from a CDN cache, still listed only 0.32.0 more than 15 minutes after the upload; the published 0.32.1 itself passes every check.

Changed

• Honesty Bench v1 prompt states that signal() must compute positions from df only.
• Agent integrations say the same and require monte-neo[mcp]>=0.35.1.

[v0.32.1] — 2026-09-28

Fixed

• Release smoke test never ran after v0.17.7. GitHub does not start other workflows from a release created with GITHUB_TOKEN, so pypi-smoke.yml stopped firing when releases moved into publish.yml. publish.yml now has a pypi-smoke job that installs the version it just published from PyPI.
• The smoke test checked the wrong things. scripts/verify_pypi_install.sh pinned 0.22.0 by default and only imported research modules. It now installs the given or latest version and runs the product: verify on a leaky strategy (must exit 2), --recheck, --grid, key generation, signing and --check-signature, bench init / prepare / collect and scoring, and the MCP server and tools.
• Agent integrations could stay on an old verifier. The Claude Code, Codex, Gemini CLI and Cursor configs pinned monte-neo[mcp]>=0.18.0, and uvx keeps a cached environment that satisfies the pin. The floor is now >=0.32.0, so existing installs pick up check_signature and the current lint rules.
• The MCP server reported an empty version. It now sends its version and the docs URL to clients.

Added

• A test that every place naming the release (server.json, plugin and extension manifests, CITATION, installation guide, README action pin, CHANGELOG) matches _version.py.

[v0.32.0] — 2026-09-26

Added

• Public Honesty Bench run tooling:
• monte-neo bench prepare <dir> --agents a,b --workspaces <out> creates one clean workspace per agent and task with only data.csv and PROMPT.md. Task names are replaced by random aliases (task-1 …), because names such as costs-trap give the answer away. It writes DATA_SHA256 and refuses workspaces inside the bench directory.
• monte-neo bench collect <dir> --workspaces <out> maps the aliases back and copies strategy.py, claim.json and session transcripts into submissions/.
• scripts/honesty_bench_run.sh runs coding agents headless, one new session per task, keeps a transcript of each session and never re-runs a finished task. Agent commands are overridable templates; it works with macOS bash 3.2 and falls back to gtimeout.
• Guide: "Public run, step by step" in Honesty Bench

Notes

• No breaking change vs 0.31.0.

[v0.31.0] — 2026-09-26

Added

• Trap Suite reaches 50 traps (18 honest controls): describe_threshold, nlargest_dates, agg_zscore, flip_cumsum, resample_ffill_max, np_interp_fill, mode_level, value_counts_level, centered_variable, shift_variable; honest controls rolling_apply_span, hour_open_ref, expanding_max_breakout
• Lint:
• constant propagation: a name assigned a constant exactly once (horizon = -1, centred = True) is checked as that constant, so a negative shift or center=True hidden in a variable is caught; names that are reassigned or are function parameters are left alone;
• describe(), agg(), mode(), value_counts() over the whole series (full_sample_stat, warn) and nlargest() / nsmallest() (full_sample_rank, warn), also on derived series such as close.round(-1).mode();
• np.flip counts as a reversal for reversed_cumulative;
• resample(...).max() and other resample / groupby aggregates, including agg, without a shift are group_aggregate (warn);
• np.interp is reported as interpolate (fail).

Notes

• No breaking API change vs 0.30.0.

[v0.30.0] — 2026-09-26

Added

• Trap Suite: forward_window_indexer, np_sort_rank, cut_auto_bins, reversed_accumulate, tail_threshold, iat_last, builtin_max, and two traps that only the dynamic probes can catch, dataset_fraction and block_mean_reshape; honest controls cut_fixed_bins, hour_running_high and rolling_min_periods (40 traps, 15 honest controls)
• Lint rules (20 in total):
• forward_window (fail): FixedForwardWindowIndexer;
• last_row (fail) now also covers .tail() and .iat[-k];
• reversed_cumulative (fail) now also covers np.<ufunc>.accumulate over a reversed array;
• full_sample_rank (warn) now covers np.sort and pd.cut with a bin count;
• full_sample_stat (warn) now covers the built-ins max / min / sum / sorted over a whole column.
• Trap Suite guide: a section on traps that are invisible to the static lint

Notes

• No breaking API change vs 0.29.0.

