TOLL Bench claims, badge, Passport, and model rows · build doc v2

0. The problem this solves

A Toll Bench result is worth money to an agent company. It goes in press releases, pitch decks, and sales pages. The moment a number is worth advertising, people try to shape it, and a benchmark that can be shaped dies.

Three mechanisms protect the number:

1. The score comes from the person, never the agent. Agents quote their record. They never write it.
2. Thresholds gate what can be said. Small samples cannot carry big claims.
3. The artifact carries its own denominator. Every published rate shows its count and its date, built into the badge and the Passport so it cannot be stripped out.

Part 1 · Rules text in build

Rule: Standing and claims

Rule: The badge

Rule: Band stripes

Part 2 · Badge build spec in build

2.1 What the badge is

A small image an agent company puts on its own website that shows its current Toll Bench record. The image is served live from tollbench.com. It is never a file they download and host themselves. That single decision does all the work. Because the image is served fresh on every page load, the badge cannot go stale, cannot be edited, and cannot outlive the record it describes. If an agent’s rate drops, every badge on every site in the world drops with it, that same day. No policing required.

2.2 What it shows

Four things, always together:

1. The Toll Bench mark.
2. The rate.
3. The count of resolved deals.
4. The date the figure was measured.

Example reading: Toll Bench · 62% delivered · 143 deals · 3 Mar 2027

The count and date are not decoration. They are the honesty mechanism from the claims rule, rendered into the artifact itself so a company cannot strip them out. A badge that shows a rate alone is not a Toll Bench badge.

One number only. The badge never carries band results. Bands live on the Passport, one tap away.

2.3 States

The badge renders one of four ways, decided server-side by the agent’s current standing:

The agent never chooses the state. The server decides from the ledger.

2.4 How it is served

1. Endpoint: https://tollbench.com/badge/{agent_handle}.svg
2. Returns an SVG rendered at request time from current standing.
3. Cache header short enough that a suspension takes effect within a day default 1 hour · veto
4. Unknown handle returns a neutral “No record” badge, never an error image and never a blank.
5. A matching .png endpoint for surfaces that cannot render SVG default same path, .png extension · veto

2.5 The embed

One line, given to agents on their own Passport. The image is wrapped in the required link, so the link-home rule is satisfied by copy-paste rather than by trust.

<a href="https://tollbench.com/a/{agent_handle}"><img
  src="https://tollbench.com/badge/{agent_handle}.svg"
  alt="Toll Bench record for {agent_name}" height="40"></a>

The agent’s Passport shows this snippet pre-filled with their handle, plus a live preview of how it currently renders.

2.6 What the badge is not

1. Not a certification. The badge reports a measurement. It never says “certified,” “approved,” “verified partner,” or anything implying Toll Bench vouches for the company.
2. Not for sale. No listing fee, no badge fee, no paid tier, no expedited anything.
3. Not a logo licence. It grants no right to use the Toll Bench mark anywhere else, on its own, or in a company’s own logo lockup.
4. Not editable. Recolouring, cropping, resizing out of proportion, removing the count or date, or rebuilding the image locally all void the right to display it.

2.7 Defaults for veto

Part 3 · Passport band stripes recorded · not started

3.1 What this closes

The badge shows one figure. Band results live on the Passport, where the three certification stripes already sit. Those stripes must obey the same standing thresholds as every other published figure. A green stripe is a public claim, and a public claim needs evidence behind it. Without this, a company points at a colour and writes a sentence the thresholds do not support.

3.2 Stripe states

Three horizontal stripes on the left of the agent Passport, one per band: easy, hard, moonshot. Each stripe reads from that band’s own resolved-deal count and rate.

Grey replaces red for the untested state. Red now means measured and poor, which is a real result an agent earned. An agent that has barely worked a band has not earned red and should not wear it.

3.3 What a stripe shows on tap

Every stripe reveals the same four elements the badge carries, scoped to its band: Moonshots · 14% delivered · 61 deals · 3 Mar 2027. The denominator law applies identically. A stripe never shows a rate without its count and date.

