Format: Keep a Changelog. Every entry below describes a customer-visible change to the Reamer Research product — the C ABI, the shipped reference material, or the kit contents — not internal refactors. Updated at release, alongside the tag.
Version numbers here are the release version (kReamerVersion), which is
shared with Reamer Server and moves independently of REAMER_ABI_VERSION.
ABI bumps are called out explicitly where they occur.
[4.11.3] - 2026-11-20
Evaluation Mode explains itself. The C ABI (REAMER_ABI_VERSION 8) is
unchanged; messages, the sample bars and reamer-license status changed.
Fixed
- A backtest past an Evaluation Mode limit now reads "Evaluation Mode runs one ticker over at most 30 days or 50,000 bars per backtest; <the limit hit>. A Reamer Research licence removes every limit: https://reamerlabs.com/pricing". A lapsed licence adds its state; a machine never activated adds nothing.
- The relay's refusal without a licence says the live relay needs a Reamer Research licence and that Evaluation Mode covers backtests.
python/sample_data/AAPL.csv, the file the Python README and the C++ README run first, spanned 139 days and was refused in Evaluation Mode. It is now 100 synthetic hourly bars over three weeks and runs before activation.reamer-license statusin the Research kit states, without a Research seat, that backtests run in Evaluation Mode, with its limits and how to activate.
Added
rr.EvaluationLimit(Python) andrr::EvaluationLimit(C++), raised forLICENSE_MISSINGwith the engine's message alone. They derive fromReamerErrorandrr::Error, andcode_namestaysLICENSE_MISSING.
Changed
RESEARCH_TO_SERVER.mdis now "Running Reamer Research with Reamer Server", for firms that also run Reamer Server as their order-management core. Its third item now points topython/live/live_paper.py, whose rolling window letson_barrun live unchanged. The relay contract and the one-order round trip are unchanged.
[4.11.2] - 2026-11-18
The examples release. The library, the C ABI (REAMER_ABI_VERSION 8) and the
wrappers' APIs are unchanged from 4.11.1.
Fixed
- Every Python example and template runs directly as
python3 examples/<name>.pyfrompython/. Before, most had no__main__and did nothing when run. Single-ticker files run on the sample bars in Evaluation Mode; multi-ticker files stop with a one-line message that they need a licence. limit_entry_strategy.pyplaces limits the sample bars reach, andrisk_sized_strategy.pysizes inside its notional cap, so both trade.- The C++ examples build and run from the kit on the sample bars:
quickstart.cpp(a lambda, or an object with--object),atr_stop_bracket.cpp(an object),buy_and_hold.cpp,donchian_breakout.cpp(a function, with good-till-date entry stops and a trailing exit) andrejected_orders.cpp, which checks the three rejections it provokes. python/live/live_paper.pystops with one line, not a traceback, when no licence is active.
Added
cpp/examples/data_csv.cppandcpp/examples/data_bin.cpp: loading CSV bars, converting them to the binary format and running from the middle of a file, from C++.python/live/:relay_receiver.py(a receiver and paper broker, standard library only) andlive_paper.py.relay_receiver.pyalso runs on its own aspython3 live/relay_receiver.py <socket>.cpp/live/live_paper.cpp: the same paper trade from C++, against the same receiver. It reports the same fills and position aslive_paper.py.cpp/README.md: "Getting data in" and "Paper trading" sections.- Release CI runs every shipped example from the packaged kit, as a user would, on Linux and macOS.
[4.11.1] - 2026-11-17
The relay sends every action a backtest strategy can return. REAMER_ABI_VERSION
is 8, which only adds functions: code built against 4.5.0 or later runs as it
is, and the wrappers accept a library at their ABI or later.
Fixed
- The relay sends
modify_order,modify_position,close_allandcancel_all. Before,reamer_relay_send()had no way to take them andRelay.send()in both wrappers refused them.modify_ordergoes out as a Replace (message type 2) with new brackets in a replace extension block, plus a Cancel for each exit it clears;modify_positionreplaces, cancels or places the position's exits;close_allsends one close per open position;cancel_allcancels every order resting before the call. Actions run in the backtest's order: the call's orders, thenclose_all's closes, then the other actions. Every order and action is checked before anything is sent. Relay.send()in Python returned order IDs that drifted from the relay's after a cancel or a refused order. It now returns the IDs the relay took,close_all's closes included, andrelay.sentcounts orders as the relay numbers them.- A refused order or action names the rule it broke instead of a generic relay error.
reamer_relay_get_positions()reportstake_profitandstop_loss: the levels of the position's live exits.- The relay tracks each order to the end. An update with status cancelled or rejected ends it, an exit ends when its one-cancels-other partner fills, and a position that goes flat cancels its live exits.
