Autonomous research system for measurable, computer-executable research.
Rocket Booster Recovery (Rust) Praxist task
This directory is the Rust edition of the high-concurrency Rocket Booster Recovery server configuration and can be launched as a Praxist task harness. The directory name is retained for task compatibility. The evaluator, candidate-controller API, 6DoF simulation, metric computation, and resource observer run on CPU in Rust; they do not use Python, NumPy, JAX, XLA, CUDA, a GPU, or a virtual environment.
The research policy remains 16 peers over 30 generations, DIG in Gen0 only, QD in every generation, a 75% constructive-mix soft constraint, no online literature search, and multi-metric Pareto retention in the frontier and incubator.
Frozen Baseline And Success Standard
The baseline under assets/baseline/ contains controller.rs,
controller_config.json, and variant.json. Every research candidate must
retain those three root files, but may add local .rs modules. Candidates may
not add a Cargo manifest or third-party dependency.
The complete protocol contains 12,288 landing trajectories across three strata and 1,024 roll-disturbance trajectories. The joint success predicate is evaluated at interpolated first landing-leg contact. It requires no more than 5 m lateral error, 1 m/s center-of-mass and lowest-contacting-leg sink speed, 0.3 m/s lateral speed, 1.5 degrees tilt, 0.02 rad/s absolute roll rate, and 0.03 rad/s pitch/yaw rate norm, with strictly more than 2% of the initial 7,000 kg main fuel remaining. Post-contact spring and damper response receives no score credit.
The Rust complete baseline reproduces 495 / 12,288 = 4.0283203125%.
Complete evidence and the compact ledger are under assets/baselines/; they
are baseline assets, not historical run records. Runtime-created
experiments/ and scratch/ directories are excluded from version control.
Rust Toolchain
Rust 1.85 or newer and a native cargo on PATH are required. The root
vendor/ directory and Cargo.lock support --locked --offline builds
without a host-private Cargo cache. The Praxist orchestrator runs in its own
environment; this task has neither a simulator venv nor a task venv. A shared
server may select a managed Rust toolchain through the ordinary shell
environment before Praxist starts, but machine-specific absolute paths do not
belong in the task harness.
Build and verify from the example root:
cargo --version
cargo build --release --workspace --locked --offline
cargo test --release --workspace --locked --offline
cargo run --release --locked --offline \
--manifest-path task_GPU_server/Cargo.toml \
--bin rocket-booster-recovery-task-eval -- \
--variant-dir task_GPU_server/assets/baseline \
--mode complete \
--output-dir task_GPU_server/scratch/manual-baseline-complete \
--threads 16
The four evaluator modes are canary, development, roll-diagnostic, and
complete; Praxist normalizes the roll evidence label to roll_diagnostic.
Only a complete 13,312-unit result may mature, serve as a parent, or support a
normal generation close.
A complete, protocol-clean result sets promotion_eligible=true, making it
eligible for durable evidence, incubator retention, and generation-close
counts. This does not make it a confirmed champion. The confirmed lane also
requires confirmed_performance_gate_passed=true: overall, hard-OOD, and
worst initial-radius-stratum success rates must all be nonzero. A complete but
weak result can therefore remain auditable evidence and a Pareto parent
without masquerading as a strict champion or blocking the maturity quorum.
Validate And Launch With Praxist
resolve_dir="$(mktemp -d)"
praxist resolve "$PWD/task_GPU_server" --run-dir "$resolve_dir"
praxist start --task-path "$PWD/task_GPU_server" \
--cohort 16 --generations 30 --daemonize --json
Provider configuration, agent runtime, and credentials belong to the operator
environment and are not stored in the harness. Every formal peer evaluation
must launch through the central scheduler's cpu_evaluation profile using a
supported command template from the prompt. The candidate static audit is not
an operating-system sandbox; do not run untrusted candidate code on a host
with sensitive credentials or mounts.
Candidate Isolation
A peer may modify only the three root files in its run-owned variant and
optional local .rs modules. The frozen launcher parses and content-addresses
the full Rust source tree, performs TOCTOU verification, rejects unsafe,
custom #[path], file/network/subprocess access, dynamic includes, learning
frameworks, and evaluator/data references, then compiles the candidate in an
isolated Cargo cache under scratch/. The generated crate provides Serde
derives directly. Candidates may use DynamicControllerConfig or
VariantConfig<LocalParams> to add a small set of mechanism parameters under
variant_params without redeclaring roughly one hundred frozen fields. The
generated runner can call only the task's limited state, math, and control
output API. The frozen harness always owns the plant, data selection, contact
interpolation, success gate, and evidence classification.
Generation synthesis starts after the full 90-minute formal peer research window, aligned with the validated Python harness clock. Early findings, contributing peers, or mature-result quorum do not remove time reserved for implementing, falsifying, ablating, and completely reevaluating complex mechanisms. Gen0 DIG occurs before this timer and is excluded from the 90-minute interval. These clock settings affect orchestration only; they do not change trajectories, success gates, or baseline metrics.
Fair-Evolution Boundary
This task contains no prior run, peer workspace, champion package, historical
score trajectory, or cross-language winning-mechanism prior. Researchers may
use only the frozen Rust baseline, generic directions bundled with this task,
and evidence created after the current run starts. They must not search for,
read, translate, or reuse old controllers, configurations, or reports, nor use
Git history, adjacent checkouts, or external paths to recover answers. The
complete policy is in assets/research_independence_policy.md.
Every variant.json must contain four research_independence declarations,
all set to false; the frozen evaluator rejects missing or non-clean
declarations. The declaration must agree with session-access evidence. If a
PI or Chair detects access to old results, the candidate and descendants must
be marked as leakage and excluded from parenting and promotion. This fairness
boundary is independent of multi-file Rust candidates, variant_params, the
pinned Cargo toolchain, and the two infrastructure retries retained for first-
run interface or compilation-environment failures.