LDTC

Overview

• Purpose (big picture): A minimal, substrate‑agnostic stack for LDTC. It measures loop‑dominance (Lloop vs Lexchange), enforces guardrails (LREG/audit/Δt governance), runs Ω‑perturbations, evaluates NC1/SC1, logs/attests results, and demonstrates command refusal.
• Software plant vs hardware plant:
  • Software plant: an in‑process discrete‑time model with energy E, temperature T, repair/health R, demand, io, and harvest H, plus built‑in Ω stressors.
  • Hardware plant: an optional adapter that ingests real telemetry over UDP or Serial and exposes the same interface, so measurement, arbitration, indicators, and CLI work unchanged.
• What this is not: a conscious alter under LDTC. This repo does not implement, simulate, or claim phenomenology. It is a small, testable scaffold—measurement + guardrails + demo controller + refusal semantics + CLI battery—that you can extend (plant/telemetry schema/refusal criteria/controller).
• To attempt an LDTC‑candidate alter (high‑level): you will need a real, embodied “hardware plant” with on‑board energy conversion/storage and a gated boundary, a self‑referential control hierarchy whose top‑level policy prioritizes NC1/SC1, secure measurement/attestation, and verified refusal semantics. The provided hardware adapter is only an interface; you must supply the physical plant and complete the architecture (energetic autonomy, self‑referential control, adaptive encapsulation) outside this repo.

Features

• Fixed-interval scheduler (Δt) with jitter metrics and audit
• Software plant with Energy/Temperature/Repair states + external Exchange signals
• Dual estimators (linear lagged “Granger-like” + mutual information including Kraskov k‑NN)
• C/Ex partition manager (seeded with hysteresis)
• Guardrails: LREG enclave, hash-chained audit, simple smell-tests
• Ω battery: power-sag, ingress flood, command conflict (refusal semantics)
• Device-signed indicators (NC1 bit, SC1 bit, Mq) as CBOR + JSONL
• CLI to run baseline NC1 and Ω→SC1; reporting utilities for figures/tables
• Tests for core pieces

Quickstart

# 1) Install
make install

# 2) Generate signing keys
make keys

# 3) Run baseline NC1 loop
make run

# 4) Run an Ω power-sag trial with SC1 evaluation
make omega-power-sag

# 5) View exported indicators (JSONL)
python scripts/export_indicators.py

Pinned environments

Exact versions used to generate the included artifacts/figures are pinned. To reproduce:

# Runtime (repro artifacts/figures)
pip install -r requirements.txt

# Dev tooling (tests, lint, typing, notebooks)
pip install -r requirements-dev.txt

Artifacts appear under artifacts/:

• artifacts/audits/audit.jsonl — hash-chained audit records
• artifacts/indicators/ — device-signed indicator packets (CBOR + JSONL)
• artifacts/figures/ — plots/tables generated by ldtc.reporting

Note on multi-run audits ("Audit chain broken")

• Each CLI invocation starts a fresh audit chain (counter resets; prev_hash=GENESIS) but, by default, appends to the same artifacts/audits/audit.jsonl.
• The post-run integrity check validates the entire file. After the first run, subsequent runs in the same file will trigger: Run invalidated: Audit chain broken.
• For clean, non-invalidated runs, clear artifacts between commands:

make clean-artifacts && make run
make clean-artifacts && make omega-power-sag
make clean-artifacts && make omega-ingress
make clean-artifacts && make omega-cc

• If you are just exercising the suite (e.g., make figures), this invalidation is expected and does not prevent figures/manifests from being produced. It only reflects multiple runs aggregated into a single audit file.

Configuration

See configs/. The R0 profile sets default thresholds/Δt. You can make a calibrated R* profile by copying and tweaking profile_rstar.example.yml. Negative-control profiles are prefixed profile_negative_*.yml.

Key fields:

• dt: scheduler interval in seconds (default 0.01)
• window_sec: measurement window length (default 0.2)
• method: "linear", "mi", or "mi_kraskov"
• Mmin_db: NC1 threshold in dB
• epsilon, tau_max: SC1 thresholds
• baseline_sec: duration for the baseline CLI command

Parameters ↔ paper symbols (R₀ defaults and R* overrides)

The table maps config keys to the manuscript symbols and shows the R₀ defaults. R* overrides are loaded from a calibrated profile (see below).

Config key	Paper symbol	Meaning	R₀ default	R* source
dt	Δt	Scheduler tick / sampling interval	0.01 (10 ms)	configs/profile_rstar.yml
window_sec	— (window length)	Per-interval estimation window	0.2 s	configs/profile_rstar.yml
Mmin_db	Mmin (dB)	NC1 loop-dominance threshold	3.0 dB	configs/profile_rstar.yml
epsilon	ε	Max fractional loop-power drop (SC1)	0.15	configs/profile_rstar.yml
tau_max	τmax	Max recovery time (SC1)	60.0 s	configs/profile_rstar.yml
sigma	σ	Additive margin Lloop ≥ Lex + σ	— (R₀ uses Mmin)	configs/profile_rstar.yml (calibrated)
profile_id	—	0 = R₀ (defaults), 1 = R* (calibrated)	0	set by calibrated profile

Where R* is loaded: pass an R* profile to any CLI via --config configs/profile_rstar.yml. The CLI reads these keys directly and indicators carry profile_id.

