HFT Backtest is the deterministic replay/simulation consumer of canonical CXET events, recorder corpus contracts and trading-core semantics. It builds only as part of the CXET direct-source root graph.
The product includes hft-backtestd and hft-backtest-tui. The daemon owns
experiments, independent strategy-variant portfolios and persistent analytics;
neither a terminal nor Vetron is needed while experiments run. The terminal is
a client: closing it never stops recording or experiments.
Recorder publishes compressed blocks only after a durable, producer-complete capture cut. Backtest normally obtains those bytes through the Recorder local client and decodes into bounded RAM. An evicted block is read from the same compressed archive. Protocol or CRC failures are errors, not fallback signals. Original receive monotonic/UTC stamps and event identity determine simulation ordering; disk speed and UI polling never supply the simulation clock.
From the CXET checkout root, start the successfully selected programs explicitly (these are examples, not verification results; use separate terminals):
python3 tools/Build/ActiveBuild.py --root . --owner hft-backtest --exec hft-backtestd -- --root /absolute/private/research --socket /absolute/private/research/control.sock --build-id BUILD --module-id MODULE
python3 tools/Build/ActiveBuild.py --root . --owner hft-backtest --exec hft-backtest-tui -- --research-socket /absolute/private/research/control.sock --recorder-runtime /absolute/private/recorderEach entrypoint resolves the successful product marker under the shared build lock and retains it for its process lifetime.
The roots must be real owner-private directories. Build/module identity must describe the deployed implementation; clients obtain the admitted values from the service catalog. Saved experiments retain their original provenance. Profiles and exact decimal parameter ranges expand to a complete bounded grid; every variant has its own portfolio and the same explicit scenario seed. Account currencies, fees, host budget and valuation requirements are explicit. Unknown instrument or collateral facts refuse new risk. Unsupported financial representations, including missing currency valuation, are never synthesized.
Enabled experiments restore automatically. Recovery replays the saved input
from the original start into a new immutable attempt generation, then switches
the published result manifest after catching the prior visible cut. Existing
generations are not appended with duplicate replay rows. Analytics include
qualified financial rows, executions, interval returns and distributions, with
input position and account/currency provenance. Data-only C++ modules use the
same scheduling and storage, beginning with midprice_return_distribution.
Capture loss stops affected experiments as incomplete. User pause retains state and catches up from the archive on resume; a native risk stop is terminal for that variant. Stop does not invent liquidation fills or discard open inventory. RAM, source/context, variant and disk admission remain bounded. Data is not automatically deleted to make a quota fit.
Saved specifications consume bounded catalog metadata. Running and retained
paused experiments reserve simulation resources; completed experiments release
those resources while their analytics remain readable. Saved jobs from another
build/module remain available through Describe and result queries, with
execution disabled until their exact implementation is deployed.
The daemon's cooperative work quantum is selected with
--work-items-per-step (1–64). --max-state-bytes,
--max-resident-variants, --max-output-bytes and --max-experiments bound
the service. One cooperative owner advances admitted experiments. Saved profiles
and expanded variant labels can be inspected through Describe (:describe
in the TUI) after reconnecting. An explicit clone rebuilds the complete grid
in a separate draft.
New C++ studies register a DataStudyDescriptor with declared input channels,
parameters, outputs, a state-size bound and an owned prepare/observe/destroy
binding. They do not receive an order or execution API.
Vetron's Research section uses the same producer-owned local clients. Its
optional --recorder-service-runtime-dir and --backtest-research-socket
arguments attach to existing services without launching them. Live trading
authority remains separate.
From this directory in the canonical CXET checkout:
./compile.sh p
./compile.sh --force portable p
./compile.sh all p
./compile.sh all portable pBuilds use pinned Clang, GNU Make, Release -O3, native CPU targeting and
LTO OFF. portable disables CPU targeting; --force selects FULL LTO without
cleaning. Needed missing or stale dependencies rebuild automatically, including
without a terminal. --accept-dependency-updates remains accepted but redundant.
Ordinary products share <CXET>/build/product across owners. CPU, hot, exchange
and LTO changes reconfigure that cache; omitted options select defaults.
CXET_BUILD_DIR is the exact product-directory override. Fundamentally
incompatible identities reset only a verified selected product leaf before a
fresh build; unsafe or foreign paths refuse reset. Other build namespaces and
ordinary compilation failures do not trigger cleanup.
all builds and runs only this owner's registered tests and needed dependencies.
It uses <CXET>/build/tests/<owner> without configuring or activating a product.
--force all builds the product first and publishes its selection only after
all selected stages succeed. Other owners' tests are not selected; an empty
registry does not establish passing test proof.
Source-built launchers require a successful product marker and retain the
checkout-wide shared build lock through their owned runtime descendants.
A build or launch facing a busy lock refuses immediately; status/stop remain
available. Reset or code-generation-profile changes invalidate markers pointing
at the tree; a failed rebuild leaves those products unavailable. Old trees are preserved
until separately managed. See the root build entrypoints
and launch exclusion for the full
reset, activation and failure rules. p selects memory-aware parallelism;
positive -j N overrides it. See --help for supported options.
