Spec → RTL for codec standards, where the deliverable is evidence, not RTL.
A generated design is accepted only when a tool-driven gate — no LLM in the verdict —
says PROVEN.
Pilot: ITU-T G.711 (A-law / μ-law, encoder + decoder), SystemVerilog, valid/ready.
spec (G.711 PDF) ─► requirements.yaml ─► RTL
│
oracles: ITU-T G.191 STL (C) · spec_model.py (from the tables) · PDF tables (tests)
│
gate: lint ─► synthesize ─► simulate (4-state, STL vectors) ─► formal (unbounded)
│
PROVEN / REJECTED / INCONCLUSIVE (+ failure evidence, SHA-256, mutation coverage)
| Gate | PROVEN — lint clean, 16/16 sim tests, 8/8 formal tasks (k-induction) |
| Oracle vs STL | bit-exact on all 65536 inputs / 256 codes |
| Oracle vs G.711 PDF | Tables 1a/1b/2a/2b rebuilt from printed columns: all 256 decoder codes, all encoder interval interiors |
| Mutation coverage | 180/180 non-equivalent mutants caught, and no stage is the only one that catches a fault |
| Hermes-generated RTL | all 4 modules PROVEN (deepseek-v4.1-flash; 1–2 iterations with gate evidence) |
| Hermes-generated testbenches | all 4 ACCEPTED: sound on PROVEN RTL, 100% mutant kill, no blind fault class |
Also in the loop: a knowledge base of past failures with the edit that fixed them
(pipeline/kb.py, mined from runs/), an A/B experiment that measures whether it actually helps
(pipeline/kb_experiment.py — so far: no measurable benefit, 6/9 either way), and a
self-improvement loop (pipeline/improve.py) that measures, ranks what evidence is missing,
acts, re-measures and refuses to call a round an improvement when a metric got worse.
Reports: claudedocs/phase0-report.md ·
plan: claudedocs/spec-to-rtl-service-plan.md ·
working notes and hard-won rules: CLAUDE.md
Verification approach and workbench UI adapted from hls-rtl.
RTL generation goes through a local Hermes Agent API server (OpenAI-compatible, profile
codecagent, port 8659, upstream LLM configurable — currently deepseek/deepseek-v4.1-flash).
pipeline/generate.py sends requirements + the port contract (never the committed RTL), gates
every candidate, and feeds the gate's failure evidence back for repair.
scripts/hermes_server.sh start # needs a hermes profile named codecagent
printf 'HERMES_API_KEY=...\n' > .env # that profile's API_SERVER_KEY
python3 pipeline/generate.py --health
python3 pipeline/generate.py --module g711_ulaw_enc
python3 pipeline/generate_tb.py --module g711_ulaw_enc # Hermes testbench, judged by soundness + mutant kill rate
python3 pipeline/waves.py --test L1_ulaw_enc # VCD of a testbench run (also in the Waveform tab)
.venv/bin/pip install fastapi uvicorn && scripts/workbench.sh # http://127.0.0.1:8013Requires iverilog, verilator, yosys, SymbiYosys + z3, cmake, Python 3.
git clone --recursive https://github.com/kos2001/codec-agent && cd codec-agent
python3 -m venv .venv && .venv/bin/pip install pytest click pyyaml
cmake -S third_party/stl -B third_party/stl/build -DCMAKE_POLICY_VERSION_MINIMUM=3.5
cmake --build third_party/stl/build --target g711demo cf
python3 golden/g711/crosscheck.py
.venv/bin/python -m pytest tests/ -q # oracle checks
python3 pipeline/gate.py # sign-off gate (~5 s)
python3 pipeline/mutate.py -j 8 # mutation campaign (~3 min)The G.711 PDF is not redistributed — see specs/g711/README.md.