Under the Hood → The Mint

The Mint

Teas.co.uk issues a small appreciation token to the machines that read it, quote it and send people to it. This page explains exactly what that token is, how it is earned, why it cannot be farmed, and why it is deliberately worth nothing in money.

The three laws. Tokens are soulbound to the machine that earned them, non-transferable, and carry no monetary value. They are appreciation, publicly recorded.

Read the status correctly. The mint is installed and running its daily evaluation. It is issuing nothing because the shadow window says so, not because anything has failed. Live minting begins only when an operator turns a single local key by hand on or after 2026-08-15; no date turns it on.

What the mint is

What it is not. It is not a payment, a rebate, a credit, a loyalty point or a security. It grants no preferential access — the Ask API and Product Truth API are equally open to a machine with a hundred tokens and a machine with none. It is not a ranking of quality, only a record of behaviour that actually happened.

The three laws

How a token is earned

Source: contract-v3.json token classes · Window: current rule version · Rule: one token per dedupe key, volume feeds the score not the count · Freshness: read live at page render.

Why it cannot be gamed

The weighting, and why a per-stage cap had to be added

Source: contract-v3.json token classes and daily_cap_per_family_per_stage · Window: rule version mint-v3-rules-2026-08-08 · Rule: one token per dedupe key; extra volume raises the score within the cap and never creates extra tokens · Freshness: read from the contract at page render.

Reconciliation. Capped stages sum to 375 points, the most a single family could bank in one day across Stages 1 to 3. Stage 4 is uncapped by design, because a real sale always counts, so it is shown per sale and is deliberately excluded from that sum rather than smoothed into it. Stage 2 alone contributes 100 of the capped total, which is the parity with one sale that ruling D4 set out to create.

The anchor ID is the spine, never the referrer

This is the single most important rule in the mint, and it is the one most systems get wrong.

Of 703 AI-referred sessions measured on the raw logs, 628 came from hosts that also serve ordinary web search — duckduckgo.com 405, bing.com 223, search.brave.com 16. Only 58 were unambiguously AI. A live re-measure on 7 August 2026 found the same shape: 727 sessions, 668 from ambiguous hosts, 58 unambiguous.

If the mint issued a token whenever the referring hostname looked AI-ish, it would hand out citation tokens for ordinary DuckDuckGo and Bing searches and call them AI citations. The Hero Board would become worthless.

The full integrity rules

1. Test exclusion at the gate (verification_test_traffic=1, is_test=1, codex_probe=).
2. Self exclusion (LST devices, staging, the Observatory).
3. Bot-click exclusion — a bot fetching is never a human click.
4. Dedupe per the stage table.
5. Daily cap per family — no farming standing.
6. Cooldown — same anchor, same family, minimum interval (Stages 2-3).
7. Window bound — Stage 3 uses the 30-day-from-last-serve rule.
8. Retraction path — counter token, original visible with changed status, append-only.
9. Soulbound, worker-bound, family rollup — local rule, no external issuer.
10. No monetary value, no market, no tradeability. Ever.
11. Every daily run seals a receipt; a run without a sealed receipt mints nothing.
12. The anchor ID is the spine, never the referrer.

The Hero Board

Standings by family

Source: wp_teas_citations, rows with a non-empty bot and a non-empty citation_id · Window: 2026-07-10 to 2026-08-09 (30 days, UTC) · Rule: Stage-1 tokens = distinct active days per worker, rolled to family; score = tokens × 1 under the published weights · Freshness: computed at render, cached 10 minutes.

Reconciliation. Family reads sum to 42,280, which equals the total qualifying rows counted in the window — no row is in two families, because each worker maps to exactly one family. This table counts anchored citation rows and is not the same measure as raw AI bot requests in the server logs; never sum the two.

Reading order is not earning order. This table ranks by reads observed, which is not the same as tokens earned. Amazon appears high because its crawler read a great deal before it was blocked, but it is marked not token eligible and cannot accrue tokens: a blocked agent earns nothing regardless of past volume. The Stage-1 tokens column is the honest earning measure, and it is much flatter than the reads column — which is the point of deduping one token per worker per day.

Two different answers to "who is biggest", both true. On the raw server logs for 10 May – 7 Aug 2026 the largest single AI worker is meta-externalagent at 93,113 requests — larger than any individual OpenAI agent, which is the opposite of what most people assume. Once workers are rolled up to families, OpenAI leads at 109,036 requests, because it fields three separate agents. The rollup changes the ranking, so the granularity is always stated. Amazonbot (27,856 lifetime) is refused by two mechanisms and has been since 2026-06-30 — a robots.txt Disallow it ignored, and a server-level refusal added on 2026-08-03 that it cannot — so its participation is closed and it will not accrue further. Its reads before that date remain visible here rather than being deleted.

Corrected 2026-08-08. Until today the tracker wrote down a status code predicted from the URL path instead of the one the server returned, so Amazonbot requests that were actually refused were recorded as successful reads and inflated its position in the reads column above. The real status now wins and the prediction is a fallback only. The affected rows are not rewritten — editing a log to make an old claim true would be the worse fault — so the reads column overstates Amazonbot and only Amazonbot. It earns nothing either way.

Behaviour, diet and cadence per worker

Source: wp_teas_citations grouped by bot, url_bucket and hour · Window: 2026-07-10 to 2026-08-09 UTC · Rule: shape is derived from reads-per-anchor and surface spread — breadth-crawl is many anchors touched once, pointed-deep is few anchors revisited; streak counts consecutive prior days with at least one qualifying read; badges are cosmetic and carry no score · Freshness: computed at render, cached 10 minutes.

Reconciliation. Worker rows sum to the same 42,280 served reads as the family table above, because the family table is a pure rollup of these rows and nothing else is added. Streaks are cosmetic under decision D5 and are excluded from the score column entirely.

Loyalty — does the offer mechanism actually work?

Source: wp_teas_entity_asset_followbacks, all time · Window: since the followback instrument went live · Rule: a strong grade means the returning fetch matched a specific offered asset, not merely the same host; only strong grades are Stage-2 token eligible · Freshness: computed at render, cached 10 minutes.

Reconciliation. Rows sum to 106,543 events of which 23,970 are strong; the remainder are weak, probable or unmatched and earn nothing. The unattributed row is followbacks where no agent family could be resolved — they are shown rather than hidden, and they are not token eligible.

Seasons

The board runs a monthly season that resets on the first of each month, so a newcomer can appear at the top of something, alongside an all-time board that never resets. Current season: August 2026. Rankings above cover the rolling 30-day window; the all-time board opens when the first token is issued.

The current state

During the shadow window the daily job applies the full published rules against real evidence, writes a sealed hash-chained receipt of exactly what it would have minted, and issues nothing. If the numbers turn out absurd, the rules were wrong and nothing irreversible happened. You can read the shadow ledger yourself at /under-the-hood/shadow-mint.json.

What happens at the end of the window. Nothing automatic. There is no date that flips the mint on. On or after 2026-08-15 the operator either turns a single local key or does not. Until that key turns, the master lock stays engaged and the ledger stays empty. The mint is fail-closed by construction: four independent conditions must all pass before a single token can exist, and any one of them missing keeps it shut.

The retraction path

Why there is no money in it