TrueLoop Compute: reproducible headline results

Evidence package for “Minimal sufficient control of component-observable hardware for hybrid digital–analog computing.”

The paper’s systems claim is specific: a digital host can delegate a high-arithmetic transform to component-observable analog hardware without taking back a growing calibration problem. The maintenance layer uses one parallel component-resolved acquisition and one complete actuation frame per cycle, with linear controller work and bounded state per channel. It does not claim that analog hardware replaces digital control, that scalar-feedback devices satisfy the contract, or that an end-to-end hybrid AI product has already been demonstrated.

• Paper and permanent archive
• Final-manuscript evidence map
• Pinned repository states
• Recovery scope and known gaps
• Machine-checkable archive validation

Evidence in this repository

The original three hosted-runtime campaigns remain at the repository root and in results/:

Experiment	Entry point	Headline result
Variational parameters under tight shot budgets	exp1_variational_shot_budget.py	At 12 acquisitions the runtime reaches 1.5× the sampling floor; finite differences complete no update.
Photonic mesh phase hold	exp2_photonic_mesh_hold.py	Hold error remains approximately 0.055 from 512 to 1,000,000 channels; the 1,000,000-channel model recovers from a full scramble in one acquisition.
Phased-array calibration hold	exp3_phased_array.py	Beam-gain loss remains approximately −0.27 dB across 256–4,096 elements while scalar-feedback control approaches the uncalibrated floor.

manuscript_2026/ adds the two final pre-registered campaigns and the independent public implementation disclosed in the paper:

Final figure	Public source	Headline result
Fig. 8, independent scale ladder	manuscript_2026/recovered/1_scale_ladder/ and 3_shared_harness/tlref.py	The public reference law reaches usability in four acquisitions and holds a 0.0496 floor from 100,000 to 100,000,000 channels. Diagonal Broyden reaches usability in three and has the lower 0.038 floor; this in-class result is reported without qualification.
Fig. 9, analog compute stack	manuscript_2026/recovered/2_analog_stack/ and exact runners in transcript_extracted/	In a depth-8, width-1,024 stack, retained control preserves 0.786 elementwise and 0.754 structured top-1 agreement under drift, above both free-running and the registered oracle-recalibration arm.

The recovery directory preserves disk originals and verbatim transcript extractions. The historical PROVENANCE_LEDGER.md is retained unchanged; the final-manuscript renumbering is corrected separately in FIGURE_TO_SOURCE.md.

Reproduce the root experiments

python3 -m pip install -r requirements.txt
export TRUELOOP_KEY=...        # free key: https://trueloopcompute.com
python3 exp1_variational_shot_budget.py
python3 exp2_photonic_mesh_hold.py core frontier
python3 exp3_phased_array.py

All scripts checkpoint into results/*.json and resume on rerun. run_all.sh executes the free-key-compatible set. The archived large-n hosted-runtime cells above the public service limit remain verifiable from their result records.

Reproduce the public reference-law ladder

The independent reference implementation is not the proprietary hosted or offline runtime. It is a direct implementation of the disclosed law and auxiliary-state structure.

cd manuscript_2026/recovered
PYTHONPATH=3_shared_harness python3 1_scale_ladder/ladder.py 100000 31 ref
PYTHONPATH=3_shared_harness python3 1_scale_ladder/ladder.py 100000 31 broyden
PYTHONPATH=3_shared_harness python3 1_scale_ladder/ladder.py 100000 31 nothing

The exact launch record for every published rung is in transcript_extracted/ladder_launch_commands.sh. The 100-million-channel campaign used the final checkpointed runner ladder8_ckpt.py; it requires substantial memory, storage and runtime and should not be launched casually.

Verify the archive

python3 manuscript_2026/validation/verify_archive.py

The verifier checks the two preregistration anchors, parses every retained JSON/JSONL record, validates the final plotted aggregates, checks current figure hashes, compiles all Python sources, validates recovered shell syntax, and scans tracked content for common active-secret patterns.

What is public, and what is not

Public here:

• the disclosed reference law and bounded per-channel auxiliary state;
• plant, baseline, campaign, metric and figure-generation sources;
• committed preregistrations, seeds, thresholds and raw/final result records;
• exact immutable commits for the earlier hardware and simulation campaigns;
• recovery provenance and explicit missing-artifact disclosures.

Not public here:

• the proprietary hosted/offline runtime or its undisclosed refinements;
• deleted 100-million-channel checkpoint arrays (their sizes and SHA-256 hashes survive in the recovery ledger);
• an original machine-readable 100,000-channel per-seed file, because that rung was emitted to subprocess stdout and only its aggregate was persisted;
• physical end-to-end hybrid AI hardware results, which the manuscript does not claim.

License

MIT for repository source and data; see LICENSE. The proprietary TrueLoop runtime is a separate service; see NOTICE.