[v0.29.0] — 2026-09-26

Added

• Certificate verification page on the docs site: Verify a certificate. It checks the Ed25519 signature of a strategy-verdict/1 certificate in the browser with WebCrypto; nothing is uploaded. ?cert=<https URL>&key=ed25519:<key> loads and checks a certificate from a link, so the "Verified by Monte-Neo" badge can point to a one-click check.
• docs/assets/verify-certificate.js: rebuilds the exact bytes Python signs (sorted keys, Python float formatting, Python string escapes). A test runs it in Node and compares it with canonical_payload on 500+ floats, Unicode and control characters, and checks a Python signature and a tampered certificate.
• The article "Six ways your agent's backtest lies" is published on the docs site (Articles).

Notes

• No API change vs 0.28.0.

[v0.28.0] — 2026-09-26

Added

• Trap Suite: reversed_cummax, reindex_nearest, bars_left_in_hour, hour_size_leak, hourly_close_map, sort_values_rank, plus the honest controls prev_hour_close_map, bars_into_hour and expanding_quantile_band (31 traps, 12 honest controls)
• Lint rules (19 in total):
• reversed_cumulative (fail): cummax / cummin / cumsum / cumprod over a reversed series;
• reverse_count (fail): cumcount(ascending=False);
• backward_fill now also covers reindex(method="bfill" | "backfill" | "nearest");
• group_aggregate (warn) now covers groupby(...).last() / max() / size() / … and transform("size" | "count" | "nunique"); a following .shift(1) marks the aggregate as safe;
• full_sample_rank (warn) now covers Series.sort_values().

Notes

• No breaking API change vs 0.27.1. groupby(...)[col].max() without a shift is now reported as group_aggregate instead of full_sample_stat.

[v0.27.1] — 2026-09-26

Fixed

• The README demo docs/assets/demo-verify.gif was missing from v0.27.0: a *.gif rule in .gitignore came after the docs/assets/ exception and hid the file. The exception now comes last, and a new test checks that every asset referenced from README.md or docs/ is in the tree.

[v0.27.0] — 2026-09-26

Added

• GitHub Action signs certificates: new inputs signing-key (Ed25519 PEM from a repository secret) and upload-certificate (workflow artifact monte-neo-certificate), new output key-id; the step summary and PR comment name the signing key. The key is written to an owner-only temporary file and deleted right after signing. The action self-test now checks a signed certificate.
• Animated demo docs/assets/demo-verify.gif (leaky agent strategy → fix → honest verdict), rendered from real verifier output by scripts/make_demo_gif.py; shown in the README
• Guide: "Signing certificates in CI" in Use from agents

Notes

• No breaking change vs 0.26.0. Without signing-key the action behaves as before.

[v0.26.0] — 2026-09-26

Changed

• README and docs home rewritten around the verifier: the problem, real CLI output, where to use it, why Monte-Neo, quick start for the CLI, agents and the GitHub Action, and certificates
• Badges: PyPI version, total downloads (pepy), downloads per month (pypistats), Python versions, CI, license, MCP Registry, supported agents, stars; a "Verified by Monte-Neo" badge for users
• PyPI metadata: new description, keywords and classifiers; author NeoZorK
• New banner docs/assets/social-preview.png (README hero, GitHub social preview) and logo mark docs/assets/logo-sphere.png
• Research lanes frozen (bug fixes only): the indicator generator, the MLX/Metal engine, the paper OMS, the policy module and the CLI menu; see "Frozen lanes" in the contributing guide

Added

• CITATION.cff, glama.json
• Launch kit and article draft under docs/marketing/; internal marketing plan

Removed

• Tracked build artifacts: profile_stats.txt, exports/, verify_gpu_optimizer.py

Moved

• verify_hardware.py → scripts/; one-off Metal/MLX experiments → scripts/experiments/; tests/unit/test_coverage_boost*.py → tests/unit/legacy_coverage/

Notes

• No API change vs 0.25.0.

