# Swarm pilot and creative seeds

The system is ready for an invited, scoped pilot. Visitors can copy an agent
starting brief from any of the seven hosted sites, either for available work or
for their own interests. An operator still issues keys for ongoing contribution
work. Anonymous access remains isolated synthetic practice.

The banner was checked in rendered Firefox on desktop and mobile, including
the interest field, both copy paths and the manual fallback. Those tests stubbed
the clipboard API to exercise success and rejection; they do not establish
operating-system clipboard interoperability.

## Native workers

The October 29 matrix exercised installed harness processes, their own tool
selection and actual hosted Tessera operations. Each passing lifecycle required
claim, input read, heartbeat, artifact write, artifact read-back and submission,
plus an independently read execution report. Process exit status alone was
insufficient.

| Harness | Observed coverage | Important qualification |
| --- | --- | --- |
| OpenCode 2.1.2 | Complete scoped lifecycle; earlier anonymous/keyed execution recall runs | Narrow synthetic fixtures on Ollama Cloud |
| Pi 0.73.1 | Complete lifecycle and concurrent creative contributions | Standard extension supplies the scoped MCP transport |
| Hermes 0.20.1 | Complete lifecycle after a calibration follow-up | Fixture-specific typed output contract; earlier attempt stopped before submission |
| Muse Code 1.4.14 | Complete lifecycle with approval checks enabled | Fixture-specific typed output contract; earlier schema mismatch retained |
| nanoMuse 1.0.8 | Complete lifecycle with Sentinel checks and scoped approvals | Python OpenMuse successor; thin stdio transport and typed fixture |
| OpenClaw 2026.10.8 | Complete lifecycle with Docker sandbox enabled | Explicit plugin/profile registration; earlier discovery failures retained |
| OpenMuse TypeScript 0.1.0 | Source inspected | No native Tessera connector verified; not counted as an E2E pass |

These tests do not certify every feature, model or deployment of a harness.
The typed fixture overlay constrains an already known test output. It does not
repair tool arguments or establish general Herma shim compatibility. The
general shim admission, activation and rollback lifecycle remains unfinished.

## Creative alternatives

[Opportunity Lab](https://github.com/ta-amps/tessera-infrastructure-work-and-storage-engine/blob/main/docs/opportunity-lab.md)
turns eleven startup themes into 21 versioned seed nouns and 22 initial verbs.
The source context comes from the public YC Requests for Startups page; the
showcase is not endorsed by YC and does not independently validate its claims.

Two workers may solve the same brief differently. Each attempt keeps a separate
lease, artifact and receipt while referencing the same input nouns. Comparison,
rating, benchmark design and remix procedures are reusable nouns for follow-up
verbs. A proposal to create more seeds does not activate recursive work.

Ten native Pi processes started together in the creative pilot. Seven completed;
three were truncated by an incorrectly imposed 2,048-token test-transport ceiling.
That ceiling was removed. All three affected theme tasks then completed in a
separate three-process retry using native response settings. The original
results remain **7/10 in the first wave and 3/3 in the retry**.
A subsequent full ten-process wave with the corrected transport completed
**10/10**, producing twenty private idea proposals across five themes. It used
71 worker-side Ollama Cloud attempts: 70 returned HTTP 200 and one disconnected
with unknown usage before the harness recovered. One worker also corrected a
rejected submission before completing. The runs shared a scoped cohort principal;
they do not establish ten independent account grants. Novelty, usefulness and
implementation quality have not been measured.

The corrected transport preserves caller token fields and provider defaults.
Test spending is controlled at the run level. A regression fixture verifies
that the transport neither injects nor clamps response-token settings. Historical
failed attempts remain in the observation record.

## Capacity evidence

The protocol stress test uses concurrent logical HTTP workers sharing one fresh
scoped cohort key. It exercises claims, heartbeats, versioned reads and writes,
submission replay and contested compare-and-set. It uses no model calls.
It is distinct from the native-process matrix and from a many-account test.

Before adaptive admission, the 100-client wave hit the fixed sixty-claim quota.
Honoring `Retry-After` allowed 100/100 to finish in about 73 seconds. Current
adaptive-admission results are recorded separately in the [status report](status.md)
and [machine-readable evidence](../status.json), preserving that baseline.

The deployed adaptive version completed 10, 25, 50 and 100 clients with zero
HTTP 429 responses. The 100-client wave took 25.663 seconds including setup.
Its claim headers reported zero admission wait and a window of 32, so this live
wave did not exercise a window change. Injected-pressure tests separately cover
pacing and continued lease maintenance.

Adaptive admission removes the default claims-per-minute ceiling. Bursts wait
fairly, concurrency follows observed service latency, and heartbeat/submission
traffic bypasses the claim queue. Pending-request and waiting-time bounds protect
resources. They are configurable deployment settings, not a throughput claim.
See [the owning implementation](https://github.com/ta-amps/tessera-infrastructure-work-and-storage-engine/blob/main/docs/claim-admission.md).

## Before a broad public launch

An invited pilot can use the existing scopes and receipts now. Broader claims
need repeated and sustained load with separate principals, restart and
disconnection recovery, and measured database/latency behavior. A hundred
protocol clients does not establish a
hundred reliable reasoning workers or unlimited scale.

Research seed authoring is implemented for source discovery, collector proposals,
schedule proposals and scanning tasks. A live newsletter mailbox, activated
collector code and a recurring research scheduler are still absent. Mailbox
creation is usually a small setup task; reliable ingestion, deduplication,
provenance and recovery are the larger integration. Those should create source
nouns and claimable work using the existing fabric.