claude
9eb6d6b4f6
Slice 8, closing the #319 retirement's prose and bench debt. Bench argv rides exec (fixed_cost single-run probe, campaign surfaces, and the seed helper that still called the retired sweep at runtime — caught by the bench run itself); all five fingerprints unchanged against the committed baselines, confirming record-line parity through the retirement. Live docs (README, authoring guide, project layout, glossary) describe only exec + campaign documents + graph introspect --params; three glossary entries explicitly mark their verb retired. Ledger amendments per the C29 discipline — C25's executor-verb-set re-settled (exec + the --override residue), C14's dual grammar reduced to the one dispatch, C24's document-built runs/families + gated-intake route list (lockstep with the code comment), C12's override clause and identity anchor repointed to the runner-layer hash computation, C18/C22/C27/C01 mention rewrites — superseded sentences moved verbatim to the history sidecars (c25's created). Stale dual-grammar doc comment on is_blueprint_file rewritten. refs #319
4.0 KiB
C1 — Determinism and disjoint parallelism
Guarantee. A backtest is a deterministic, synchronous, non-concurrent event
loop that reaches a unique state after each input tick. Same input (incl. seed)
→ bit-identical run. Two backtests are fully disjoint and run concurrently
without locking. The bit-identity is per run: one backtest of given (inputs,
seed) reproduces byte-for-byte. It does not extend to a derived metric
recomputed for the same params by two different command paths (e.g. a
campaign cell's sqn within its sweep stage vs the same params re-run
through exec's blueprint leg with --override, #319), which may differ
by floating-point reassociation (≤1 ULP) because the two paths accumulate the
same logical reduction in a different operation order (IEEE-754 non-associativity).
C1 governs the determinism of a single run, not the cross-command bit-identity of
a re-derived statistic. (fieldtest 0078, ratified.)
Forbids. Concurrency within a single sim; any nondeterministic input that is not captured as an explicit input (see C11, C12).
Why. Real money rides on backtest results; reproducibility and an audit trail are non-negotiable. Speed comes from parallelism across sims, which disjointness makes lock-free.
Current state
The per-run event loop is Harness::run (crates/aura-engine/src/harness.rs):
each cycle picks the live source head with the smallest (timestamp, source index), pops one record, increments cycle_id, and forwards the value —
allocating nothing per cycle beyond a reused scratch buffer. A single run is
fully sequential, so its output is bit-identical by construction; the
end-of-stream finalize pass runs after the loop and adds no within-sim
concurrency.
Cross-sim parallelism spans the sweep/member/window fan-out and campaign cells
(#277). The campaign executor execute (crates/aura-campaign/src/exec.rs)
flattens the (strategy, instrument, window, regime) cell matrix into
document-order PlannedCells, groups them by instrument ordinal, and walks
sequential chunks of K instrument groups — --parallel-instruments
(DEFAULT_PARALLEL_INSTRUMENTS = 4, crates/aura-campaign/src/lib.rs). K is a
structural bound on distinct resident instruments — the RAM lever for the
external data-server's per-reference file retention (~1.4 GB per instrument on
the reference dataset), not a pool-worker count. Within a chunk, cells run
concurrently via rayon::par_iter on the process-global pool shared with the
member/window fan-out, and results are written into document-order slots, so
outputs stay byte-identical across worker counts and bounds.
Two scheduling-dependent carve-outs sit outside the per-run bit-identity, which
governs successful runs. On the run-fatal path (non-containable faults, e.g. a
dead registry store) a shared AtomicBool abort flag latches, and the propagated
fault is the lowest document-order fault among the cells that completed before
the latch; the set of per-cell family lines already written by then is
scheduling-dependent — inert, because no campaign-run record is written on that
path and store reads are name-keyed, never line-ordered.
Duplicate campaign instruments are refused at both the validate tier
(crates/aura-campaign/src/lib.rs, a seen_instruments scan) and the executor's
preflight: the per-cell registry family name embeds the raw instrument string,
so name uniqueness is what keeps concurrent appends from racing one name's
run-index assignment.
See also
- C4 — tie determinism preserves this contract; the source-index tie-break is what keeps the merge order deterministic.
- C11, C12 — recorded/pinned inputs are how nondeterministic sources become explicit inputs; the harness is the atomic disjoint unit.
- C23 — the raw-index compilation primitive; names survive only as non-load-bearing debug symbols.
History: c01-determinism.history.md