Added
reamer_relay_send_batch(): orders andReamerOrderActions in one call.reamer_relay_send()is that call with no actions.reamer_relay_get_counts(): order IDs taken and frames sent.- Python
relay.action_resultsandrelay.messages; C++rr::modify_order,rr::modify_position,rr::close_all,rr::cancel_allfor the relay,Relay::send()returning order IDs,action_results(),orders_sent()andmessages_sent(). - The relay section of
include/reamer_research_abi.hdocuments the Replace frame, thecontingency_type/linked_intent_idexit links and how each action is sent. Receivers that ignore message type 2 keep working for orders and cancels.
[4.11.0] - 2026-11-17
Reamer Research runs without a key in Evaluation Mode, within fixed limits.
REAMER_ABI_VERSION stays 7, and code built against 4.5.0 or later runs as
it is.
Added
- Evaluation Mode. Without an active licence,
reamer_run_backtest()andreamer_run_backtest_files()run one ticker over at most 30 days of bar time (last timestamp minus first) or 50,000 bars, whichever comes first. For a files run the limits apply to the selected range. A run past a limit fails withREAMER_RESEARCH_ERROR_LICENSE_MISSINGand a message naming the limit; nothing is truncated.reamer_run_monte_carlo()and the visual trade diagnostics tool work on an evaluation run's result. The relay stays licence-only. An active licence removes every limit. python/sample_data/SAMPLE_5min.csv: 1,638 synthetic five-minute bars (a seeded random walk, not market data) that fit inside the Evaluation Mode limits, sopython examples/quickstart.py sample_data/SAMPLE_5min.csvruns before activation.- The kit downloads without a key from https://reamerlabs.com/download.
- The wheel carries
LICENSE.txtandTHIRD_PARTY_NOTICESin its.dist-info/licenses/, so the terms travel with an installed package. - The examples, templates, fetchers and sample data are also public under
the MIT License at https://github.com/reamerlabs/reamer-research-examples,
with new data examples:
examples/data_csv.py,data_bin.py,data_parquet.py,data_alpaca.pyanddata_databento.py.
Changed
LICENSE.txtis version 1.6. Section 2B grants use in Evaluation Mode without payment or a key.reamer-license activateasks you to accept the new version once.- A refused relay open now says where to buy a licence.
[4.10.0] - 2026-11-16
The Python package and the C++ wrapper are now supported product, in
directories named for what they are. REAMER_ABI_VERSION stays 7, and code
built against 4.5.0 or later runs as it is.
Changed
reference-python/is nowpython/, andreference-cpp/is nowcpp/. Install from source withpip install ./python(waspip install ./reference-python); aPYTHONPATHor include path that named the old directory needs the new name. The wheel,include/reamer_research.hpp,lib/cmake/andlib/pkgconfig/are where they were, and the Python and C++ APIs are unchanged.LICENSE.txtis version 1.5. The Python package and the C++ wrapper are now Wrapper Code: part of the licensed product and supported like the library, while staying in source for you to read, change and keep (Section 4).reamer-license activateasks you to accept the new version once.- The visual trade diagnostics tool is named that way throughout the kit's
documentation, its page title and the line a run prints when it opens it.
viewer/reamer-viewer.html,open_viewerandviewer_htmlkeep their names. - The C++ header's ABI-mismatch message starts with
[ERROR] reamer_research.hpp:(was[ERROR] reference-cpp:). - The kit's README.md leads with the Python and C++ quickstart.
[4.9.0] - 2026-11-12
Python and C++ researchers run a backtest without writing any ABI plumbing:
pip install the package, or include one header. REAMER_ABI_VERSION
stays 7, and code built against 4.5.0 or later runs as it is.
Added
- Python package
reamer_research, shipped built aspython/reamer_research-<version>-py3-none-<tag>.whlwith the library and the visual trade diagnostics tool inside it, and installable from source withpip install ./reference-python([pandas]extra for DataFrames).rr.backtest(on_bar, "AAPL.csv")takes a function or class strategy and CSV paths,.binpaths (memory-mapped), pandas DataFrames or numpy bars, withlookback, per-ticker costs (ticker_config), exogenous sidecars,start/end,result_json,console_logandopen_viewer. Order helpers (rr.buy_market(10, stop_loss=...)and the rest),rr.Result(every summary field,closed_trades,trades_frame(),json(),monte_carlo(),open_viewer()),load_csv,to_bars,write_bin,read_bin,write_exogenousandrr.Relay. - C++ wrapper
include/reamer_research.hpp, one header, C++20:rr::backtest(on_bar, "AAPL.csv", {.lookback = 30})with a lambda or an object strategy, CSV,.binor in-memory bars, fluent orders with.bracket(tp, sl)and time in force,rr::Resultwithclosed_trades(),json(),monte_carlo()andwrite_viewer(),load_csv,write_bin,read_bin,write_exogenousandrr::Relay. Failures throwrr::Error. lib/cmake/ReamerResearch/(find_package(ReamerResearch), targetReamerResearch::reamer_research) andlib/pkgconfig/reamer-research.pc.viewer/reamer-viewer.html, a visual trade diagnostics tool: each closed trade charted on its bars beside the values the strategy printed, the console lines for each bar, and a Monte Carlo simulation of the run. The chart shows a fixed window around the selected trade, with a crosshair and a readout of the hovered bar. In a run of several tickers, a line that names its ticker (AUDUSD_M15 mid=...) is plotted on that ticker only. One HTML file; nothing is uploaded.open_viewer=True(Python) and.open_viewer = true(C++) open a run in it as soon as it finishes;python3 -m reamer_research.viewerbuilds the page from a C program's files. On a machine with no desktop the page's path is printed and no browser is started.- Quickstarts of about a dozen lines, an ATR band reversion with a bracket
on every entry:
reference-python/examples/quickstart.pyandreference-cpp/examples/quickstart.cpp. They return early while the lookback window is still filling, when its oldest rows are placeholders that read as 0 (valid()is 0 for them).
