Built 30 October 2026 on the prototype 1 reference site. Roadmap gate: the dashboard distinguishes search, training and agent traffic and attributes leads; the citation probe runs unattended for four weeks.
What was built
| Piece | Where | What it does |
|---|---|---|
| Event store | packages/events |
One events table beside the content; every surface records page views, Markdown fetches, robots fetches, leads, conversions and probe observations with traffic class, agent name, signature status, session and referrer. Queries for traffic by class, agents, pages, actions, lead attribution, funnel and citations. |
| Signed-agent verification | packages/edge/src/webbotauth.ts |
Web Bot Auth: RFC 9421 HTTP Message Signatures with Ed25519, keys published in the agent's /.well-known/http-message-signatures-directory, keyid as the RFC 7638 thumbprint. Signer for Sysbot's own agents; verifier for the site. A verified signature names the agent by its directory host and outranks any user-agent string. |
| Request classification | apps/web/lib/traffic.ts, middleware.ts |
Edge headers or user agent first, then signature verification; an opaque session cookie for attribution; assistant referrals (ChatGPT, Perplexity, Claude, Gemini, Copilot, Grok) recognised from the Referer. |
| Operator dashboard | apps/web/app/ops |
Traffic by class, assistants and agents (claimed vs signed), pages retrieved, lead attribution by first touch, open leads with mark won/lost, actions, citation probe results. Gated by OPS_TOKEN (ADR-002). |
| Citation probe | apps/probe/src/citation |
Asks five assistants (Claude, ChatGPT, Perplexity, Gemini, Grok) the buyer's questions from citation.config.json, scores cited / mentioned / surfaced / absent with rank and competitors, writes a dated report to apps/probe/citations/, posts observations to POST /api/events when INGEST_TOKEN is set. Assistants without a key are skipped and listed. |
| Schedules | .github/workflows/citation-probe.yml, apps/probe/schedule/ |
Weekly Monday run on GitHub Actions (commits observations back) or as a launchd agent on a Mac. |
| Harness identity | apps/probe/src/identity.ts |
The harness's browser and fetch agents now sign their requests as agents.sysbot.test with a persistent key and serve their directory locally, so the dashboard shows what a verified agent looks like next to claimed ones. |
Evidence

Local run, 30 October 2026: one harness pass (three agents, five tasks), six known crawlers and fetchers reading five paths each, one person referred from Perplexity who requested a quote that operations then marked won, and one citation probe with Claude. The dashboard separated search crawlers, training crawlers, user-directed agents, unknown bots and people; showed the harness agent as signed and every user-agent-only crawler as claimed; attributed the won lead to the Perplexity referral and its first page; and attributed two agent-submitted quotes to the signed agent. The probe found the fictional business cited in 0 of 5 answers and listed the directories and dealers that were cited instead, which is exactly the report a real customer would want.
Gate status
- Dashboard distinguishes search, training and agent traffic and attributes leads: met (above).
- Probe runs unattended for four weeks: started, not yet met. First scheduled run Monday 2 November 2026;
four weekly runs complete on 23 November 2026. Requires either a GitHub remote with
ANTHROPIC_API_KEYas a repository secret or the launchd agent on a Mac that stays awake. Until then the probe measures the fictional reference business; setsite,brandandqueriesinapps/probe/citation.config.jsonto a real site to make the four weeks count. On 31 October 2026 the reference site itself was deployed to https://northwind.sysbot.in andsitenow names it (docs/publishing/reference-site.md), so the runs measure a reachable site.
Findings worth carrying into the blueprint
- Claimed versus verified is the first column that matters. Every published crawler arrives as a user-agent string anyone can type; the only row the dashboard can vouch for is the signed one. Web Bot Auth verification cost under a millisecond per request with the directory cached, and the harness agents needed one header set per run because the signature covers the authority, not the path.
- Sessions mean nothing for crawlers. Cookie-less fetchers produce one session per request, which inflates session counts for bot classes. The dashboard should report requests for bots and sessions for people; the funnel already attributes leads correctly because leads come from sessions that do carry the cookie.
- Headless automation without a name is an unknown bot, and that is correct. The prototype 1 baseline agent and a raw Playwright session both land there. A site cannot tell a scripted buyer from a scraper unless the agent identifies itself; this is the commercial argument for the agent gateway product.
- Citation probes need unbranded questions. A mention on a question that already names the brand is noise; the scorer ignores it. Rank among citations and the competitor list are the actionable numbers.
- Assistant referrals are rare but attributable today. A
Refererfrom an assistant domain is enough to build the "assistant referral" funnel bucket without any vendor integration.
Known limits
- Only Claude ran in the first probe; the other four adapters are written against each vendor's current API shape and are exercised only when their keys exist.
- Gemini returns redirect URIs for grounding sources; the adapter scores by the source title's domain.
- The dashboard is single-tenant and single-deployment; period filters only.
- The
OPS_TOKENandINGEST_TOKENgates are prototype gates, not the security model.