QC 4.1 Light

Español · AI agents: AGENTS.md

The call felt fine. The transcript tells another story.

QC 4.1 Light is an open, evidence-first sales-call diagnostic: one transcript in, one structured report out — with quotes, a breakpoint, corrections, and a speakable recovery line.

Quick start

git clone https://github.com/mgallmur-glitch/qc41-light.git
cd qc41-light
npm install   # optional — for CLI/MCP only
qc41-light demo

The offline demo validates and renders a synthetic report — no API key, no network.

Shareable samples: examples/synthetic-report.md · ES examples/synthetic-report.es.md

Related MCP server: audio-transcriber

Install from GitHub Packages

Package visibility on GitHub Packages defaults to private for user accounts. Make it Public at: https://github.com/users/mgallmur-glitch/packages/npm/package/qc41-light → Package settings → Change visibility.

npm install @mgallmur-glitch/qc41-light
# or globally:
npm install -g @mgallmur-glitch/qc41-light

Requires a GitHub PAT with read:packages (and login to npm.pkg.github.com), or install from the release tarball / clone.

Install surfaces

MCP tools: analyze_sales_call, get_qc41_light_prompt, get_schema.

Skill / harness path

Hermes Agent — one-command install

hermes skills install --yes https://raw.githubusercontent.com/mgallmur-glitch/qc41-light/v0.3.5/SKILL.md

Then invoke it in a new Hermes session:

/qc41-light Analyze this authorized, redacted sales-call transcript and return validated JSON.

The URL installation includes the canonical prompt, report schema, validator, renderer, privacy rules, and IP boundary. Package version: 0.3.5. Report schema version: qc41-light-0.2.

For any other assistant, provide SKILL.md + prompts/analyze-call.md, then validate:

python3 scripts/validate_report.py report.json
python3 scripts/render_report.py report.json > report.md

Adapters: Claude Code · Codex · Hermes · generic — under adapters/.

Eval harness

python3 eval/generate_corpus.py   # 50 synthetic EN/ES calls
python3 eval/run_baseline.py      # reference vs naive baseline

Run the baseline locally and consult eval/README.md for the current reference and naive-score definitions.

What one call returns

Boundary

Useful and bounded. Proprietary QC methodology, taxonomies, scoring formulas and team systems stay out of this package — see docs/IP_BOUNDARY.md.

Privacy

No telemetry in this package. CLI/MCP analysis is local BYOK (env keys). Redact real calls before sending transcripts to any model — docs/PRIVACY.md.

Status

Current package and skill release: v0.3.5. The report schema contract remains qc41-light-0.2 for compatibility. These versions are intentionally separate: package changes do not alter the JSON shape unless the schema version also changes. CLI · MCP · eval are on main.

License

Apache-2.0. See LICENSE.