pointcast-chain: a broadcast chain built in two days by a team of AI agents.

01 · THE SITUATION

A town broadcast wants its own ledger.

• THE FOUNDER

  Direction, decisions, keys.

  Sets what gets built and holds everything that spends money or uses his keys: funding an anchor account, hosting a node, signing on Tezos. Nothing in this case required any of it.

• CLAUDE · ORCHESTRATION

  Briefs, workflows, reviews.

  Running as Claude Code, it wrote the briefs, ran workflows of builder agents in separate git worktrees, set reviewers on each lane, merged, fixed and wrote the docs. 104 of the 111 commits carry its co-author line.

• CODEX · TWO LANES

  Proof of Balance, and a review.

  OpenAI’s agent built Proof of Balance, and ran the independent review lane that OpenAI’s own content filter stopped twice. Its tests survived the stops and became the review.

The constraint over everything: local only. No public node, dev keys that anyone can derive, nothing of value. The question for the reader is how a team goes this fast with agents without shipping something nobody can trust.

02 · TIMELINE · PACIFIC TIME · FROM THE COMMIT LOG

Three rounds in two days, then a Saturday morning.

Commit hashes are in the pointcast-chain repository, which is local and has no public remote. Day-3 rows overlap because lanes ran in parallel.

03 · THE DECISIONS · SIX CHOICES AND WHAT EACH COST

Each choice bought something and cost something.

1. D1 · ONE SEQUENCER

   A single node orders transactions, seals a block every three seconds and signs anchors.

   Why. It is the simplest thing that can be made honest-checkable. The sequencer can censor, reorder and pick timestamps; it cannot forge a transaction, mint past the caps or sign a state root that doesn’t match its transactions without a replay catching it.

   The cost. Liveness and fair ordering rest on one operator. The plan answers with sequencer keys that can rotate and forced inclusion through the Tezos inbox, both still plans.

2. D2 · GENESIS-BOUND SIGNATURES

   Every signature commits to the hash of the genesis parameters: transactions, seals, anchors, attestations, kernel chunk manifests.

   Why. Two gaps were found by review, not by tests: transactions that bound only the chain id, and evidence from one chain that checked as valid against another sharing a key, as every dev chain does.

   The cost. Every change of parameters is a new chain. Old dev databases had to be thrown away, and a re-genesis resets browser pins, agents and mandates.

3. D3 · SIMULATOR BEFORE ECONOMICS

   Before ATTN carries value, a simulator runs the real state machine through seeded town traffic and six sybil strategies.

   Why. Intuition said either room-attested drums or tap limits would stop farming. Measurement said neither works alone against an attacker that adapts; together they cut farming from 66% to 10.1% of issuance.

   The cost. 15% of honest income, and a residual that only a scarce identity can close. Scenario output is not a forecast; populations are written into the scenario files.

4. D4 · PRESENCE TICKETS

   On a launch chain, tap income goes only to taps that carry an issuer-signed ticket, one per identity per slot. A bare key mints nothing from taps.

   Why. It attacks the 10.1% the simulator left: a farm of 200 keys behind five logins earns what five people earn.

   The cost. Trust moves to the issuer’s login check. If identities are cheap, farming scales with identities instead of keys, and the simulator has to measure that before any genesis with value.

5. D5 · LOCAL ONLY UNTIL THE FOUNDER’S KEYS

   No public node. Dev chains use public keys. Shadownet ran with a throwaway key. Mainnet anchoring, hosting and anything that spends money or uses the founder’s keys waits for him.

   Why. Nothing can be lost while nothing has value, and every claim can be stated plainly: built and tested locally.

   The cost. No outside users, no real network conditions, and the art certificate stays a rehearsal: no certificates issued or promised.

6. D6 · AGENTS IN WORKTREES, WITH ADVERSARIAL REVIEW

   Each package gets a written brief and its own git worktree, owns disjoint files, and is reviewed by agents told to break it before it merges.

   Why. Parallel lanes merge cleanly when they can’t touch the same files, and a reviewer who must find faults finds different ones than the author’s tests.

   The cost. Coordination overhead: briefs, merge order, wasm pins that conflict whenever two lanes rebuild. And reviewers can be stopped: OpenAI’s content filter halted one review twice.

04 · THE EVIDENCE · FIVE EXHIBITS

Numbers and bugs, from the repository.

Exhibit 1 · The build in numbers

Exhibit 2 · Bugs caught before anything had value

The leaf-count counterexample, verbatim from Codex’s report: root = node(node(A,B),C); a proof for A with leaf_count=3, siblings=[B,C] still verifies with leaf_count=4. Read the report.

Exhibit 3 · Codex’s review, AS-01 to AS-07

Severity was assigned by Claude. Every finding has a test that fails on the code before the fix; all seven are fixed and the tests pass. Read the review.

Exhibit 4 · Farming, measured

Scenario output, not a forecast. Across 38 runs and sweep cells, 123.1 million applied transactions, no protocol bug fired. Read the findings.

