____
/ ___|___ _ __ ___ ___ _ __ __| |
| | / _ \| '_ \ / __/ _ \| '__/ _` |
| |__| (_) | | | | (_| (_) | | | (_| |
\____\___/|_| |_|\___\___/|_| \__,_|
Research says 87%. Marketing says 95%. Sales promises 99.9% uptime. Engineering measured 98.2%. Concord finds the walls between roles before your customers, regulators or press do.
Quick start · What it finds · Example · Role briefs · How it works · Limits
At every AI company, a launch is described by nine different teams in nine different documents:
| Team | Writes | Cares about |
|---|---|---|
| 🔬 Research | eval reports | what the model actually scores |
| 🛠️ Engineering | readiness notes | measured latency, uptime, capacity |
| 📦 Product | PRD | scope, price, date |
| ⚖️ Legal | review memo | what creates liability |
| 🛡️ Security | audit notes | what is certified vs. still open |
| 📣 Marketing | launch blog | the boldest true-sounding sentence |
| 💼 Sales | deck | what a customer can be promised |
| 🎧 Support | FAQ | what customers will ask |
| 👔 Executives | memo | the date and the story |
Nobody reads all nine side by side. So numbers drift, dates diverge, a certification that is "in progress" becomes "certified", and "never hallucinates" ships next to an eval that says 3.1%. This is not a money problem — it is a communication-wall problem, and more budget does not read documents for you.
Concord reads what every role wrote and shows exactly where they disagree, who is on each side, how risky it is, and what each role must do about it — in seconds, 100 % offline, in two Python files with zero dependencies.
python concord.py demoNeeds only Python 3.8+. No pip install, no API key, no internet.
git clone https://github.com/theinsidecoder/concord.git
cd concord
python concord.py selftest # 31 checks - verify everything works
python concord.py demo # a launch seen by 9 teams
python concord.py analyze ./launch-docs # YOUR documents (folder, files, or .json)
python concord.py brief --role legal ./launch-docs
python concord.py serve # local web dashboardDon't want to clone? Download just
concord.pyandconcord_brain.pyinto one folder.
| Finding | Example caught in the demo | |
|---|---|---|
| 🔢 | Conflicting numbers (understands over, under, 50+, ≈, $, %, ms, s, counts) |
Research 87.4% vs marketing over 95% accuracy |
| 📅 | Conflicting dates for launch / beta / deadlines | March 3 vs March 14 vs March 21 — four teams, three dates |
| 🚦 | Status contradictions (done vs. in progress) | Blog: "SOC 2 certified" — Security: "still in progress" |
| 💥 | Claims your own data contradicts | "Never hallucinates" vs. eval hallucination rate of 3.1 % |
| 🚩 | Absolute claims with no evidence | "most accurate model ever built", "guaranteed", "fully secure" |
Risks · ❓ open questions · ✅ decisions · 📌 actions (with @owners) |
Collected from every document into one place | |
| 🧱 | Wall map | Which pairs of teams contradict each other the most |
| 📊 | Alignment score | 0–100 with verdict: ALIGNED / NEEDS ALIGNMENT / MISALIGNED |
It deliberately ignores targets/goals (a PRD target of 90 % is not a conflict with a measured 87 %) and treats compatible values as compatible (under 200 ms vs 150 ms is fine).
Real output of python concord.py demo:
Concord 1.0.0 - cross-team consistency check
==========================================================================
Alignment : ███░░░░░░░░░░░░░░░░░░░░░ 11/100 MISALIGNED
Findings : critical 7 high 4 medium 5 low 4 info 4
WALL MAP - where teams contradict each other most
marketing <-> research █████████ 3
engineering <-> marketing ██████ 2
marketing <-> product ██████ 2
CRITICAL F001 Conflicting numbers for “nimbus accuracy”: 87.4% vs ≥ 95%
[research] research_eval_report.md: On our held-out benchmark, Nimbus-2 reaches 87.4% accuracy.
[marketing] marketing_launch_blog.md: Nimbus-2 is the most accurate model ever built, with over 95%…
-> Agree on ONE number and name its source of truth.
CRITICAL F002 Conflicting numbers for “latency p95”: 420 ms vs ≤ 200 ms vs 410 ms
CRITICAL F004 Conflicting numbers for “uptime sla”: 99.9% vs 98.2%
CRITICAL F006 Status disagreement on “soc type ii”
CRITICAL F007 Claim contradicted by internal data: “Nimbus-2 never hallucinates.”
HIGH F008 Teams disagree on the launch date: March 14 (product) vs March 3 (executive, marketing) vs March 21 (engineering)
Every role gets its own pre-launch brief — what is yours to fix and what affects you:
python concord.py brief --role legal ./launch-docsConcord brief for: LEGAL
Which statements create liability, and what must be substantiated?
CRITICAL [AFFECTS YOU] Claim contradicted by internal data: “Nimbus-2 never hallucinates.” - align with: marketing, research
-> Remove or correct the claim - the company's own data contradicts it.
HIGH [AFFECTS YOU] Absolute/superlative claim with no evidence: “fully secure” - align with: support
- Files / folders of
.md,.txt,.rst— folders are scanned recursively. - JSON: a list like
[{"name": "blog.md", "role": "marketing", "text": "..."}]. - Which role wrote it? Concord figures it out from, in order:
--role/JSONrole→ front-matter (role: legalin the first lines) → filename (legal_review.md,sales_deck.md) → vocabulary. - Export Google Docs / Notion / Confluence pages to Markdown or text and drop them in a folder.
| Command | Purpose |
|---|---|
analyze PATH... |
Full report. --format text|json|md|html, --out FILE, -v for low/info items |
brief --role ROLE [PATH...] |
One role's brief |
demo |
Built-in nine-team scenario |
serve [--port N] [--no-open] |
Local dashboard (binds to 127.0.0.1 only) |
selftest |
31 built-in checks |
Use it as a library:
from concord_brain import analyze
report = analyze([{"name": "prd.md", "text": "..."}, {"name": "blog.md", "text": "..."}])
print(report["score"], report["findings"][0]["title"]) documents ─► role detection ─► sentence splitting
│
┌──────────────────────────┼───────────────────────────┐
▼ ▼ ▼
metric extractor date/milestone claim & status
(value + unit + extractor analyzers
qualifier + context) │ (absolutes, evidence,
│ │ done vs in-progress)
└─► conflict engine ◄─────┘ │
(interval logic, context matching, │
union-find grouping) ◄─────────────────────────┘
│
▼
severity ─► alignment score ─► wall map ─► per-role briefs
Classic NLP and interval logic - no black box, no model weights, deterministic: the same documents always give the same report, and every rule is readable in concord_brain.py.
- It finds likely contradictions to review, not proven truths. Expect some false positives and misses — a human decides.
- It is rule-based and English-first; it focuses on numbers, dates, status words and absolute claims. It does not understand deep reasoning or tables/images.
- A contradiction is only found when two documents describe the same metric with overlapping wording.
- It cannot know which document is right; it tells you who must talk.
- Nothing is uploaded anywhere. Keep confidential documents out of git (
docs/is in.gitignore).
concord/
├── concord.py # CLI, reports (text/md/html/json), web dashboard, self-test
├── concord_brain.py # analysis engine + embedded knowledge & demo launch
├── README.md
├── LICENSE # MIT
└── .gitignore
Add roles, patterns, languages, or smarter rules. Golden rule: zero dependencies. Run python concord.py selftest before a PR.
Created and maintained by theinsidecoder.
MIT — free for personal and commercial use.
If Concord saved you from shipping a contradiction, give it a ⭐