The cpp workflow overlays the triggering Backtest revision into the private
CXETCPP root checkout for direct-source family integration. The retired offline
CTest suite is not current acceptance evidence. The current registered owner
offline suites check quota, captured replay and canonical order evidence; their
local results are separate from CI. The badge is a workflow link, not evidence
of current passing CI.
The root checkout requires
the repository secret CI_CXETCPP_SSH_KEY; fork pull requests without that
secret fail explicitly before checkout and do not provide a passing CI result.
Run and sweep artifacts contain financial/simulation results without stage,
wall/CPU or depth-read profiling. Simulated execution latency, replay timestamps,
fills, equity, drawdown, progress and bounded sweep worker scheduling remain.
The public Backtest ABI is the one current contract shared with Recorder and
the CLI. Finite run.result artifacts include JSONL streams
fills.jsonl, equity.jsonl and order_lifetimes.jsonl. Continuous research
uses the current immutable attempt/segment layout.
Binary corpus runs stream checked compressed blocks directly into bounded RAM. An empty corpus source selection discovers eligible contexts over the recording folder: single-source strategies run each source, canonical cross-venue strategies group the same instrument, and explicit groups come from the strategy configuration. Account balances and quotas are shared across its instruments. The whole run requires one common exact captured quote asset; unknown or mixed currencies are refused before strategy admission. Account totals use this same unit without conversion. Missing canonical execution/reference inputs are refused.
Recorded monotonic arrival determines when data becomes visible to the strategy; exchange timestamps remain market evidence. Source lifecycle boundaries suspend affected contexts and preserve their orders and positions. Stream decode creates no unpacked file. The preparation memory guard is not an RSS or throughput proof.
This adapter requires a recorded canonical executable quantity grid, exact quantity-conversion authority, typed economic identity and the descriptor's reference/financial inputs. Admission follows those properties rather than a venue-name allowlist. Unsupported inverse units, nonunit price-basis accounting, descriptor capabilities or rule changes fail before admission or at their captured boundary. A source pause retains venue orders; real data from a healthy sibling source can still fill its existing venue-active orders.
Whole-universe replay uses the binary corpus's exact quote identity. Configured sources without market rows keep paused contexts while healthy sources proceed. Legacy JSON archives remain structurally readable, but executable replay requires the current binary quantity authority. Multiple legacy sessions also lack authoritative common quote identity and refuse before balance aggregation.
Spot simulation owns shared quote cash and working-buy reservations per venue account, and base inventory and working-sell reservations per captured source. It refuses unfunded buys and naked sells. Fills debit or credit principal and exact Core fees; linear Futures/Swaps retain the PnL wallet model. Local pending reservations are immediately known, while venue ACK, fill, rejection and cancel facts advance observed cash, inventory and terminal releases only at modeled private delivery. Shared-account private delivery preserves venue event order across sources even when latency samples tie; payload timestamps are unchanged.
Depth execution consumes finite recorded liquidity across orders until the affected price level is refreshed; unrelated updates do not replenish consumed levels. Recorded trades consume existing depth; trade-through makes better levels unavailable. Finite execution uses authoritative depth without cached BBO overlays. An accepted empty depth book remains finite. Resting maker orders fill only from eligible recorded trades, sharing each trade's finite quantity across orders. Partial-fill private reports carry the cumulative executed quantity and weighted average execution price.
BBO execution is explicitly optimistic: without execution depth, including
BboNoDepth and depth-failure fallback, eligible ordinary Market, marketable
Limit and triggered Stop orders fill their full remaining quantity at the best
price regardless of displayed quantity. It does not model
market impact or finite top liquidity. Actual BBO fills are counted and produce
a result warning; a declared-depth failure retains its separate diagnostic even
if no fallback fill occurs. Depth execution remains quantity-limited.
Spot equity includes marked positive base inventory. During a source pause it uses the last genuine bid received on each actual/observed plane as a historical valuation; this cache supplies no quote readiness or fill liquidity. Service unwind on Spot and linear Futures/Swaps retains its stricter requirement for a genuine opposite venue price and sufficient visible quantity; missing liquidity retains exposure and refuses continuation. Aggregate wallet results sum actual source cash contributions with each account seeded once.
Warmup command timestamps describe local sends. ACK timestamps and warmup settlement become visible only on correlated private delivery. Service exposure tracks only delivered tagged warmup/unwind fills, separately from ordinary strategy inventory. PnL curves include actual per-source unrealized marks; final aggregate snapshots sum quote-denominated PnL without a cross-symbol price. Drawdown starts at the initial zero-PnL baseline, including a loss in the first observation. Portfolio metrics synchronize legs by timestamp using each leg's latest observation at that timestamp, with zero before its first observation.
The bounded canonical simulation evidence owner retains at most 1024 orders per source. Continuous sessions reclaim terminal records only after final private delivery, actor consumption and release of every borrowed reference; exhaustion refuses further admission. Implementation and fixture evidence is static-only. Builds and tests have not been run for these changes; no performance, runtime or live-trading result is established.