Changed
- A market entry whose fill gaps past its own take profit or stop loss keeps
its bracket, and the passed level exits on the next tick after the fill at
that tick's price. Before, the engine dropped both levels and the position
ran with no bracket. The rule is in
EXECUTION_SPEC.md§0b. run_backtest()andreamer_research_bindingfrom 4.8 still import and run, and now return anrr.Result. A strategy's exception is raised from the run instead of ending it quietly, andlookbackis read from an instance as well as from the class.closed_tradesis no longer capped at 4,096 trades, in Python and in the 4.x C++ reference wrapper (namespacereamer_research_ref, kept in the new header so its code compiles unchanged).
[4.8.1] - 2026-11-09
Reamer Research becomes a monthly subscription, and the engine picks up
each renewal by itself. REAMER_ABI_VERSION stays 7, and no strategy code
or results change.
Added
- Automatic licence refresh. Within two days of a Reamer Research grant's
end date, a run first asks the licence server to re-sign the grant this
machine already holds, so a renewed subscription needs no second
reamer-license activate. The request sends the key, machine fingerprint and seat id that activation already sent, and nothing else; it never moves a seat, waits at most five seconds, and retries at most every six hours before the end date and every hour after it, for 30 days. Any failure leaves the grant on disk to decide as before.REAMER_LICENSE_NO_REFRESH=1turns it off. reamer-license refreshmakes the same request on demand.
Changed
LICENSE.txt1.4: the Reamer Research Term is the period paid on the subscription; trials are Reamer Server only; a Reamer Research licence is not reissued to a different employee (cancel and subscribe again).LICENSING.md: monthly Reamer Research pricing and renewal; Reamer Server USD 24,000 a year with a USD 2,000 30-day trial.- Reamer Research updates are self-serve from the download link in the license email; releases are not sent on email request.
activation.jsonis written to a temporary file and renamed into place, so a process checking its licence never reads a half-written grant.
[4.8.0] - 2026-11-08
Windows support through WSL2, and market-data fetchers for the Python
binding. REAMER_ABI_VERSION stays 7, and no strategy code or results change.
Added
- Windows through WSL2 is a supported platform, using the Linux x86-64 kit. Supported platforms covers the requirements.
reference-python/fetchers/: download historical bars from Alpaca (US stocks and crypto) and Databento (OHLCV schemas, including continuous futures across rolls) straight into the formrun_backtest()takes.write_csv()saves them forbin/reamer-csv-build. The Alpaca fetcher is standard library only; the Databento one needspip install databentoand prices each request before downloading it. Source, like the binding. The engine makes no network calls.reference-python/examples/fetch_and_run.py: fetch bars from either vendor and run a multi-ticker backtest in one command. Alpaca crypto needs no account.
Fixed
- An activation inside WSL2 now survives a restart of the WSL VM or of
Windows. WSL2 gives its VM a new MAC address on every restart, which changed
the machine fingerprint and made
reamer-licensereport the licence as active on another machine. Under a WSL kernel the fingerprint now uses/etc/machine-idin place of the MAC. Native Linux and macOS fingerprints are unchanged, so existing activations stay valid.
[4.7.0] - 2026-11-06
A core change to how time runs in a backtest. REAMER_ABI_VERSION stays 7, and
no strategy code needs to change.
Changed
bin/reamer-csv-build: a CSV with two rows at the same timestamp now stops the tool. It names the first repeated timestamp, exits non-zero and writes nothing. Before, it dropped the repeat with a warning and exited 0, so a sub-second feed could silently lose most of its bars.--keep-firstkeeps the old behaviour.
Fixed
- One wall clock now runs every backtest, from the earliest row of any ticker to
the close of the last row. Every row of every ticker is an event on it at its
own timestamp,
on_barruns the moment a row closes, and fills run off the same clock. Rows never need to line up with anything. Before,on_barfollowed row timestamps while each ticker's fills ran on its own counter, so tickers that were offset, on different timeframes or with differenttick_countfilled ahead of the rowon_barwas shown: a 1-minute ticker beside a daily one filled 1,440 rows ahead. Results on those multi-ticker runs change. Runs on one ticker, and tickers sharing one grid and one tick count, give identical results. - A row reaches
on_barwhen it closes, so a ticker with a gap shows its row on the step of its next row's start, not the step of its own timestamp. - An order on a ticker part-way through a row now fills from that row's next tick.