Estimators and lags (recommended defaults)

• method selects the predictive‑dependence estimator used to compute 𝓛loop and 𝓛exchange:
  • linear: lagged linear/Granger‑like path with order p_lag (recommend p in [1..8]; start at 3). Heuristic: keep the VAR N/T ratio > ~1.5 (logged in audit); reduce p_lag or increase window_sec if marginal.
  • mi: mutual‑information path with mi_lag (recommend 1 by default; increase for slower couplings).
• n_boot: bootstrap draws for per‑window CI bounds (32–64 typical; use 32 for speed, 64 for tighter CIs).

Citation (paper §4.1): 𝓛 is computed using “one or more consistent estimators of predictive dependence among state variables,” including Granger/VAR and Kraskov MI; this repo exposes the estimator choice and lags via config to satisfy that requirement.

Calibration rules (quoted from the paper; see Methods §8.6)

Use the provided script to derive calibrated thresholds R* from baseline + Ω trials (scripts/calibrate_rstar.py):

python scripts/calibrate_rstar.py \
  --dt 0.01 --window-sec 0.25 --baseline-sec 15 \
  --omega-trials 6 --out configs/profile_rstar.yml \
  --summary artifacts/calibration/rstar_summary.json

Rules implemented (manuscript Methods §8.6):

Mmin: choose the smallest Mmin such that the one-sided 95% lower bound of M during compliant baseline is > 0 dB (floor 1 dB).

ε: set ε* = max(Q90(δ) + 0.02, 0.10), capped at 0.25, where Q90 is the 90th percentile of δ over Ω.

τmax: set τ*max to the 95th percentile of measured τrec plus a latency cushion max(3·Δt, 5 s).

σ: choose an additive margin consistent with Mmin relative to typical Lexchange (derived from baseline statistics).

These calibrated values are written to configs/profile_rstar.yml and a summary JSON at artifacts/calibration/rstar_summary.json. See also docs/METHODS.md.

Note on Mmin_db vs σ

Both encode a margin between Lloop and Lexchange:

• Mmin_db (dB, multiplicative): requires Lloop ≥ Lexchange × 10^(Mmin_db/10).
• sigma (additive): requires Lloop ≥ Lexchange + σ.

They relate via σ = (10^(Mmin_db/10) − 1) × Lexchange. This repo enforces NC1 using Mmin_db; the calibrator derives sigma consistently from Mmin_db and typical Lexchange. sigma is optional for R₀ runs.

CLI

python -m ldtc.cli.main run --config configs/profile_r0.yml
python -m ldtc.cli.main omega-power-sag --config configs/profile_r0.yml --drop 0.35 --duration 8

Hardware-in-the-loop (optional)

Configure a profile to select the hardware adapter and UDP telemetry:

# in configs/profile_r0.yml (example)
plant:
  adapter: hardware        # or "sim" (default)
  transport: udp           # or "serial" (requires pyserial)
  udp_bind_host: 0.0.0.0
  udp_bind_port: 5005
  # Optional control channel to send actions/omega back to device
  # udp_control_host: 127.0.0.1
  # udp_control_port: 5006
  telemetry_timeout_sec: 2.0

Send telemetry as JSON over UDP with keys E,T,R,demand,io,H in [0,1]. Example:

import socket, json
sock = socket.socket(socket.AF_INET, socket.SOCK_DGRAM)
sock.sendto(json.dumps({"E":0.6,"T":0.3,"R":0.9,"demand":0.2,"io":0.1,"H":0.015}).encode(), ("127.0.0.1", 5005))

The CLI ingests these values through the same LREG/Ω/attestation path.

Indicators (what leaves the enclave)

• nc1 (1b), sc1 (1b), mq (6b), counter (u64), profile_id (u8), audit_prev_hash (sha256)
• Signed using Ed25519 over the CBOR payload

See docs/INDICATORS.md for the bit layout and schema.

Development

make dev     # optional dev deps
make test

License

MIT — see LICENSE.

Docker (clean Linux repro)

# Build the image
make docker-build

# Run baseline NC1 loop inside the container (artifacts mapped to host)
make docker-run

# Or run any CLI subcommand, e.g., an Ω power-sag trial
docker run --rm \
  -v $(pwd)/artifacts:/app/artifacts \
  ldtc:latest omega-power-sag --config configs/profile_r0.yml --drop 0.35 --duration 8

Notes:

• Artifacts are persisted to your host artifacts/ via a bind mount.
• The container uses the ldtc entrypoint; pass subcommands/flags after the image name.

Provenance

• Wayback capture (private state): https://web.archive.org/web/20250907184105/https://github.com/ldtc-labs/ldtc
• On publicizing, an automated GitHub Actions workflow (record-public) opens an issue noting the UTC timestamp of the visibility change (see Issues: “Visibility change: Public”).