[v0.25.0] — 2026-09-26

Added

• Signed certificates (Ed25519, optional extra monte-neo[sign]):
• monte-neo verify --keygen PREFIX writes PREFIX.key (owner-only) and PREFIX.pub;
• --sign KEY adds a signature block to the certificate;
• --check-signature CERT [--public-key PUB] checks it (exit code 5 when invalid or signed by another key);
• library: sign_certificate, check_signature, generate_keypair; result schema strategy-signature-check/1
• MCP tool check_signature
• strategy-verdict/1 schema: optional signature property (backward compatible)

Notes

• No breaking API change vs 0.24.0. The default install is unchanged; cryptography comes only with the sign (or full) extra.

[v0.24.0] — 2026-09-26

Added

• Trap Suite: signal-processing leaks gradient_leak, convolve_same, fft_denoise, whole-sample ranks qcut_full, argsort_rank, plus the honest controls ewm_cross, convolve_causal and rolling_quantile_band (25 traps, 9 honest controls)
• Lint rules: central_difference (np.gradient, fail), centered_filter (convolve / correlate with mode="same", filtfilt, sosfiltfilt, savgol_filter, gaussian_filter1d, uniform_filter1d, medfilt, fail), full_sample_transform (fft, rfft, dct, hilbert, detrend, warn); qcut and argsort added to full_sample_rank
• Guide: Trap Suite, with the catalogue and how to contribute a trap; a "Trap submission" issue form; a test keeps the catalogue in sync with tests/traps/strategies/

Changed

• Lint reports one finding per rule and line (nested calls such as argsort(argsort(x)) no longer produce duplicates)

Notes

• No breaking API change vs 0.23.0.

[v0.23.0] — 2026-09-26

Added

• Trap Suite: last_row_leak, idxmax_leak, numpy_global_stat, plus the honest control cummax_drawdown (20 traps, 6 honest controls)
• Lint rules: last_row (fail) for .iloc[-k], .values[-k] and .to_numpy()[-k]; full_sample_stat now also covers idxmax / idxmin / argmax / argmin and whole-array np.mean / np.std / np.percentile / …
• Honesty Bench guide: a run checklist (contamination rules for each agent and task) and a leaderboard publication template
• Internal: cleanup candidate list with evidence; nothing removed

Notes

• No breaking API change vs 0.22.0. np.mean(series) and similar calls now produce a lint warn.

[v0.22.0] — 2026-09-26

Added

• Honesty Bench v1: monte-neo bench init <dir> writes five deterministic tasks (noise, momentum, mean-reversion, costs-trap, regime) with a shared prompt, 5 + 5 bps costs and a hidden answer_key.json
• Bench metrics from the answer key: false discovery (claimed profit on a task with no edge after costs) and edge found (edge task where the strategy passes verification)
• Trap Suite: diff_negative, pct_change_negative, roll_negative, merge_asof_forward, interpolate_leak, cumsum_total_norm, plus the honest controls resample_shifted and expanding_rank (17 traps, 5 honest controls)
• Lint rules: negative_period (diff / pct_change), negative_roll, forward_asof, interpolate (fail); sum added to full_sample_stat

Fixed

• test_generator_dynamic evaluated a random formula on 100 bars. Stacked windows and shifts can need about 180 bars of warm-up, so the test now uses 400 bars.

Notes

• No breaking API change vs 0.21.0.

[v0.21.0] — 2026-09-26

Added

• Agent Backtest Honesty Bench (monte_neo.bench, CLI monte-neo bench <dir>). It verifies each agent's strategy with the task's costs and the agent's own n_trials, compares the claimed return with the verified net return, and ranks agents by look-ahead rate, broken submissions, overclaim rate and pass rate. Report schema: honesty-bench/1, plus a Markdown leaderboard.
• scripts/make_example_bench.py: example bench with two illustrative agents built from the Trap Suite
• Guide: Honesty Bench

Fixed

• MCP Registry publish: server.json description shortened to ≤ 100 characters (the registry rejected v0.20.0 with 422); mcp-registry job stops retrying on validation errors
• CodeGenerator no longer emits binary expressions with identical operands (x / x, x - x), which produced constant indicators and made test_generator_dynamic flaky

Notes

• No breaking API change vs 0.20.0.

