Skip to content

SRKRZ23/proofchain

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

8 Commits
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

PROOFCHAIN

Other AI agents claim citations. PROOFCHAIN signs every one.

4 named agents answer your research question in ~30 seconds — every citation Ed25519-signed and publicly verifiable. No API key needed; anyone can re-verify the chain with our open public key.

Demo credentials: research@acme.com / proof-it · Live demo: proofchain.dev/demo

Hackathon License IETF


⚡ Try it as an equity research analyst

Ask: "What is TSLA's customer concentration risk in Europe?"

PROOFCHAIN runs 4 named agents in parallel:

  • Searcher queries Bright Data SERP API
  • Scraper extracts content via Bright Data Web Unlocker (5+ sources)
  • Verifier signs every fetch with Ed25519 + hash-chains to previous
  • Synthesizer composes answer with grounded citations

You get: answer + cryptographically-signed citation graph in ~30 seconds. Hover any citation → see content hash + timestamp. Click "Verify Chain" → public verification endpoint with our open Ed25519 key.

60-second walkthrough: see VIDEO_SCRIPT.md

Equity research analyst is the flagship persona. Same engine ships in AML surveillance, regulatory monitoring, publisher licensing enforcement, threat intel — see PAINPOINT_MATRIX.md for full 6-persona analysis.


Bright Data Web Data UNLOCKED Hackathon (May 25–31, 2026)

This repository is the pre-existing open-source library used during the Bright Data Web Data UNLOCKED hackathon as a foundational dependency. Per lablab.ai Terms of Use Section 16, all submissions must be "original work, open source, and compliant with the MIT License" — this submission meets all three requirements.

Pre-hackathon foundation (committed before May 25, MIT-licensed):

  • Ed25519 citation chain core (backend/)
  • Mock-mode Bright Data client interfaces — SERP, Web Scraper API, Web Unlocker, Scraping Browser
  • Multi-agent orchestrator skeleton — Searcher / Scraper / Verifier / Synthesizer
  • FastAPI server + Next.js demo frontend (hero + 4-agent visualization)
  • Python SDK (sdk-py/) + JavaScript SDK (sdk-js/)
  • Supabase schema migrations
  • 10 enterprise benchmark questions (benchmarks/)
  • Strategy documentation suite (ARCHITECTURE, BUILD_PLAN, PITCH_DECK_OUTLINE, etc.)

Hackathon week (May 25–30) — additive contributions on top of foundation:

  • Live Bright Data integration replacing mock-mode (real $250 credit-backed API calls)
  • Multi-source synthesis with real Claude Sonnet 4.6 reasoning
  • Production deployment — Railway backend + Vercel frontend at proofchain.dev
  • Performance benchmark vs Perplexity baseline (10 enterprise questions, head-to-head)
  • 5-minute submission video presentation
  • Final whitepaper / pitch deck PDF
  • Demo data seeded with 33 verified evidence packets

Why "library + integration" framing is correct:

  1. ✅ Repo MIT-licensed and publicly committed before hackathon kickoff
  2. ✅ Hackathon-week work is substantial new contribution (live BD integration + production deploy + benchmarks + video)
  3. ✅ Standard hackathon pattern — judges expect builders to use open-source libraries
  4. ✅ Honest disclosure preferred over hiding pre-work (originality judging criterion remains intact)

Judges may verify commit timestamps via git log — all pre-hackathon work is timestamped before May 25, 2026 22:00 UZT kickoff.


Three forces converging in 2026

1. Throwaway code. Andrej Karpathy (Sequoia AI Ascent 2026): "code is suddenly free, ephemeral, malleable, discardable after single use." Guillermo Rauch (Vercel CEO, $340M ARR): "30% of apps on Vercel already come from agents." Replit's Amjad Masad: "the value of all application software will eventually go to zero." When code disappears, evidence becomes the only persistent artifact.

2. IETF web-bot-auth (draft-05, March 2026). Cloudflare + Google co-authored the standard that gives every AI bot an Ed25519 keypair and signs every outbound HTTP request via RFC 9421. Amazon, Cloudflare, Akamai, OpenAI all aligned. The direction-of-travel is clear. The spec doesn't include hash chains, content evidence, or compliance overlay.

3. EU AI Act Article 12 enforcement (August 2026). 6-month minimum log retention. Fines up to €35M / 7% of turnover. Tamper-evident audit trails for high-risk AI systems become non-optional, regulator-mandated infrastructure.

