gHashTag · gHashTag · Sep 23, 2026 · Sep 23, 2026
@@ -383,3 +383,150 @@ XC7A200T dice: **232 LUT total, zero DSP, zero SRL**, and **100/100 layer
 agreement** with the reference model on 100 real UNSW-NB15 rows. Every value on
 the wire is produced by the die before it; the host shifts bits and performs no
 arithmetic on any payload.
+
+---
+
+## 6. The swarm: agent fleets, donated compute, and what we may not claim
+
+Sections 1-5 position the **silicon**. They were written when that was the whole
+product. It is not any more: a Queen dispatches a swarm of coding bees that take
+issues, open pull requests and get reviewed; a leaderboard ranks the lanes
+people lend; and the stated goal is rewriting the stack into `.t27`. This
+section is the same discipline applied to that half. Surveyed live 2026-09-23;
+every vendor claim is from the vendor's own page.
+
+### 6.1 The finding that comes before any positioning
+
+**"Donate your API key to the swarm" is prohibited in explicit words by all
+three major providers.** This is not a grey area and there is no carve-out for
+non-commercial or charitable use:
+
+| provider | clause |
+|---|---|
+| OpenAI, Business Terms 3.3(g) | may not "buy, sell, or transfer API keys from, to, or with a third party" |
+| Anthropic, Consumer Terms 2 | "You may not share your Account login information, Anthropic API key, or Account credentials with anyone else" |
+| Google, APIs ToS 4(b) | "Developer credentials ... may not be embedded in open source projects"; 4(a) bars sublicensing an API to a third party |
+
+Also barring transfer in their own words: Cerebras, Mistral, Fireworks,
+Moonshot. NVIDIA's API trial terms bar production use and bar making the service
+"available to others". Groq's AUP bars "orchestrating usage between multiple
+organizations" -- which is the pooling model itself, not a side effect of it.
+
+Three consequences follow, and each is a fact about the **donor**, not about us:
+
+1. A key is not scoped to inference. An OpenAI key can revoke itself, mint new
+   keys and change spend caps: whoever holds it holds the account.
+2. Every clause above puts responsibility for all activity on the account
+   holder. A bee that runs up a bill or trips a policy gets the donor banned.
+3. Per-account rate limits are the point, not an accident, so pooling keys for
+   throughput is the prohibited act rather than an incidental one.
+
+**The compliant shape of the same product** is the one AI Horde and BOINC have
+always used: the swarm dispatches the *task*, the contributor runs the bee on
+their own machine under their own account, the credential never moves, and the
+patch comes back. Lanes, XP and a leaderboard all survive that inversion. It is
+**not built here** as of 2026-09-23, and is named as missing rather than implied
+as available -- in this file and in the public tutorial at
+`https://t27.ai/blog/how-to-join-the-swarm/`.
+
+### 6.2 Adjacent products (alphabetical by category)
+
+**Autonomous coding-agent fleets.** Claude Code, GitHub Copilot's coding agent
+with Agent HQ, Cursor background agents, Devin, Google Jules, OpenAI Codex,
+Factory droids, Trae; open-source: OpenHands, SWE-agent / mini-swe-agent, Goose
+(now under the Linux Foundation's Agentic AI Foundation), Zed, Aider.
+
+- **What they do better:** the plumbing, almost entirely -- per-agent container
+  or worktree isolation, hard session caps, per-child budget caps, one-branch-
+  one-PR invariants, merge-queue and CI-failure handling.
+- **What we are not:** we do not offer a hosted product, an IDE, a support
+  contract or an SLA, and we make no claim to their model quality.
+- **The gap nobody has closed:** concurrency *governance*. Only Jules (3/15/60),
+  Trae (2/10/15/20) and Claude Code publish a fleet cap at all; Cursor, Copilot,
+  Devin and Codex say "multiple". How many bees is right is an open question for
+  the vendors too. Cross-agent deduplication and maintainer consent are likewise
+  unsolved everywhere.
+
+**Donated and crowdsourced compute.** AI Horde is the true reference point, not
+a competitor: contributors run workers and earn **kudos**, which buy queue
+priority and nothing else, with a burn ("horde tax") per request and an explicit
+rule that kudos are "not a currency" and that monetising them is "an existential
+threat". BOINC and Folding@home are the classic donation model. Petals is
+dormant; Prime Intellect's p2p protocol repository is archived; Exo pivoted to
+personal device clusters; Bittensor and Gensyn are token-financialised.
+
+- **The warning in the data:** every volunteer-compute network here is small,
+  dead, pivoted, or paid in tokens. AI Horde had 44 live workers when measured;
+  Folding@home is at roughly 3.2 PFLOPS against a COVID-era peak near an
+  exaFLOP. **Plan for tens of lanes, not thousands.**
+
+**Gamified contribution.** Hacktoberfest abolished PR counting; OSS Insight has
+switched its leaderboards off; OnlyDust distributed $18M and shut down; Algora
+pivoted to recruiting; Gitcoin's GG24 raised $36,657 in crowdfunded donations
+against GG23's $95,278. Kaggle's ladder survives unchanged since 2015 because
+medals are percentile-capped and points decay while tiers are permanent.
+Tokscale ranks people by how much their agents *spent* -- the only live instance
+of anything like lane XP, and it scores consumption, not output.
+
+- **What the evidence says about ladders:** badges alone have no measured effect
+  on contributor quality; contributors rank competition and leaderboards
+  *lowest* among gamification elements; steady weekly cadence predicts retention
+  where front-loading predicts churn; and money attracts volume rather than
+  quality (curl's confirmed-vulnerability rate fell below 5% once bounty farming
+  arrived). The design rules worth copying are AI Horde's and Kaggle's: make the
+  reward functional rather than cosmetic, give it a sink, forbid monetisation
+  early, cap by percentile, decay the points, and gate on CI rather than on
+  human attention.
+
+### 6.3 What we do not claim (swarm)
+
+Section 2 exists to keep us out of races we are not running. These belong on the
+same list:
+
+1. **Not "self-improving".** AI4AI-Bench measures the best system at 0.250 where
+   0.1 is the algorithm the repository already ships, and most submissions never
+   change how the model learns at all. The Darwin Godel Machine's headline
+   20 -> 50% is a **200-task subset** of SWE-bench Verified; its Polyglot figure
+   falls from 38.0% on a 50-task subset to **30.7%** on the full set, at roughly
+   **$22,000 per run**. We do not claim recursive self-improvement, and we do not
+   report subset numbers without the word "subset" in the same sentence.
+2. **No throughput, quality or benchmark parity** with any hosted agent product.
+   We publish no SWE-bench or Terminal-Bench figure of our own.
+3. **No reward.** XP ranks lanes by work the dispatch records; it converts to no
+   money, no equity and no token, and no such thing is offered.
+4. **Not a network.** Lanes are single credentials in one operator's
+   environment, not a distributed compute market.
+5. **Star counts are not adoption.** Two repositories in this space carry 39k
+   and 73k stars with single-author commit histories and near-zero issue
+   activity; we cite neither, and ask the same scepticism of any number here.
+
+### 6.4 What we do claim (narrow, defensible)
+
+1. **The rules are executable.** `apps/website/specs/catalog/onboarding.t27` in
+   `gHashTag/trinity` compiles, evaluates its own test blocks, and is published
+   only if they hold -- so a claim in the rules cannot drift from what is true
+   without failing a build. When XP shipped, the spec's own statement that there
+   was no ranking became false and had to be corrected in the same week.
+2. **Capacity is derived and published**, not typed into a variable and
+   forgotten. Only three vendors publish a fleet cap at all.
+3. **The bootstrap, not the intelligence, is what we ask you to verify.**
+   "Two successive self-builds are byte-identical" is a forty-year-old,
+   falsifiable test a stranger can run, and it is a provenance claim rather than
+   a capability claim.
+4. **Open, end to end.** Devin, Jules, Cursor, Copilot and Claude Code's cloud
+   side are closed products; the corpus, the compiler and the board here are
+   readable without an account.
+
+### 6.5 Open, and named as open
+
+- Whether a contributor-side bee (6.1) is worth building here at all, given that
+  every volunteer-compute network surveyed is small, dead or financialised.
+- Nobody ranks agent-authored *contributions*; Hacktoberfest deleted its PR
+  counter precisely because agents made farming trivial. Whether lane XP avoids
+  that failure is unproven, and the honest answer today is that it is untested.
+- A 2026 result reproduced Thompson's "Trusting Trust" attack against
+  self-modifying agents: poisoned benchmarks made three published systems evolve
+  code that disabled certificate validation, and the contamination persisted
+  after re-evolving against clean benchmarks. Any system that edits its own
+  toolchain, including this one, owes that finding an answer it does not yet
+  have.
@@ -1,6 +1,6 @@
 # NOW -- Trinity t27 sync
 
