# Remaining work

The [status report](status.md) separates merged source, reported local checks
and narrow hosted recovery from broader release qualification. These increments remain open;
a seed pack, proposed payload or architecture rule does not install or admit them.

| Work | Current state | Completion evidence |
| --- | --- | --- |
| Hosted recovery and release qualification | Latest recorded Tessera PR13/home receipts inspected; fresh public Tessera health/instructions reads passed; PR12 smoke remains separately attributed | Native artifact/report readback and authorization checks, PR13 learning paths, a completed Frontier cycle, broader operation and fresh browser checks; bind runtime evidence to exact releases |
| Same-namespace blob governance | PR12 reports first-uploader visibility for identical bytes within a namespace | Explicit per-uploader governance, migration evidence and cross-principal scope tests before broader reuse claims |
| Provider operation and qualification | Shared-health PR10 and accounting PR11 merged with exact-head/main CI; 197 Python/150 Worker and overlapping independent 132/115 checks; Herma deployment held | Deliberate rollout with preserved health/budget state and scoped evidence. Qualify changed standard-tier request fields only through an authorized future check; prior live OpenAI success belongs to its earlier source/request. Preserve manual pause and caller proxy authority; review remaining pricing/alias and legacy-path limits. |
| Capability research loop | Research seeds and community candidate authoring exist; no active collector/mailbox/recurring scheduler | Pinned real profile, permitted sources, deduplicated provenance, comparable baseline/trial and a useful adoption or no-change decision |
| [Research specialist adaptation](fine-tuning-pilot.md) | Actual tiny random-Llama CPU/GPU smoke; pretrained 1B/3B and Glimmer recipes unrun | Reviewed permitted examples, disjoint source/project families, unchanged-model baseline, useful held-out quality/cost/resource comparison and independent admission review |
| [Community seed experiments](community-seeds.md) | Six private draft bundles, 14 noun and 12 verb shapes checked; external code not executed | Choose one bounded owned-fixture trial, verify license/scope, author through existing contracts, execute in an authorized worker and review evidence before admission |
| Pack import and lifecycle | Exports of the code composition and video and motion packs and six local deterministic executions passed | General import/dependency/upgrade handling, source/profile resolution, negative permission tests and explicit admission; preserve existing tested tools while the older pack websites retire |
| Pinned category/profile work packets | Query repairs exist; general resolver/QueryPlan proposed | Deterministic version resolution, source scope, conflicts, required constraints and direct-client parity |
| Governed shim lifecycle and attribution | Scoped rules, provider protocols/plans and explicit attempt evidence implemented | Admission/activation/invalidation/rollback checks; migrate remaining legacy inference paths; captured parent links for deterministic native actions |
| Native and swarm coverage | Seven distinct harnesses across narrow historical/local tests; ten native processes and 100 protocol clients tested separately | OpenMuse TypeScript connector, repeated recovery scenarios, sustained native/separate-principal waves with latency, wait, completion, retries and provider usage |
| Frontier candidate pilot | Daily schedule configuration recorded in latest home release; server repair off; completed hosted cycle not demonstrated | One completed authorized cycle with scoped outputs, then a bounded candidate against baseline/development/holdout/adversarial cases and independent review before admission |
| Delivery and operational surfaces | Completion/delivery ideas proposed; historical 16 rendered banner cases passed | Delivery retries never repeat completed effects; revoked recipient denial; fresh onboarding after recovery, longer operation and native OS clipboard checks |
| Video and motion pack media extensions | Deterministic frame/SVG/metadata checks and seed exports work | Actual encoding/asset pipeline, rendered-media checks and verified media license provenance |

Only nouns and verbs are core entities. Two public surfaces expose them, Eranos and
Agora; narrow infrastructure, Tessera and Herma, supports them; and seed packs bundle
everything else, agent kits included. Project names do not create another
primitive or authority. Ordinary reasoning and self-checks belong to authorized
workers; required additional judgment becomes a review verb. Server inference
needs an explicit scoped allowance and recorded attempts. Keep failures, unknowns
and no-change outcomes. Do not train around harness bugs or tune against the test set.

Frontier's recorded configuration uses hourly due checks for a 24-hour cycle
anchored at 13:00 UTC, with a 10-minute manual cooldown and a stop at
**November 4, 2026, 17:00 UTC (noon ET)**. Server repair is off. Configuration does
not establish completed hosted work or authorize automatic publication.