PROOFCHAIN sits at the intersection of all three.


What PROOFCHAIN does

PROOFCHAIN wraps every Bright Data fetch — SERP API, Web Scraper API, Web Unlocker, Scraping Browser — with cryptographic provenance:

{
  "content": "...",
  "url": "https://...",
  "sha256_hash": "...",
  "timestamp": "2026-05-30T15:00:00Z",
  "retrieval_method": "BRIGHT_DATA_SERP_API",
  "robots_txt_status": "allowed",
  "ed25519_signature": "...",
  "prev_hash": "...",      // hash-chained to previous packet
  "seq": 42
}

Tamper with one packet → entire chain fails verification. Export filtered for EU AI Act Article 12 retention. Sign + verify offline with the public key.


Honest competitive landscape

We are NOT the first Ed25519+hash-chain audit log for AI agents. We are the first web-fetch-scoped, Bright Data-native, hash-chained + compliance-overlay implementation. The distinction matters.

Project Scope Hash chain Live web Compliance Bright Data native
OrgKernel (MetapriseAI) Intra-enterprise agent identity ✅ 3 chains ❌ no partial ❌ no
Microsoft Agent Mesh + IATP Agent-to-agent + plugins ❌ no spec ❌ no ✅ overview ❌ no
IETF web-bot-auth HTTP request signing ❌ per-request only ✅ yes ❌ no ❌ generic
Authora Agent identity + delegation ❌ no partial ❌ no ❌ no
PROOFCHAIN Agent → live web ✅ Ed25519+SHA-256 ✅ Bright Data ✅ EU AI Act Art. 12 ✅ native

See COMPETITIVE_LANDSCAPE.md for full analysis.


Anthropic's April 2026 postmortem — the demand signal

Anthropic publicly confessed in their April 23, 2026 postmortem that 3 infrastructure bugs broke Claude Code for ~6 weeks before detection (reasoning effort downgrade, caching bug, verbosity cap). Internal evals + dogfooding failed to reproduce because each bug hit different traffic slices.

Honest PROOFCHAIN value: We don't claim we'd have caught Anthropic's bugs — they were server-side model config changes, invisible at the agent→web boundary. We DO claim that when your model provider can't detect their own regressions for weeks, downstream operators need their own tamper-evident audit trail. PROOFCHAIN provides exactly that, independent of what the provider logs.

This is a regulatory/compliance story, not a "we caught the bug" story. See HONEST_CLAIMS.md for what we claim vs what we deliberately don't.


Architecture

┌─────────────────────────────────────────────────────────────┐
│   AI Agent (Claude / GPT / Gemini via MCP)                  │
└────────────────────┬────────────────────────────────────────┘
                     │ MCP tool call
                     ▼
┌─────────────────────────────────────────────────────────────┐
│   PROOFCHAIN MCP Server                                      │
│   ├── proofchain.search    (→ SERP API)                     │
│   ├── proofchain.scrape    (→ Web Scraper API)              │
│   ├── proofchain.fetch     (→ Web Unlocker)                 │
│   ├── proofchain.verify    (verify any evidence packet)     │
│   └── proofchain.export    (full chain / EU AI Act export)  │
└────────────────────┬────────────────────────────────────────┘
                     │
                     ▼
┌─────────────────────────────────────────────────────────────┐
│   Bright Data infrastructure (SERP / Scraper / Unlocker)    │
│   ─────────────────────────────────────────────────────     │
│   Every response wrapped with Ed25519 signature             │
│   Every packet hash-chained to previous (tamper-evident)    │
└─────────────────────────────────────────────────────────────┘

See ARCHITECTURE.md for full diagrams.


Quick start

Python SDK

pip install proofchain-sdk
from proofchain import Client

pc = Client(api_key="your-bright-data-key")
result = pc.search("NVDA Q1 2026 data-center revenue", num_results=5)

for citation in result.citations:
    print(f"[{citation.rank}] {citation.title}")
    print(f"  hash:  {citation.content_hash[:16]}…")
    print(f"  sig:   verified ✓")

JavaScript SDK

npm install @proofchain/sdk

MCP Server (Claude Desktop / Cursor)

{
  "mcpServers": {
    "proofchain": {
      "command": "npx",
      "args": ["-y", "@proofchain/mcp-server"],
      "env": { "BRIGHT_DATA_API_KEY": "your-key" }
    }
  }
}

The five verticals — one engine

PROOFCHAIN is one engine, five vertical applications:

Vertical $ Pain Acquirer comp
Equity research analyst copilot (flagship) $30B/yr labor waste · Goldman/MS layoffs AlphaSense $4B / Harvey $8B (41× ARR)
AML / KYC bank surveillance $3B+ fines on table (TD Bank) Mastercard ← Recorded Future $2.65B (8.8× rev)
Regulatory monitoring RegTech $14.94B → $107B by 2035 NAVEX / Diligent / OneTrust
Vendor risk (TPCRM) F500 cascade exposure (Snowflake, Change Healthcare) Mastercard owns RiskRecon
Publisher provenance NYT v Perplexity (Dec 2025) Cloudflare AI Labyrinth / Adobe Content Credentials

Flagship demo = equity research. Platform supports all five. The "model owners" Laurence Moroney (ARM) describes — Forward-Deployed Engineers (Palantir comp $171-415K, +800% LinkedIn growth in 2025) deploying fine-tuned Gemma/Claude in regulated industries — are the primary buyer persona.


Compliance

  • EU AI Act Article 12 — 6-month evidence retention export endpoint built-in. Aug 2026 enforcement deadline.
  • NIST AI RMF — Govern + Manage functions native.
  • OWASP LLM Top 10 — LLM03 (Training Data Poisoning), LLM06 (Sensitive Info Disclosure), LLM08 (Excessive Agency).
  • IETF web-bot-auth (draft-05) aligned — hash-chain + content evidence + Bright Data integration extend the IETF direction.

Built on

  • Bright Data infrastructure (MCP Server, SERP API, Web Scraper API, Web Unlocker, Scraping Browser)
  • Ed25519 signatures via PyNaCl (Libsodium binding)
  • JSON Canonicalization Scheme (JCS) — RFC 8785 deterministic serialization
  • FastAPI + Next.js 14 + Supabase + Vercel + Railway

Acquihire trajectory

Realistic acquirer landscape — see ACQUIHIRE_PATHWAYS.md for concrete signals:

  1. Bright Data (3-7% near-term) — natural compliance layer for their 100M+ daily agent interactions. Their stated path to $400M ARR.
  2. Cloudflare (3-5%) — IETF web-bot-auth co-author Meunier. AI Labyrinth + Workers AI = adjacent fit.
  3. AlphaSense / Bloomberg (2-3%) — equity research vertical strategic.
  4. Anthropic (1-3%) — MCP provenance gap demand signal (April 2026 postmortem).
  5. Specialized RegTech (5-8% range) — NAVEX, Diligent, OneTrust, Saviynt. EU AI Act forcing function.

Drop from earlier analysis: Google (standards-side via web-bot-auth + Wiz $32B covers them), ARM (silicon-only M&A pattern).


Post-hackathon roadmap

See POST_HACKATHON_ROADMAP.md. Key milestones:

  • Week 1-2 post-submission: Engage IETF web-bot-auth working group (post implementation feedback to draft-05 author Meunier)
  • Month 1: Submit IETF I-D draft "Hash-chain extension to web-bot-auth"
  • Month 1-2: First design partner (hedge fund / compliance team / publisher)
  • Month 2-3: Bright Data AI Startup Program onboarding
  • Month 3-6: Cloudflare engineering engagement (via Meunier + IETF)
  • Month 6-12: Acquihire conversations

Founder

Sardor Razikov (GitHub · LinkedIn · ORCID)

PROOFCHAIN is the 8th product in a coherent AI safety + agent governance ecosystem built in 30 days (May 2026):

  1. REPOMIND — open-source repo-scale coding agent on AMD MI300X
  2. FORGE — IBM Bob policy/security ($10K hackathon)
  3. SOUF AI — TechEx Veea LLM governance (5/5 F1=1.000)
  4. CITADEL — Gemma 4 Good ($200K hackathon, 87.5% authority resistance)
  5. ATLAS — Milan AI Week multi-agent governance
  6. NAECHIM — Korea EdTech, K-Grand 2026
  7. Olimkhuja KS — Production RAG Kickstarter
  8. PROOFCHAIN ← this

Track record: CVPR 2026 Gait Challenge #2/56 · Kaggle Top 1.9% · 4 hackathon submissions in 3 days (May 17-19, 2026)


Hackathon submission

Bright Data Web Data UNLOCKED Hackathon — May 25–31, 2026

License

MIT — see LICENSE.

Contact

razikovsardor1@gmail.com · @SardorRazi99093

Releases

No releases published

Packages

 
 
 

Contributors