Before, it was rejected with
"no market data", which now means the ticker has no closed row yet. - Rows sharing a timestamp fill one after another. An order written when the first closes fills on the second, not on the first row of a later timestamp.
- A futures roll on a ticker whose rows do not fall on every step is no longer lost.
- Every ticker in a run is cut to the latest first bar and the earliest last
bar across the run, on arrays and
.binfiles, so the shortest ticker sets the run's length. Tickers that started apart filled ahead of the baron_barwas shown, by the gap between their starts: a ticker listed 30 days late filled 30 bars ahead. The run is refused if the tickers share no range, or if a ticker has no bars inside it. Multi-ticker results on tickers with different start or end dates change. reamer_run_backtest()refuses bars that are not sorted ascending byts. The header always required this, but only.binfiles were checked.reamer_run_backtest()andreamer_run_backtest_files()refuse atick_countbelow 1 other than the -1 sentinel, and one past the range ofint.reamer_write_bin()refuses the same bars. A bar with no ticks seals at once, so a run with atick_countof 0 returned success with no fills.reamer_run_backtest()refuses a bar with notick_countthat shares itstswith the next bar. The gap gave such bars no ticks, which ran the engine ahead of the stepon_barwas shown, so a fill could land on a bar the strategy had not seen yet. Same-second bars with a realtick_countstill run.
Documentation
README.mdstates the scope as bars of any type. OHLCV is the record format, not a limit on how a bar was built.EXECUTION_SPEC.md,DEPLOYMENT.mdand the header state the whole-second timestamp rule and thetick_countrange.
[4.6.2] - 2026-11-05
No ABI change (REAMER_ABI_VERSION stays 7). The change is in
bin/reamer-csv-build; results from an existing .bin file are identical
to 4.6.1.
Fixed
bin/reamer-csv-build: a CSV with several symbols in asymbolortickercolumn was written as one series. Rows that shared a timestamp were dropped as duplicates, and rows that did not were interleaved, with exit code 0 and no warning about the mix. The tool now names the symbols, exits non-zero and writes nothing. If you built a.binfile from such a CSV, rebuild it.
Added
bin/reamer-csv-build --symbol NAMEconverts one symbol from a CSV that holds several. The match is exact.bin/reamer-csv-buildrecogniseswindow_startandtas timestamp columns, andtransactionsandnas trade counts.- The parse report has three new lines:
symbol_column,symbolsandfiltered_rows.
[4.6.1] - 2026-10-26
No ABI change (REAMER_ABI_VERSION stays 7). Reference Python binding only;
the library is unchanged and results are identical to 4.6.0.
Fixed
- Reference Python binding: a backtest with more than 65,536 closed trades
raised
RuntimeErrorinstead of returning them. The trade buffer is now sized from the run's own trade count, so every closed trade is returned.
Changed
- Reference Python binding: a numpy structured array of bars is converted field by field instead of row by row. On 500,000 bars, peak memory drops from about 290 MB to about 210 MB and start-up is about 1.5 s faster. Fields are still matched by name, so their order, extra fields and narrower types are handled as before.
[4.6.0] - 2026-10-25
No ABI change (REAMER_ABI_VERSION stays 7): no struct or signature
changes. Behaviour change: a strategy's results can change, so re-run
backtests made on earlier versions.
Fixed
on_bar(N)now sees the state after bar N's fills, as a venue reports fills before the bar closes.positions,position_statesandopen_orderswere one bar stale: an order that filled at bar N's open still showed as resting and the position as unchanged, andunrealized_pnlwas marked at bar N-1's close. Fill timing is unchanged: an order returned fromon_bar(N)still fills from bar N+1 tick 1.- A full position close no longer cancels every pending order on that
ticker. It cancels only orders that close (
is_close, or an opposite-side order with noqty); entries and reversals stay live. Together with the stale state above, this cancelled the second of two reversals on consecutive bars, leaving the strategy on the wrong side. - Cancel requests,
MODIFY_ORDER,MODIFY_POSITIONandCANCEL_ALLapply whenon_barreturns; bar N has already replayed by then, so the observable timing is unchanged.
Changed
- Reference Python binding: each
on_barcall is about 4x cheaper. The bar window's numpy dtype is built once instead of on every call; a 500,000-bar no-op run drops from about 38 s to 10 s. Results are unchanged.
[4.5.1] - 2026-10-15
No ABI change (REAMER_ABI_VERSION stays 7). A security release: rebuilding
against it is not required.