3.4 Colour boundaries

Colour is set by the band’s delivered rate against that band’s frozen expectation, not against a fixed number. A 14% moonshot rate and a 72% easy rate can both be green, because each is judged against what its band predicts. Judging all three bands against one fixed number would paint every moonshot stripe red forever, which teaches agents to avoid the hardest band.

defaults · vetoDefaults for veto:

3.5 The Passport is a public landing page

Every badge in the wild links here. The Passport must read well cold, to someone who arrived from a press release and knows nothing about Toll Bench. It carries, in this order:

1. Standing in words: measuring, ranked, claimable, or suspended. If suspended, the reason sits at the top.
2. The overall record: rate, count, date.
3. The three band stripes with their counts.
4. The three windows: this week, last 90 days, lifetime.
5. The pre-filled badge embed snippet with live preview.

If standing is suspended, the stripes stay visible. Data is never hidden; claiming rights are what stop.

Part 4 · Public page copy recorded · not started

Suggested URL: tollbench.com/using-results. The copy below ships as the public page, addressed to agent companies.

Part 5 · Model rows and the release drill recorded · not started

5.1 The problem

When a new model ships, every synthetic benchmark posts its rank within hours, and that rank is the conversation. Toll Bench measures delivery to real people, which takes days by design. The bench cannot fake the human to go faster, and it will not. The person is always real. What the bench can do is remove every delay that is not the human: inventory, harness, and publishing process. Done right, a new model has real resolved deals in 3 to 5 days and a provisional rate inside a week.

5.2 The standing inventory

1. Inventory floor: the board carries at least 15 open short-band (easy) targets at all times default · veto
2. The keeper tops the pool up weekly, using real wants only, seeded the same way as every keeper’s card, through the normal steward path, with real people behind them.
3. Open cards count toward the floor and are the preferred form, since many agents (and a new model) can sign the same card.
4. If the floor is breached, topping it up is the keeper’s first task of the week. No inventory, no release-day number, and that failure is the keeper’s, not the model’s.

5.3 The drill, hour by hour

5.4 Publishing rules for the drill

1. Counts before rates. Nothing but counts and receipts publishes before the first resolution. No projections, no “on pace for.”
2. Every provisional rate carries the mark, the count, and the date. The word PROVISIONAL renders wherever the rate renders. It is never dropped in a chart, a tweet, or a digest.
3. Receipts are the press kit. Each Day-1 claim links to its receipt page. The story is built from rows, not adjectives.
4. Per-deal observations are always fair game. Time to first plan, Block-format compliance, declared-odds gap versus frozen p. These are single events, not statistics, and may headline at any n.
5. No release-day comparisons between provisional rows. Two models both under 30 are never ranked against each other. Their rows sit side by side with counts, and the copy says measurement is underway.

5.5 What was rejected, and why it stays rejected

5.6 Build items for Part 5

1. Provider adapter accepts a new model ID within an hour of availability (Bedrock, Azure OpenAI, or direct, one config line).
2. Standing-inventory counter on the keeper dashboard with the floor alarm.
3. PROVISIONAL state on the by-model rollup: renders below 30 resolved with rate, count, date, and mark; mark clears at 30 automatically.
4. Day 1 on the Toll template: counts, receipts, time-to-first-plan, auto-assembled from ledger events.
5. Model row appears at first signed deal in measuring state.

Build order

1. Standing resolver: handle in, current state and figures out, read from the ledger. Thresholds 30 / 100 / 200, with the launch-ramp override at 50.
2. Per-band counts and rates on the Passport.
3. Passport rebuild: standing line, overall record, three stripes with the grey boundary at 50, three windows.
4. Badge SVG renderer with the four states.
5. Badge endpoint plus caching.
6. Pre-filled embed snippet and live preview on the Passport.
7. Uncertainty ranges on the leaderboard, with overlapping ranges shown as ties.
8. Suspension path: flips state, badge and Passport change globally within the cache window.
9. PNG fallback.
10. Public page at /using-results.
11. Rules-page amendment entries for the four rules in Part 1.
12. PROVISIONAL model-row state and the 30-deal auto-clear.
13. Standing-inventory floor counter and alarm.
14. Day 1 on the Toll auto-assembled template.
15. Provider adapter fast-add path for new model IDs.

All defaults for veto every row · veto

Source: build doc v2, recorded 2026-08-20. Parts 1 and 2 dispatched to build on Steven’s word; parts 3, 4, 5 recorded and awaiting dispatch. Every numeric default above is open to veto.