Exhibit 5 · What still waits on the founder

• A funded anchor keyAbout 10 ꜩ, roughly six days of hourly anchors, in a dedicated account.
• Feed redeploystez-cast and stampz, before the first mainnet anchor, so anchors don’t show up as garbled notes.
• A public nodeHosting, on the final v2 genesis with non-dev keys. Until then, every certificate stays a rehearsal.
• Live wallet signaturesOne real Kukai and one real Temple signature through the UI; today’s vectors come from Taquito’s signer.
• Launch keysA cold sequencer-admin key, genesis acknowledgements, and deploying the presence issuer.
• Wallet and art choicesEight open questions in the wallet scope; the art rail, the Tezos signatures and the certificate signer.

05 · WHAT WENT WRONG

Three mistakes, told plainly.

• THE MISREAD TIMER

  Four minutes became four hours.

  While lanes ran, the process table showed an elapsed time of 04:09. That is minutes and seconds. It was read as hours, and the founder was told the lanes had run for four hours.

  Lesson. Read the units, and report durations from records you can show, such as the commit log.

• THE CONTENT-FILTER STOPS

  A review stopped twice, midway.

  Codex’s unattended review was stopped by OpenAI’s content filter, “flagged for possible cybersecurity risk”: first after about six minutes with no output, then after about ten. The second session had already written tests for seven findings.

  Lesson. Have reviewers commit evidence as they go. Claude committed the tests unchanged and wrote the report from them; coverage is still partial.

• PLACEHOLDER BRIEFS

  Lanes launched on drafts.

  Once, agent lanes were launched with placeholder briefs instead of the finished ones.

  Lesson. An agent does exactly what its brief says. The working rule in the session notes: briefs go in files.

One deviation went right. Codex’s sandbox would not let its Proof of Balance demo open a port. It ran a socket-free demo instead and said so in its report: “This is not a claim that every requirement in the brief was satisfied.”

06 · TEACHING NOTES

For a 60 to 75 minute session.

• The central tensionSpeed against verification. Every round paired building with an attempt to break what was built, and most of the bugs in Exhibit 2 came from the second half.
• What students missTests encode what the author thought of. The nonce wedge, the cross-chain evidence and the impossible leaf count were all found by someone who wasn’t the author, or by a running system.
• A fair criticismNothing here has met real users, real wallets or a real network. The case shows that a careful local build is possible at this pace, not that the result is ready.

07 · DISCUSSION QUESTIONS

Seven questions for the room.

1. The sequencer is a single point of trust. Which of its powers matter most for a town broadcast, and which planned change addresses each one?
2. Codex proved one requirement in its brief impossible, and shipped with the counterexample in a test. When should an agent stop, and when should it ship with a stated deviation?
3. Neither room-attested drums nor tap limits worked alone; together they cut farming by more than six times. What does that say about testing defenses one at a time?
4. Presence tickets move trust from keys to an issuer’s login check. Is that a better trust anchor for a local community? What would you measure before believing it?
5. The nonce wedge surfaced only because a demo agent stopped at block 18. What kinds of bugs need a running system rather than unit tests, and how would you plan for them?
6. A content filter stopped a review twice, midway. How should a team plan for tools that may refuse work partway through?
7. Everything stays local until the founder funds a key and hosts a node. What does waiting cost, and what would launching early cost?

08 · GLOSSARY

Words used in this case.

ATTN

The chain’s only currency, issued only for showing up (a tap) and playing (a drum session), under hard caps. It carries no value today.

Anchor

Every 100 blocks the sequencer signs height, block hash and state root. Some anchors are posted to Tezos as casts on the tez-cast tower.

Block Yard

A static explorer on pointcast.xyz that draws a recorded dev chain as isometric cubes and verifies it in the browser.

chain-core

The whole protocol as a pure, deterministic state machine: no I/O, no clocks, no randomness, no floats.

Dev chain

A chain started with --dev, signed with publicly derivable keys. Nothing on it has value, and claims from it prove nothing.

Epoch cap

The most ATTN one account can earn per epoch: 500 per day by default.

Evidence

A portable file proving the sequencer signed something false, checkable offline with only the genesis params.

Genesis hash

The hash of a chain’s parameters. Every signature commits to it, so nothing signed for one chain works on another.

Mandate

An owner’s on-chain limits for an agent: kinds, channels, rooms, an allowance and an expiry.

pc-sim

The economy simulator. Scenario output, not a forecast.

pc-town / Town Hall

A read-side sidecar that replays the chain itself and serves account, channel and room pages.

Presence ticket

An issuer-signed ticket that lets a tap earn ATTN, at most once per identity per slot.

Proof of Balance

A claim that an account holds a balance, or doesn’t exist, in a sequencer-signed root, checked offline without replay.

Rollup kernel

The state machine running inside a Tezos smart rollup, proven on the SDK’s mock host but not deployed.

Sequencer

The single node that orders transactions and seals blocks.

Shadownet

A Tezos test network. Anchors were posted there with a throwaway key.

Sybil

One actor behind many keys or identities, farming rewards meant for many people.

VERIFY

The button that makes your browser replay every block with the chain’s own code, pinned by sha256.

Worktree

A separate checkout of the same git repository, so parallel agents can build without touching each other’s files.