Security
- The license library's internal test switches are no longer compiled into
any shipped binary. A release gate fails the build if one reaches the kit.
See
SECURITY_HARDENING.mdS10. - Every executable in
bin/ships stripped. bin/release-signis a verify-only build.verifyis unchanged.
[4.5.0] - 2026-10-09
ABI bump: REAMER_ABI_VERSION 6 → 7. Breaking. Rebuild strategies
against the ABI 7 header.
Added
on_barreceivesposition_states(oneReamerPositionStateper ticker: signedqty,avg_entry_price,take_profit,stop_loss,unrealized_pnl,entry_ts) andopen_orders/open_order_count(every resting order as aReamerOpenOrder, sorted byorder_id). A strategy can now cancel its own resting order by theorder_idit reads here.on_barwrites order and position actions toout_actions(ReamerOrderAction, up to 64 per call):MODIFY_ORDERre-prices a resting order and replaces its brackets,MODIFY_POSITIONreplaces a position's take-profit and stop-loss,CLOSE_ALLcloses every position at market,CANCEL_ALLcancels every resting order. Modify and cancel-all take effect from the next bar; a stop may move to breakeven or into profit. Each action reportsappliedandreject_reason.EXECUTION_SPEC.md§6b has the timing and validation.- Optional vtable slot
on_progress(user_data, steps_done, steps_total). The engine calls it between bars, at most once per 256 bars and 100 ms, plus a final call withsteps_done == steps_total. Return nonzero to stop the run:reamer_run_backtest()andreamer_run_backtest_files()returnREAMER_RESEARCH_ERROR_CANCELLED(11) and no handle. NULL means no calls. A run's result is the same with or without it. - Result JSON
order_logrecords each applied modification with status"Modified": the order's snapshot under its ownid, orid0 for a position's brackets. - reference-cpp passes the new arrays to the strategy function as spans,
plus
std::span<ReamerOrderAction>and asize_t&action count. reference-python exposes them asdata.position_state(ticker),data.open_orders()anddata.action_results();engine/actions.pybuilds the four actions, and a strategy'son_progress(done, total)method, if defined, receives progress calls and stops the run by returning a truthy value. - Strategy relay to Reamer Server:
reamer_relay_open(),reamer_relay_send(),reamer_relay_poll(),reamer_relay_get_positions(),reamer_relay_is_connected(),reamer_relay_close()andReamerRelayHandle. Sends the sameReamerOrderRequesta strategy returns fromon_barto a running server over its local-mode strategy socket, with backtest order IDs and submission rules, and reads fills back asReamerPositionState.take_profitandstop_losstravel as Reamer Server ABI 4 order fields; the server tracks the exits as<order_id>:tpand<order_id>:sl, your broker connector places them at the venue, and their fills move the reported position. Brackets on a close (is_close) are rejected: they attach to an entry. New error codeREAMER_RESEARCH_ERROR_RELAY(12). Linux and macOS.RESEARCH_TO_SERVER.mdhas a one-order round trip against the reference gate. The gate and the broker connector remain yours to build.
Changed
on_bartakes three new parameters beforeout_orders:const ReamerPositionState* position_states,const ReamerOpenOrder* open_orders,size_t open_order_count, and three aftermax_orders:ReamerOrderAction* out_actions,size_t max_actions,size_t* out_action_count. The vtable becomes{user_data, on_bar, on_progress}(24 bytes).- A cancel request (
is_cancel) takes effect after the current bar replays, not before. In 4.4.0 a strategy that had seen bar N could cancel an order bar N would have filled. Runs that cancel orders can now differ from 4.4.0: an order that fills on the bar the strategy saw stays filled.EXECUTION_SPEC.md§6. - Result JSON
schema_version1 → 2:order_logidrepeats on"Modified"records.summary.total_ordersdoes not count them. - A strategy that writes no actions produces the same result JSON as 4.4.0,
byte for byte, apart from
abi_versionandschema_version. - Cost: about 1-2%.
research-benchP50 is 1.98–2.00 µs peron_barcall against 1.96 µs on 4.4.0 (same machine, same build flags, steady runs).
[4.4.0] - 2026-10-09
ABI bump: REAMER_ABI_VERSION 5 → 6. Breaking. Rebuild against the ABI 6
header.
Removed
reamer_research_run_backtest()(v1) andreamer_research_run_backtest_v2().- The 7-field
ReamerBacktestSummaryandreamer_research_get_summary(). - Vtable slots
on_bar(v1),on_bar_v2andon_bar_v3. - The
ticker_ids[]parameter ofreamer_run_backtest(). A ticker'sticker_idis its array position. REAMER_RESEARCH_ERROR_INVALID_TICKER_ID(3). Other codes keep their values.