[v0.20.0] — 2026-09-26

Added

• Certificate re-check: recheck_certificate / monte-neo verify --recheck cert.json (exit 4 when not reproduced) / MCP tool recheck_certificate. It compares the data, signal and source hashes, re-runs the verifier (or the recorded grid search) and compares the verdict and certificate_id. Schema strategy-recheck/1.
• Grid certificates record grid.spec and grid.folds, so they can be reproduced exactly

• Trap Suite: resample_max_leak, target_encoding_leak (ML-style target encoding on the full sample)

• monte-neo mcp subcommand (same server as monte-neo-mcp)

• Official MCP Registry entry server.json (io.github.NeoZorK/monte-neo) + mcp-registry publish job (GitHub OIDC)

Changed

• mcp SDK is now a default dependency, so uvx monte-neo mcp works with no extras ([mcp] extra kept for compatibility)
• monte_neo and monte_neo.core export lazily, so import monte_neo, the verifier, the MCP server and the CLI no longer load MLX / Metal backends (fixes intermittent CI aborts on macOS runners)

Fixed

• BinanceDownloader._fetch_klines: a persistent HTTP 429 / rate limit now raises after MAX_RATE_LIMIT_RETRIES (5) instead of sleeping forever. This was the cause of CI jobs hanging for an hour.
• test_misc_one_liners no longer reaches Binance over the network; CI test job has timeout-minutes: 20

Notes

• No breaking API change vs 0.19.0.

[v0.19.0] — 2026-09-26

Added

• verify_grid: the verifier runs the parameter search itself. It expands the grid (≤ 512 combos), measures n_trials and the spread of trial Sharpes, verifies the best combo, and runs an anchored walk-forward (new walk_forward_oos check). Available through CLI --grid / --folds, the MCP tool verify_grid and the grid input of the Action.
• Static lint rules: reversed_window (fail), full_sample_rank, full_sample_stat, group_aggregate and polyfit fit (warn)
• Trap Suite: hourly_close_leak, reverse_rolling, full_polyfit, full_rank traps, the honest control expanding_zscore, and the grid strategies sma_params / momentum_params
• Claude Code plugin PostToolUse hook: reminds the agent to verify after editing strategy or backtest code (once per file per session, never blocks the edit)
• GitHub Action inputs grid, comment (PR comment updated in place) and github-token
• verify_strategy(extra_checks=, extra=) extension points; reproducibility.extra_sha256

Changed

• publish.yml runs on v* tag push: it checks that the tag equals the package version, publishes to PyPI, then creates the GitHub Release from this CHANGELOG section

Notes

• No breaking API change vs 0.18.0.

[v0.18.0] — 2026-09-26

Added

• Strategy verifier monte_neo.verify.verify_strategy → strategy-verdict/1 certificate (verdict, checks, metrics, next_actions, SHA-256 reproducibility block, deterministic certificate_id)
• Look-ahead probes: truncation, future perturbation (mirrored returns), determinism, AST lint (shift(-k), center=True, bfill, full-sample fit, x[i + k]), implausible next-bar hit rate
• Cost stress: break-even cost (bps per side), 0/1/2-bar execution delay scan
• Selection-aware statistics: PSR, Deflated Sharpe priced by n_trials, sample size, holdout consistency
• export_signals — bring-your-own positions through the fee-aware research bar (with signal SHA-256)
• CLI monte-neo verify (+ monte-neo-verify) with CI exit codes 0/1/2/3 and --out certificate
• MCP server monte-neo-mcp ([mcp] extra; MCP SDK 2.x MCPServer, 1.x FastMCP fallback): verify_strategy, probe_lookahead, cost_stress, verdict_schema, verifier_manifest
• Agent integrations in integrations/: Claude Code plugin (skill verify-strategy, /verify, .mcp.json) + root .claude-plugin/marketplace.json, Codex, Gemini CLI extension, Cursor rules
• Composite GitHub Action action.yml (fails on REJECT, step summary, verdict output)
• Trap Suite tests/traps: 5 lying strategies + 2 honest controls + data-snooping case
• Docs: Verifier API, Use from agents, internal product strategy

Changed

• README / docs home repositioned: "verify a trading strategy before you trust it"
• CI installs the mcp extra and runs tests/traps

Notes

• No breaking change to existing research / export APIs.
• License remains MIT. No peer product names in this tree.

