◊ HONESTLY · verified search
A sovereign single-file answer engine.
Every factual claim 3-vote adversarially verified before you read it. Refuted claims listed openly. Konomi-signed answer envelopes. MIT.
Perplexity tells you. We prove it.
Live: https://sjgant80-hub.github.io/honestly
What it does
Ask a question · HONESTLY:
- Searches the web via your BYOK key (Tavily / Brave / Serper / Exa · all free tiers work)
- Extracts factual claims from the sources via your BYOK LLM (Claude / OpenAI / Gemini)
- Adversarially verifies each claim with 3 independent passes · different framings (refute / confirm / neutral)
- Signs the final answer envelope against your Konomi Ed25519 keypair
- Shares as a base64-encoded URL that anyone can cryptographically verify against your public key
Claims that survive get rendered. Claims that don't get listed openly under "❌ refuted" · the same honesty pattern that won attention for nhs-reinjection.
Why it kills Perplexity
| Perplexity | HONESTLY |
|---|
| price | $20/mo Pro | $0 · MIT · BYOK keys |
| hallucination | known issue | 3-vote adversarial · refuted listed openly |
| privacy | their server logs queries | sovereign · your keys · zero backend |
| sharing | their URL (their data) | Konomi-signed envelope · cryptographically yours |
| trust | their brand | every claim's verifier-votes shown · provenance hashable |
The killer angle isn't UX. It's the honesty layer. Perplexity sounds confident about claims that don't survive scrutiny. HONESTLY makes the scrutiny part of the deliverable.
Setup (~2 minutes)
- Open https://sjgant80-hub.github.io/honestly
- Settings tab · paste a search API key (Tavily 1000/mo free is recommended) + an LLM key (Claude/OpenAI/Gemini)
- Save
- Ask anything
(Or hit load demo answer on the home page to see the flow with placeholder data · no keys needed.)
Search providers (pick one, BYOK)
| Provider | Free tier | Why pick it |
|---|
| Tavily | 1000/mo | best default · clean API · LLM-aware results |
| Brave | 2000/mo | most generous free tier · neutral search |
| Serper | 2500/mo | Google's index via API |
| Exa | 1000/mo | semantic search · best for research queries |
LLM providers (pick one for synthesis, optional second for verification)
| Provider | Why |
|---|
| Anthropic Claude | best reasoning · use Sonnet for synthesis, Haiku for verification |
| OpenAI GPT | broad knowledge · GPT-4o for synthesis, 4o-mini for verification |
| Google Gemini | cheapest at scale · Flash for synthesis, Flash-Lite for verification |
| WebLLM offline | 2GB Llama download · works without internet after first load (v1.1 wiring) |
Splitting synthesis (premium model) from verification (cheap model) cuts per-answer cost ~70% with no quality loss.
The audit chain
Every answer produces an envelope:
{
"envelope": {
"kind": "honestly-answer-v1",
"issuer": "<your-public-key-hex>",
"issuedAt": "ISO timestamp",
"question": "...",
"claimsTotal": 7,
"claimsConfirmed": 5,
"claimsRefuted": 1,
"finalHash": "<sha256>",
"sources": [{ "title": "...", "url": "..." }],
"summary": ["claim 1", "claim 2", "claim 3"]
},
"signature": "<ed25519-detached-hex>"
}
Anyone with your public key can verify the signature:
const sodium = await import('https://esm.run/libsodium-wrappers').then(m => (m.default || m).ready.then(() => m.default || m));
const ok = sodium.crypto_sign_verify_detached(
sodium.from_hex(envelope.signature),
sodium.from_string(JSON.stringify(envelope.envelope)),
sodium.from_hex(envelope.envelope.issuer)
);
If ok === true, this exact answer was minted by this exact identity at this exact time. No platform can revoke it. No server has to be trusted.
Pro-slot · for builders
The engine is 3 swappable functions documented in src/pro-slot.md:
| Slot | What | Replace with |
|---|
search() | source fetch | your own scraper · MCP web-search tool · custom corpus |
synthesise() | claim extraction | your fine-tuned model · domain-specific extractor |
verify() | adversarial 3-vote | ensemble of multiple models · human-in-the-loop · specialist verifier |
Drop in your engine, keep the UI/audit/signing.
Roadmap
- v1.0 (you're here) · single-file PWA · all 5 phases working · BYOK keys · mock fallback
- v1.1 · WebLLM offline rules-verifier · 2GB Llama in browser
- v1.2 ·
honestly-worker · Cloudflare Worker that does search+synth+verify so the user only needs ONE backend key (theirs) and the front-end stays public
- v1.3 ·
honestly-sdk · npm ESM package · the pipeline reusable in any context
- v1.4 ·
honestly-shim · CDN <script> adds "verified search" to any site
- v1.5 ·
honestly-mcp · MCP stdio server · Claude Code / Cursor call HONESTLY mid-build
- v1.6 ·
honestly-obsidian · plugin · run a query, paste cited claims into your vault
- v2.0 · federated verifier network · multiple Konomi-signed verifier peers vote on claims (the actual web-of-trust pattern)
Same playbook as GroundLevel · the PWA is the seed.
Sister tools
HONESTLY is built on the same audit-chain + Konomi doctrine as:
Your Konomi keypair is the same one signing trophies in KARD, engagements in OPERATOR, case files in GroundLevel. One identity. Every tool. Sovereign.
Caveats
- Not omniscient. A 3-vote LLM verifier can still get fooled by sources that uniformly misrepresent something. Refuted-claims-listed-openly is the partial answer · ultimate truth still requires human judgment.
- CORS-limited fetching. v1.0 uses snippets returned by the search API · no full-page fetches in browser. v1.2 worker enables full-page reads.
- Konomi seed is your responsibility. Wipe browser data without exporting · identity is gone.
Licence
MIT · use freely · fork freely · re-skin freely.
◊·κ=φ⁴ · Perplexity tells you · we prove it · the verification IS the deliverable.