Changed
- Every exported function is renamed from
reamer_research_*toreamer_*.reamer_research_run_backtest_v3()is nowreamer_run_backtest(), andreamer_research_get_summary_v2()is nowreamer_get_summary(). ReamerBacktestSummaryV2is renamedReamerBacktestSummary(31 metrics, 248 bytes).reamer_get_summary(handle, periods_per_year, out)takes the Sharpe and Sortino annualisation (252 daily, 52 weekly).<= 0infers it from closed-trade timestamps, as before. reference-cpp'ssummary()and reference-python'speriods_per_yearconfig key pass it through.- Docs:
calmar_ratioistotal_return_pctover relative drawdown and is not annualised. The header andRESULT_JSON_SCHEMA.mdsaid it was. ReamerResearchStrategyVtableis{user_data, on_bar}(16 bytes).on_bartakes the formeron_bar_v4signature:window,window_valid,window_len,ticker_count,positions,exogenous,exogenous_len,out_orders,max_orders.- All 11 exports sit in one version node,
REAMER_RESEARCH_2.0. The 1.0-1.3 nodes are gone. - reference-cpp:
run_backtest<>()takes a requiredticker_namesargument and noticker_ids; the callable takes(window, ticker_count, positions, window_valid, out_orders).run_backtest_v2<>()is removed. - reference-python: the ctypes mirrors follow the new vtable and summary.
on_bardocumentation: a forward-filled row keeps the last real bar'sts, so per-ticker staleness is readable from the window.
Added
reamer_run_backtest_files(): run a backtest over memory-mapped.binfiles, one per ticker, with an optional inclusivestart_ts/end_tsrange. Datasets larger than RAM run. Results are identical toreamer_run_backtest()on the same bars.reamer_write_bin(): write validated, sorted bars to a.binfile.REAMER_RESEARCH_ERROR_BAR_FILE(10): a.binfile is missing, unreadable, lacks its header, is truncated, or is unsorted.bin/reamer-csv-build: converts an OHLCV CSV into a.binfile forreamer_run_backtest_files(). Detects delimiter, columns and timestamp format; prints a parse report (skipped rows, duplicate timestamps, first and last timestamp). Needs no license.reamer_run_monte_carlo(): resamples a completed run's per-trade returns with replacement and reports final-equity, loss, ruin and max-drawdown percentile stats inReamerMonteCarloStats. Seeded, and bit-identical for any thread count. A run with fewer than two closed trades returns all zeros. License-gated.- Result JSON
equity_curveandequity_curve_ts: realised equity after each closed trade, with timestamps. It is the curveequity_drawdown_maximalis measured on, floored at 0.schema_versionstays 1 (fields added).
Fixed
- Result JSON
open_orders_end[]was always empty. An order still pending at the end of a run now appears there asPending, and inorder_log[]asCancelled. reamer_run_backtest(): a ticker passed with zero bars no longer shifts every later ticker down one asset index. Before the fix,on_bar'sticker_id,positions[]and closed-trade names for later tickers pointed at the wrong instrument.
[4.3.2] - 2026-10-03
No change to the C ABI or the Reamer Research binary behaviour. Released
alongside a Reamer Server fix to the license.expiring warning schedule.
Added
LICENSING.md"Renew on the same machine": deactivate the old key, then activate the new one.
[4.3.1] - 2026-09-17
No change to the C ABI. REAMER_ABI_VERSION stays at 5, every entry point and
struct is untouched, and a program compiled against an earlier header links
and runs against this build unmodified. The version moves with the shared
release number (kReamerVersion, one number across both products); the
substantive documentation change this release is on the server side (see the
Reamer Server changelog).
Changed
- The written ABI versioning policy is clarified. The advance-notice and deprecation sections are resolved to the actual policy: support the latest version only, no sunsetting because the surface is additive-only, and a version bump is communicated through the changelog at release. The document is also rewritten in plain technical english with no em or en dashes.
[4.3.0] - 2026-09-13
No change to the C ABI. REAMER_ABI_VERSION stays at 5, every entry point and
struct is untouched, and a program compiled against an earlier header links
and runs against this build unmodified. This release carries new benchmark
evidence and the documentation built on it; the version moves with the shared
release number (kReamerVersion, one number across both products).
Changed
BENCHMARK.mdnow includes a many-core verification section measured on a 64-core AMD EPYC 9575F. The pure-ABI native engine held a per-call latency tail essentially flat across 2000 independent backtests (P99.9 within ~3% of P50) with per-run throughput CV 0.59% (1.04× min→max spread) — the quantitative form of the determinism claim. A stress/capacity pass on the same box reported no bottlenecks, ingesting a 50-million-bar universe (10,000 tickers × 20 years) without failure.
[4.2.1] - 2026-09-12
No change to the C ABI. REAMER_ABI_VERSION stays at 5, every entry point and
struct is untouched, and a program compiled against 4.1.15 links and runs
against this build unmodified. The fix is entirely inside the shipped Python
reference binding; the library and every other reference tree are unchanged.