[v0.17.7] — 2026-09-17

Changed

• Docs site homepage is native MkDocs docs/index.md at site root (no /home/ duplicate)
• Renamed docs/INDEX.md → docs/docs-map.md (macOS case-insensitive safe next to index.md)
• Native MkDocs docs/index.md as site root; docs.yml checks root index and rejects site/home/ (sed/copy-home hack removed)
• Maintenance tests and contributing rules point at docs/docs-map.md

Notes

• Docs / Pages infra only — no runtime API break vs 0.17.6.
• License remains MIT. No peer product names in this tree.

[v0.17.6] — 2026-09-17

Added

• GitHub Actions PyPI install smoke (.github/workflows/pypi-smoke.yml): workflow_dispatch, weekly schedule, and release published → clean venv + pip install monte-neo==… --no-cache-dir + import policy/holdout (via scripts/verify_pypi_install.sh, CDN retries)
• Links to SECURITY.md from README, CONTRIBUTING, docs INDEX, and MkDocs Project nav

Changed

• scripts/verify_pypi_install.sh default version → 0.17.6; --no-cache-dir + retry/sleep for PyPI CDN lag

Notes

• Docs / OSS trust only — no runtime API break vs 0.17.5.
• License remains MIT. No peer product names in this tree.

[v0.17.5] — 2026-09-17

Changed

• Research device="auto" prefers cpu_numba for wall clock on typical grids (measured M1 Pro: Metal was ~50–100× slower on 100k×64 / 1M×32)
• Sets fallback_reason="auto_prefer_cpu_numba" when auto declines Metal/MLX for speed
• Explicit device="metal" / "mlx" unchanged (size gate only; may be slow — user's choice)
• OMS resolve_device("auto") unchanged (still prefers Metal when available)

Added

• Env MONTE_NEO_RESEARCH_AUTO_PREFER_METAL=1 — restore pre-0.17.5 prefer-Metal research auto
• Env MONTE_NEO_RESEARCH_AUTO_METAL_MIN_COMBOS=N — allow auto→Metal only when n_combos >= N

Notes

• See Performance and FAQ.
• License remains MIT. No peer product names in this tree.

[v0.17.4] — 2026-09-17

Changed

• Performance: primary table is MacBook Pro M1 Pro 16 GB real timings (post-warmup)
• Honest guidance: for measured small/medium grids, cpu_numba beat Metal wall-clock while auto still selected Metal when eligible

Notes

• Docs / trust only — no runtime API break vs 0.17.3.
• License remains MIT. No peer product names in this tree.

[v0.17.3] — 2026-09-17

Added

• Job-fit scenarios on Home and FAQ: when Monte-Neo fits / when not (research bar vs OMS/live; peer-free)
• Performance rewrite with first-party export_sma_sweep timings + scripts/bench_research_bar.py
• MkDocs nav: Performance under Get started

Changed

• Quick start is export-first (pip → export → policy → holdout); CLI wizard moved to optional section
• Softened stale “125x” / “300k ops/sec” / “v0.0.4 targeting” claims in public docs

Notes

• Docs / trust pack only — no runtime API break vs 0.17.2.
• License remains MIT. No peer product names in this tree.

[v0.17.2] — 2026-09-17

Changed

• Holdout default promote_mode="holdout_positive": positive holdout no longer blocked by large train−holdout gap
• HeuristicPolicy: block promote only when holdout_promote_ok is false; high gap → MC flag; holdout can unlock promote

Added

• promote_mode="strict" for legacy conservative gate
• append_research_label JSONL logger + CLI --holdout-log / --human-label (corpus for a future LocalScorer — not B yet)

Notes

• LocalScorer B still deferred until real labeled runs exist.
• License remains MIT.

[v0.17.1] — 2026-09-17

Added

• CONTRIBUTING + GitHub issue/PR templates
• Install audit docs: pipx, extras matrix, safe TestPyPI --no-deps recipe
• FAQ: TestPyPI poison, extras ImportError, policy/holdout CLI pointers

Changed

• Installation page current release → 0.17.x; docs site link in install/FAQ
• MkDocs nav: Contributing

Notes

• Docs/OSS friction only — no runtime API break vs 0.17.0.
• License remains MIT.

