> For the complete documentation index, see [llms.txt](https://docs.unbrowse.ai/llms.txt). Markdown versions of documentation pages are available by appending `.md` to page URLs; this page is available as [Markdown](https://docs.unbrowse.ai/whitepaper/paper-vs-product.md). # Paper vs Product Status This page maps the paper, *Internal APIs Are All You Need* (arXiv:2604.00694), to the hosted Unbrowse service that runs today. Status meanings: * `Shipped`: running in the hosted service, and its acceptance check is closed with evidence. * `Shipped, check open`: running in the hosted service; its acceptance check is not yet closed. * `Partial`: some of the claim exists, in a narrower or different form. * `Changed`: the product solves the same problem a different way. * `Not shipped`: described in the paper, not in the service. The client in this repo is a thin transport. Every row below is service behavior. ## Capability layer | Paper claim | Status | What exists today | | -------------------------------------------------------------------------------------- | ------------------- | ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ | | Agents should call first-party routes, not drive the DOM | Shipped | Learned capabilities replay over first-party HTTP. The browser is for discovery and for reads HTTP cannot finish. | | Passive indexing from normal use | Shipped | The cloud browser records every session. `browse.finish` compiles it. The first request is fulfilled while indexing. | | Learning from recorded traffic | Shipped | `unbrowse_learn` compiles HAR files or traces. Cross-session diffs find inputs; producer→consumer tracing binds tokens and ids. | | Three paths: local cache, shared graph, browser fallback | Changed | Private space, then public registry, then the cloud browser. A render fallback covers reads HTTP cannot finish. The client keeps no local cache. | | Shared route graph | Partial | A public registry of scrubbed, re-verified read-only capabilities (check open). It holds reads, not a general graph of writes. Multi-step flows live inside one capability's harness. | | Intent resolution by composite score (embedding, reliability, freshness, verification) | Changed | Discover ranks and returns health hints (p50/p95, success rate, warm/rendered/cold). Jev picks among eligible candidates, with a deterministic fallback. No embedding-weighted formula. | | Schema drift detection | Shipped | HTML where JSON was learned fails as `schema_or_auth_drift` and can fall back to rendering. A version-aware compatibility ledger excludes mismatched bindings until targeted revalidation. | | Route lifecycle | Shipped | `observed` → `candidate` → `validated`, per package digest. | | Periodic re-verification | Shipped, check open | Registry revalidation with fresh words; repeated failures unpublish. | | "HTTP 200 is not success" | Shipped | `succeeded` requires an independently verified business outcome. | | Per-route package (`SKILL.md`, `auth.json`, generated `api.ts`) | Partial | Each capability is `unbrowse/v1alpha1` harness YAML plus a SKILL.md. Each site is an MCP server and an OpenAPI 3.1 document (check open). No `auth.json` or generated `api.ts`. | | MCP support | Shipped | Remote MCP at `/mcp`, an adapter over the same runs as REST `/api/v1`, which `@unbrowse/sdk` and the CLI in this repo call. | | A2A / ANP protocol coverage | Not shipped | | ## Credentials | Paper claim | Status | What exists today | | -------------------------------- | ------------------- | ----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | | Credentials are never published | Shipped | The registry strips cookies, authorization and session values. Login-backed and write capabilities are never shared. | | Local encrypted credential vault | Changed | The vault is server-side and envelope-encrypted per workspace (shipped). A zero-knowledge Private vault, decryptable only in the owner's browser, is the newest piece (check open). | | Auth reuse across runs | Shipped, check open | Auto sign-in on replay, cached sessions, login step skipped on later runs. | ## Economic layer | Paper claim | Status | What exists today | | ------------------------------------------------------------ | ----------- | ----------------------------------------------------------------------------------- | | Adoption condition: route fee below rediscovery cost | Not shipped | Pricing is one flat meter, not priced per route. | | Metered usage | Shipped | 500 verified calls a month free, then $10 per 10,000. Only verified successes bill. | | HTTP 402 handshake and x402 settlement | Removed | Shipped, then removed in September 2026. Every call needs an account. | | Solana settlement | Not shipped | No on-chain settlement today. | | Route-level pricing | Not shipped | | | Fee splits across contributors, maintainers, infra, treasury | Not shipped | | | Contributor payouts and delta-based attribution | Not shipped | Auto-shared routes carry no payout. | | Site-owner compensation | Not shipped | | | Dynamic pricing by savings and trust | Not shipped | | ## Trust and validation | Paper claim | Status | What exists today | | --------------------------------------- | ------------------- | ------------------------------------------------------------------------------------------------------------------ | | Run history drives trust | Shipped | Health from each capability's run ledger shows in discover hints. | | Pre-publish quality gate | Shipped, check open | A scrubbed copy must answer two probe words and one unseen word, fail a nonsense query, and do so without cookies. | | Broken public tools are hidden | Shipped, check open | Discover withholds failing public tools unless the query names the site, and reports how many it withheld. | | Signed feedback, validator attestations | Not shipped | | | Independent validator markets | Not shipped | | | E2B sandbox validation | Not shipped | | | TEE attestation | Not shipped | | ## Benchmarks | Paper claim | Status | Notes | | ----------------------------------------------------------------------------------- | ------------------ | ----------------------------------------------------------------------------------------------------------------------------------------------------------- | | 94-domain benchmark: 3.6x mean / 5.4x median speedup on warmed routes vs Playwright | Paper result | Not re-run against the current service. | | 90–96% per-task cost reduction for warmed-cache execution | Paper result | Not re-run. | | \~8,000 DOM tokens down to \~200 JSON tokens on typical reads | Paper result | Not re-run. | | Warm replay much faster than browsing | Shipped (own eval) | The service's production lean eval measured median warm replay at 5.4% of median browse time on its demo site. See [Evaluation](/whitepaper/evaluation.md). | ## Read this as * The paper's core thesis (first-party routes, learn once, replay over HTTP) is what the service does. * The shared-graph idea ships as a read-only public registry, still maturing. * The route economy (per-route prices, splits, payouts) and the cryptographic trust layer do not exist. Billing is a flat meter.