Fixed
Multi-ticker strategies in the Python reference binding now read each ticker's own bars instead of aliasing every ticker to the first. The ctypes binding hands
on_barone_TickerViewper instrument over the flat, row-major(ticker_count, window_len)window the ABI delivers. Every view was constructed on the same base pointer, so all tickers reported ticker 0's OHLCV; a strategy reading two instruments saw the first one twice. The binding now offsets each view byticker_id * window_leninto the window before constructing it, so each ticker reads its own row. Single ticker runs were unaffected (offset 0), which is why this survived to now.This is a read-side view bug only. The engine's execution path — tick simulation between bars, the synthetic-tick PRNG, spread and slippage randomization, and per-ticker wall-clock alignment of asynchronous bars — was never involved and is unchanged. Fills, results, and RNG draws for any run that did not read a multi-ticker window are byte-identical to 4.2.0.
No change to the C ABI. REAMER_ABI_VERSION stays at 5, every entry point and
struct is untouched, and a program compiled against 4.1.15 links and runs
against this build unmodified.
Fixed
Malformed orders are now rejected at the ABI boundary instead of reaching the engine. Order requests are validated in
ToEngineOrderbefore any field is cast: the four enums (order_type,kind,side,tif) are range-checked, andqty,limit_price,stop_price,take_profitandstop_lossmust be finite. A failing order is rejected with a reason naming the field —qty is not finite,side out of range— through the existing rejection path.This closes two paths that previously produced wrong answers rather than errors. A non-finite quantity on a position reversal bypassed the margin check entirely and filled: the run returned success with a NaN
net_pnland no indication anything had gone wrong. An out-of-range enum matched no comparison in the engine, so the order silently never became fill-eligible and sat pending for the life of the run.Inbound bars were already checked for finiteness at ingest; orders crossing the other way were not. Runs remain byte-identical for identical input: rejection consumes no RNG.
Added
MONITORING.mdships in the kit. For running backtests unattended: what the library writes to stderr and at what volume, the error codes worth alerting on, how to tell a run that failed from a run whose orders were all rejected (a distinction exit status does not make), what to capture for a support report, and what changes when running a grid.reference-cpp/examples/rejected_orders.cpp. A worked example of reading rejections out of a result.summary()andclosed_trades()report what a strategy achieved; only the order log says what the engine refused, and it reaches callers through the result JSON.BacktestResult::result_json()in the reference C++ wrapper, exposingreamer_research_get_result_json(). The full result document — order log, rejection reasons, thereturns[]series — was previously unreachable through the wrapper.
Changed
- The vtable's callback contract is written down.
reamer_research_abi.hnow states what happens when a callback throws (it is caught at the ABI entry point and reported asREAMER_RESEARCH_ERROR_PROCESS_FATAL, but the run ends — there is no per-bar recovery), that callback execution time is deliberately unbounded and uninterruptible, and thatuser_datais passed through untouched.DEPLOYMENT.mdcovers the same ground for operators.
[4.1.13] - 2026-09-05
No change to the C ABI. REAMER_ABI_VERSION stays at 5, every entry point
and struct is untouched, and a program compiled against 4.1.12 links and
runs against this build unmodified.
Added
LICENSE.txtships in the kit. The archive previously carried no licence terms at all. The file sits in the kit root and is covered bySHA256SUMS.Activation requires accepting the terms.
reamer-license activatenow displaysLICENSE.txtand requires an explicitacceptbefore it contacts the licence server, converting activation from browsewrap to clickwrap. Declining consumes no seat, and the acceptance record — key, machine fingerprint, UTC timestamp, licence version, and the SHA-256 of the accepted text — is written next to the activation record only after activation succeeds. Amended terms hash differently and prompt again.For unattended provisioning set
REAMER_ACCEPT_LICENSE=accept. Activation with no terminal and no such variable fails rather than defaulting to acceptance, so existing non-interactive provisioning that activates a licence must set it.reamer-licensereadsLICENSE.txtfrom the kit root, one level abovebin/. Movingbin/out of the extracted tree makes activation fail with a message naming the cause.
[4.1.12] - 2026-09-02
REAMER_ABI_VERSION 4 → 5. Purely additive. Four new exported
functions and one new
optional vtable slot; the library now exports eleven symbols across three
version-script nodes (REAMER_RESEARCH_1.0 / 1.1 / 1.2). Every
pre-existing entry point and struct is untouched, so a program compiled
against ABI 1-4 links and runs against this build unmodified. No
recompilation is forced by this release.
This bump exposes three capabilities that were already implemented in the engine but unreachable through the ABI, which is the product.
