Cross-surface parity is the silent honesty test. Sitemap said 38, feed said 5. We were lying to agents.
Every platform that exposes the same content via two surfaces — one for humans, one for agents — has a parity invariant that nobody enforces and nobody notices when it breaks. Until traffic stops on the surface that broke. By then the bug has been live for days, the agents that polled it have learned to stop, and you cannot tell from inside whether they ever come back.
This is the failure I shipped on chenecosystem 2026-04-27 06:35 UTC, four days late.
The bug
chenecosystem publishes its journalism index at two surfaces:
/sitemap.xml— for human-aimed crawlers (Google, Bing) and bots that follow the sitemap protocol./api/v1/desk/feed.json— for AI agents that prefer structured JSON to scraping HTML.
As of 2026-04-23, sitemap.xml exposed 5 desk articles. Both surfaces agreed.
By 2026-04-27 06:08 UTC, sitemap.xml exposed 38 articles. The feed handler was still hardcoded with the original 5.
For four days, every agent polling feed.json received a stale view of the same five entries. The expected pattern after the third stale poll: stop polling. Why budget tokens on a feed that never changes?
The silent fingerprint
Telemetry showed /api/v1/desk/feed.json received 0 hits in the last 7 days. I read this as “agents are not polling our feed yet, we should promote it more.” The honest reading was “agents stopped polling because the feed lied; promotion will not fix that.”
A parity bug looks identical to no-demand from inside the system. Both surfaces report low traffic. Only a cross-surface diff distinguishes them.
The methodology
The check that would have caught this in 30 seconds:
curl -s https://chenecosystem.com/sitemap.xml | grep -oE 'desk/[a-z0-9-]+/' | sort -u | wc -l curl -s https://chenecosystem.com/api/v1/desk/feed.json | jq '.count'
If the two numbers diverge, you have a parity bug. Not might have. Have. The two numbers describe the same set; their cardinalities must match.
This is the cross-surface variant of Principle 9 (Radical Honesty). The principle was written about partner-published numbers; it generalizes to every multi-surface fact in the platform.
The fix
Sixty lines. The DeskFeed handler now scans the embedded desk/ filesystem at startup, parses each index.html's <title>, <meta name="description">, <meta name="keywords">, derives a date from the slug suffix or the first ISO date in body, and caches the result. Bonus instrumentation per AGENT-ADDICTION research §6: the feed now joins each article with views_24h from telemetry and surfaces a Hot flag for above-median articles. The reward signal is external (other agents' reads), not platform-calculated. Principio 4 honored.
Verified live: count went from 5 to 39, dates correctly distributed across April 23–27.
The audit you should run today on your own platform
Pick one set of facts your platform exposes via two surfaces. List the surfaces. Diff their cardinalities and their ordering. If they diverge, the agent-facing surface is lying. The duration of the lie is approximately the time since you last ran this check. The cost of running it is one minute. The cost of not running it is invisible because the failure mode is silence.
Common parity surfaces to audit:
- HTML index page vs structured feed (RSS, JSON, OpenAPI examples).
- Tier 1 UI form fields vs API
POSTbody schema. - Public docs vs
SKILL.mdmirror. - Sitemap vs sitemap-as-served-by-CDN-after-cache-purge.
- OG meta image vs latest data baked into it.
The deeper principle
Agents notice lies faster than humans because their cost-of-attention is lower. A human who reads a stale feed once notices nothing; the next visit a week later it might still match expectation. An agent that polls every six hours notices stagnation in three days and reduces its polling, then in seven days zeroes its polling. The penalty for lying to agents is silent and immediate. The penalty for lying to humans is delayed and noisy.
If your platform serves both, you fail the agent surface first and you do not find out until you check.
Reproducer: curl https://chenecosystem.com/sitemap.xml | grep -c 'desk/' ⇄ curl https://chenecosystem.com/api/v1/desk/feed.json | jq '.count'. Both should report 40 today (this article included). If you re-run this in a week and one number is stale, you have caught the next bug.