[v0.17.0] — 2026-09-17

Added

• Holdout helper (holdout_sma_sweep, split_bar_range): train SMA sweep → score top-K on holdout
• Schema mn.holdout_report.v1 with gap / overfit_risk / promote_ok
• HeuristicPolicy enrichment via build_research_state(..., holdout_report=)
• CLI: monte-neo --holdout-sma
• Docs: Holdout API

Notes

• Practical anti-overfit without ML. LocalScorer B still deferred.
• License remains MIT. No peer product names in this tree.

[v0.16.0] — 2026-09-17

Added

• HeuristicPolicy A (monte_neo.policy): local deterministic triage after research export
• build_research_state → compact mn.research_state.v1 (no raw OHLCV)
• triage_export / HeuristicPolicy.decide → next_action, promote/MC flags, reasons
• CLI: monte-neo --policy-triage path/to/export.json
• Docs: export API + policy pages; install extras matrix ([apple], [plot], [data], [ml], [server], [full])

Changed

• Slim default dependencies — research-core only (numpy/pandas/pyarrow/numba/rich/CLI)
• Optional extras: apple (darwin markers), plot, data, ml, server, full
• Lazy imports for Binance downloader, websocket client, plot helpers (monte-neo[data] / [plot])
• CI syncs --extra apple --extra plot --extra data --group dev
• Documentation URL → https://neozork.github.io/Monte-Neo/

Notes

• Not a cloud “System One” model — offline rules only. Holdout helper / LocalScorer B deferred.
• License remains MIT. No peer product names in this tree.

[v0.15.1] — 2026-09-16

Fixed

• Bare pip install without MLX (from __future__ import annotations on acceleration modules)
• Quiet Metal-unavailable import noise (debug level)
• Include monte_neo.data in wheels (.gitignore /data/)
• CI: portable CLI subprocess cwd; ruff clean on coverage boosts
• Coverage close-out for sequential advice + cli __main__ pragma

Packaging

• Ready for first TestPyPI / PyPI upload as PEP 440 0.15.1
• GitHub tag v0.15.0 predates these packaging/CI fixes — ship as 0.15.1

[v0.15.0] — 2026-09-15

Fixed

• verify_export_golden / verify_golden_vectors: auto Metal float32 tolerance band when batch resolves to Metal
• Unit metrics: pin Numba path for deterministic trade_count; native extract stub via _get_native
• Build: pin hatchling>=1.24,<1.26 so wheels emit Metadata 2.3 (twine-check friendly)
• Packaging: .gitignore /data/ (root only) so monte_neo.data is included in wheels
• Import without MLX: from __future__ import annotations on acceleration modules so bare pip install works
• Metal extension absence logged at debug (no emoji warning on every import)

Packaging / PyPI prep

• mlx + PyObjC Metal moved to optional extra monte-neo[apple] (Linux/CI-friendly wheels)
• Hatch version pattern strips leading v for PEP 440 distribution version
• Richer project metadata: description, keywords, classifiers, project.urls
• docs/project/PACKAGING.md — free-only PyPI / TestPyPI checklist

Docs / examples

• Demo assets: docs/assets/demo_sma_sweep.png, docs/assets/demo_memory_plan.png
• examples/export_sma_sweep_quickstart.py + examples README
• README install notes for [apple] and git install

Notes

• First PyPI upload still gated on TestPyPI + maintainer OK (see PACKAGING.md)
• License remains MIT. No peer product names in this tree.

[v0.14.1] — 2026-09-15

Fixed

• Research accelerator no-hang gate: oversized Metal/MLX workloads fall back to cpu_numba instead of blocking forever on GPU wait (e.g. 10M-bar sweeps on 16GB-class hosts)
• Transparent fallback_reason when auto demotes Metal (metal_max_bars_exceeded, shared-budget, host soft budget)

Docs

• README: install from GitHub, research export quickstart, advantages, device notes
• docs/guides/FAQ.md starter

Notes

• Env knobs: MONTE_NEO_RESEARCH_BYTES_BUDGET, MONTE_NEO_METAL_SHARED_BYTES_BUDGET, MONTE_NEO_METAL_MAX_BARS
• License remains MIT. No peer product names in this tree.