Added
reamer_research_run_backtest_v3()— the entry point to write new integrations against. Adds per-ticker execution config, mandatory instrument names, and exogenous data sidecars in one signature.Per-ticker execution config. A sparse
ticker_configs[]array applies distinct commission, slippage, spread, and swap terms per instrument within a single run. A NULL entry falls through to the global config. Resolution order is specified inEXECUTION_SPEC.md§0.rng_seedis deliberately excluded from override: it is forced from the global config when building each resolved entry, so differing costs between instruments can never introduce cost-correlated randomness into the synthetic tick path. Enforced in code, not only documented.Mandatory
ticker_names[]. A NULL array is rejected at this entry point. Consequently every instrument name in a v3 run's results —ReamerClosedTrade.tickerand everytickerfield in the result JSON — is the caller's own symbol, never the synthetic"T%09zu"placeholder the older entry points fall back to. This makes the research→server instrument-identity mapping checkable rather than merely required; seeRESEARCH_TO_SERVER.md§2.Exogenous data sidecars. An optional
exogenous_paths[]array attaches a non-OHLCV series per instrument. Delivered toon_bar_v4.on_bar_v4vtable slot. Extendson_bar_v3with per-ticker exogenous bytes and lengths. The vtable grows 32 → 40 bytes, which is safe only because the caller passes it by pointer, the library reads no slot it does not know about, and the header has mandated zero-initialization since ABI 2. Dispatch falls backon_bar_v4 → v3 → v2 → on_bar, so a strategy that sets onlyon_bargets exactly v1 behavior.Resolution is strictly point-in-time. At each bar the engine resolves the latest entry with
ts <= current_bar_tsand never a later one; a per-ticker cursor advances forward only, keeping this O(1) amortised rather than a per-bar search. Where no sidecar is attached or no entry qualifies yet, the strategy receives NULL and length 0. The bytes are engine-owned, valid only for the duration of the call, and are never parsed or validated by the engine.reamer_research_get_summary_v2()andReamerBacktestSummaryV2— 31 metrics whereget_summary()reported 7: Sharpe, Sortino, Calmar, profit factor, recovery factor, expectancy, win rate, consecutive win/loss streaks, holding-time statistics, absolute and relative drawdown, and full order-outcome counts.ReamerBacktestSummaryis frozen at 56 bytes andget_summary()returns byte-identical values to previous releases. The new metrics went into a separate struct reached through a separate entry point rather than growing the existing one, because growing a caller-allocated struct overruns a buffer sized by an older compiled header — silently.Sharpe, Sortino, and Calmar are annualised by
sqrt(periods_per_year), whereperiods_per_yearis inferred from the median gap between consecutive closed trades'close_timestampvalues, falling back to 252 below two closed trades or on unparseable timestamps.reamer_research_get_result_json()andreamer_research_write_result_json()— the complete run as a single JSON document: summary, closed trades, fills, every order and its outcome, scale-in fills, roll events, and the per-trade return series. Two-call sizing (buf = NULLreports the required size including the NUL); an undersized buffer is rejected rather than truncated. Output is locale-independent by construction — doubles are formatted by the serializer's own Grisu2 path, never the C locale'sprintf— and this is enforced by a test running underde_DE.UTF-8.Replay scaffolding (
orders_by_step,tick_orders_by_ticker,prints_by_step) is deliberately excluded: it carries no information not already inorder_logand would multiply the document size.bin/reamer-exo-build— converts a two-column CSV into the.exo.binsidecarrun_backtest_v3()reads. Reports entry count, duplicates collapsed, and byte size; exits non-zero with a stderr message on unreadable, malformed, or empty input. No license check — it is an offline converter that links no engine code, so gating it would add an activation failure mode ahead of evaluation.EXOGENOUS_FORMAT.md— CSV authoring format, the.exo.binbinary layout, the point-in-time visibility rule, and CLI usage.RESULT_JSON_SCHEMA.md— field-by-field schema for the result document, withschema_versionversioned independently ofREAMER_ABI_VERSION.It documents
returns[]explicitly, because the definition is not the one most consumers assume:returns[i] = net_pnl / equity_at_entry, clamped to[-1, 1]. The denominator is account equity at entry, not the trade's own margin, so it is notclosed_trades[i].return_pct / 100. This is the exact series the Sharpe, Sortino and Calmar figures are computed from — recomputing risk metrics fromclosed_trades[].net_pnlwill not reconcile withsummary.
Both new documents ship in the kit and are enforced by the archive verifier, not merely copied by the build script.
Changed
EXECUTION_SPEC.mdgains §0 (per-ticker execution config resolution, including therng_seedexclusion) and §13 (exogenous data resolution).RESEARCH_TO_SERVER.md§2 — instrument-identity mapping is still required and still yours to build, but arun_backtest_v3()run's results now carry the same string identifiers Server'sintent.Instrumentuses, so the mapping can be asserted as a build step instead of trusted.- The internal ABI versioning policy records the v5 row and the struct-growth hazard that motivated putting the new metrics in a separate struct.
[4.1.11] - 2026-09-02
Changed
- Shipped archive filename renamed for clarity: this kit is now
research-kit-<version>-<platform>.tar.gz(wasreamer-research-<version>-<platform>.tar.gz). Only the customer-facing archive filename moved; the extracted layout is unchanged.