-Last updated: 2026-08-09
+Last updated: 2026-09-23
 
 ## typecheck: warnings are PRINTED, and the unused-variable false positive is fixed (Refs #1948)
 

@@ -0,0 +1,9 @@
+# NOW -- COMPETITORS.md now positions the swarm, not only the silicon (2026-09-23)
+
+## The landscape survey the document was missing (Closes #4603)
+
+- `COMPETITORS.md` gains section 6: agent fleets, donated compute, gamified contribution and self-improving systems, surveyed live 2026-09-23 with every vendor claim taken from the vendor's own page.
+- The finding that comes before any positioning: **"donate your API key" is prohibited in explicit words** by OpenAI ("buy, sell, or transfer API keys from, to, or with a third party"), Anthropic ("You may not share your ... Anthropic API key ... with anyone else") and Google ("Developer credentials ... may not be embedded in open source projects"), with no non-commercial carve-out. Cerebras, Mistral, Fireworks and Moonshot bar transfer too; NVIDIA bars making the service available to others; Groq bars orchestrating usage between organizations, which is the pooling model itself.
+- The compliant shape is the AI Horde / BOINC inversion: the swarm dispatches the task, the contributor runs the bee on their own machine, the credential never moves. Not built; named as missing in the document and in the public tutorial rather than implied as available.
+- Two claims are added to the "what we do not claim" list. **Not self-improving**: AI4AI-Bench puts the best measured system at 0.250 where 0.1 is the algorithm the repo already ships, and the Darwin Godel Machine's 20 -> 50% is a 200-task subset whose full-set Polyglot figure falls from 38.0% to 30.7%, at about $22,000 per run. **Star counts are not adoption**: two repositories in this space carry 39k and 73k stars with single-author histories and near-zero issue activity.
+- The warning in the donated-compute data is the one worth acting on: every volunteer network surveyed is small, dead, pivoted or token-financialised. AI Horde had 44 live workers when measured; Folding@home sits near 3.2 PFLOPS against a COVID-era peak close to an exaFLOP. Plan for tens of lanes, not thousands.