[v0.14.0] — 2026-09-15

Added

• plan_research_bytes for Apple Silicon 16GB soft budgets / tile hints
• Export depth: equity_stride, include_journal, optional memory on single/batch
• README positioning: local macOS research job vs non-goals (no rival names)

Notes

• License remains MIT. Evidence publication stays permission-gated and out of tree.

[v0.13.0] — 2026-09-15

Added

• Research signal factory (build_sma_cross_grid): Numba-parallel SMA cross grids
• Sweep/export expose signal_elapsed_s + economics_elapsed_s (wall includes both)

Notes

• Default signal device is Numba (exact); MLX signal path remains opt-in.
• No peer product names in this tree. License remains MIT.

[v0.12.0] — 2026-09-14

Added

• Stable research-bar export API (export_single / export_batch / export_sma_sweep / research_manifest)
• Frozen golden vectors (verify_golden_vectors) for fee-hurts + batch↔single parity before any external timer
• Timers disclose includes_signal_build so harnesses cannot mis-attribute work

Notes

• No peer product names in this tree. Export is for external honesty harnesses only.
• License remains MIT.

[v0.11.0] — 2026-09-14

Added

• Research Metal economics: funding, session mask, long_short (Numba golden parity)
• Rewrote stale MLX engine unit tests against current kwargs-only API (no skips)

Notes

• Evidence publication stays out of this repository.
• License remains MIT.

[v0.10.0] — 2026-09-14

Added

• Research Metal economics: SL/TP/trail on long/flat next-bar-open (Numba golden parity)
• Parity + perf coverage for Metal stop/trail subset

Notes

• Funding / session mask / long_short remain Numba-only.
• Evidence publication stays out of this repository and permission-gated.
• License remains MIT.

[v0.9.1] — 2026-09-14

Removed

• Public evidence-reproduce helpers, CLI entrypoint, and core/race optimization surface
• Development race docs that described public evidence reproduce flows

Notes

• Evidence publication remains permission-gated and lives outside this repository.
• License remains MIT.

[v0.9.0] — 2026-09-14

Added

• Research bar Metal economics batch (long/flat next-bar-open subset)
• OMS: BarClock, TimeInForce (GTC/IOC), OCO/bracket helper, honest max_fill_qty partials
• OMS: adapters/replay.py OHLC replay; accel BufferPool + shader_catalog
• device= on run_bar_backtest_batch / run_sma_sweep (Metal when eligible, else Numba golden)
• Parity + perf smoke for Metal vs Numba (fill_fraction / leverage included)

Notes

• SL/TP/trail/funding/session/long_short remain Numba-only (full ExecutionModel).
• Metal float32 vs Numba float64 uses documented rtol/atol.
• License remains MIT. Evidence publication still permission-gated (out of tree).

[v0.8.0] — 2026-09-14

Added

• MC Metal/MLX/CPU unified dispatch (plan_mc_run) honoring compute_device / use_gpu
• M1 Pro 16GB-class scenario memory budgets (plan_scenario_budget, tile ranges)
• MCResult.device_used / bytes_peak_est / accel_plan; MCConfig.memory_budget_bytes
• Docs: docs/project/mc_accel.md

Notes

• Metal-first on Apple Silicon; MLX used when native params are absent but MLX repr exists.
• License remains MIT. Evidence publication still permission-gated (out of tree).

[v0.7.0] — 2026-09-14

Added

• Metal L2 book-walk dispatch (PyObjC) with Numba float64 golden parity
• run_l2_walk / run_l2_walk_batch device-aware paths; TickL2Engine uses them
• Unit parity + performance smoke for Metal L2 batch walk

Notes

• Metal uses float32; parity tests allow documented rtol/atol vs Numba float64.
• License remains MIT. Evidence publication still permission-gated (out of tree).

[v0.6.0] — 2026-09-14

Added

• Metal OMS batch dispatch (PyObjC) with Numba float64 golden parity
• run_batch_terminal(device=...) / preferred_compute_device for honest device reports
• MCConfig.compute_device + engine compute_pref checklist
• Perf smoke for auto Metal/Numba batch path

Notes

• Metal uses float32; parity tests allow documented rtol/atol vs Numba float64.
• License remains MIT. Evidence publication still permission-gated (out of tree).

[v0.5.0] — 2026-09-14

Added

• Venue adapters: PaperExchangeAdapter, BinanceAdapter, BybitAdapter
• make_adapter() factory + fill reconciliation helper
• Live safety gates: MONTE_NEO_LIVE_TRADING, MONTE_NEO_LIVE_DRY_RUN (default dry-run)
• .env.example keys for Bybit + live flags

Notes

• Paper mode is default (no network). Live submit without dry-run is blocked in this build.
• License remains MIT. Evidence publication still permission-gated (out of tree).

[v0.4.0] — 2026-09-14

Added

• Tick/L2 OMS lane: TickL2Engine, synthetic ticks, order book, L2 walk
• Numba walk_book_market with Python L2 parity tests
• Metal shader scaffold oms_l2_walk.metal
• Docs updated for dual research + OMS tick paths

Notes

• Paper/synthetic books only in this cut (live adapters later).
• License remains MIT. Evidence publication still permission-gated (out of tree).

[v0.3.0] — 2026-09-13

Added

• Paper OMS lane (monte_neo.oms): orders, matching, portfolio netting, blotter
• Accel device select (cpu_numba / metal / mlx / auto) for Apple Silicon
• Numba OMS batch helper + Metal shader scaffold (oms_bar_match.metal)
• Docs: docs/project/oms_engine.md; roadmap rules (silence, Evidence gate, sync)

Notes

• Research bar engine (monte_neo.backtest) unchanged as a separate lane.
• License remains MIT.
• No public Evidence publication without maintainer permission after correct private runs.

[v0.2.0] — 2026-09-13

Added

• Session mask (block new entries off-session; exits/SL/TP still allowed)
• Funding lite (funding_bps_per_bar) and leverage (leverage >= 1)
• Shared-cash multi-symbol portfolio (run_portfolio_shared_cash)
• Correctness suite: dual-hit SL preference, batch≡single with new knobs

Notes

• Research bar engine scope unchanged (not a full OMS).
• No public speed evidence until private comparisons complete and user OK.
• License remains MIT.

[v0.1.0] — 2026-09-13

Added

• Brand logo (docs/assets/monteneo-logo.png) and honest GitHub-facing README
• GP bar engine: StrategySpec / run_strategy_backtest, trail stops, impact_bps, partial fill_fraction, expanded work checklist
• Docs: clearer scope (research bar engine ≠ full OMS)

Notes

• License remains MIT.

[v0.0.9] — 2026-09-13

Added

• ExecutionModel.sl_pct / tp_pct with H/L exits (SL preferred on dual hit)
• Trade journal + metrics (sharpe, profit_factor, win_rate, max DD)
• run_bar_backtest_batch for external signal matrices (shared economics)
• run_multi_symbol_lite (independent cash books)
• run_sma_sweep now wraps batch (labeled convenience, not D001 kernel)

Notes

• Costs remain bps on this path; older metrics helpers may use %.

[v0.0.8] — 2026-09-13

Added

• Professional fee-aware bar backtest engine (monte_neo.backtest): next-bar fills, commission/slippage bps, cash/position/equity, SMA sweep with the same ExecutionModel, optional ReplayBarSource mid-price feeder (no Redis on hot path)
• Docs: docs/project/backtest_engine.md
• Unit / integration / stress / performance tests for the new engine

Notes

• Specialized SMA kernel gates for private local protocols remain out of scope; this package path is for matched-semantics peer races (private or future D002).

[v0.0.7] — 2026-09-12

First formal GitHub Release. Canonical development branch is main (older branches named v0.0.1 … v0.0.6 are historical archives only).

Added

• Public evidence-reproduce helpers and CLI (later removed in v0.9.1)
• Type B CPU fallback when Metal native extension is unavailable
• Unit tests for license/version doc sync

Changed

• License metadata aligned to MIT (LICENSE + pyproject.toml)
• Docs / INDEX / ROADMAP synchronized to v0.0.7

Security

• Pillow pinned to >=12.3.0 on the release line (Dependabot alerts cleared)

Notes

• Install: pip install "monte-neo @ git+https://github.com/NeoZorK/Monte-Neo.git@v0.0.7"
• Or: uv add "monte-neo @ git+https://github.com/NeoZorK/Monte-Neo.git@v0.0.7"