EVIDENCE LOG
21 OCTOBER 2026 · RESULTS FEED

What have we
actually proven?

This page separates observed behavior from plans and unknowns. Findings load from a reviewed JSON file so results can be updated without rewriting the site.

FEED STATUS · Loading…
RECENT BOUNDED VALIDATION · OBSERVED

Repair accepted; original baseline rejected.

REJECTED

Initial five-task-turn baseline rejected for missing substantive result.

ACCEPTED

Bounded two-turn repair accepted; scenario tally: 7 task activations, 1 setup activation and 1 failed auth-mode attempt.

CHECKS PASSED

15 Go top-level checks plus vet passed for UI/help. Integration checks passed 10/10.

These are scenario-specific counts and checks, not a general benchmark or savings claim.

REL01 · VERIFIED INSTALLATION

Private preview v0.1.0-preview.1 is live

Unsigned and experimental; compiled per-user binaries, no token/profile bundled. Isolated Windows installation checks and real Windows/worker operation checks passed. Authenticated release download and checksum verification passed for all four curated assets.

WINDOWS

Install, offline UI, repeat start, wrong-nonce rejection, stop/restart, upgrade, manual removal and user ACL checks passed.

LINUX WORKER

User-local install, start/restart, doctor and status passed. Existing board baseline remained unchanged.

REVIEW

39 test entries, including 2 helper fixtures; independent review cleared source after two fixed bugs. Authenticated download and checksum verification passed for all four curated assets.

Windows ZIP: 4,278,933 bytes · SHA-256 5a747969d0d49bda7f18464a72c41eb4f850b48a2bb37ac4d0a7b460bb4fdcda
Linux TAR: 4,053,451 bytes · SHA-256 7273394d4e8ac4f1790d8638757ea6cf5ba49cfa7e1b3dc8abb64a0552bccd40
Source revision 908748b78cf163bc99336d0ea5e922287ce7ac38.

Private release page ↗
RESULTS FEED

Awaiting reviewed results

No experiment numbers or aggregate claims are published until a safe results file is supplied.

PENDING

Experiment results have not been published

Reviewed, sanitized results will appear here with scope, method, date and limitations.

OBSERVED BUILDING BLOCKS

Real transports, bounded claims.

OBSERVED

Native question/action paths

Disposable fixtures demonstrated bounded answers and actions through Codex and Claude native transports. This does not prove every question is discoverable or resumable.

Curated project research · 21 Oct 2026
OBSERVED

Capability visibility has limits

App-server schema and SDK references expose version- and account-specific fields. Schema visibility alone does not prove populated telemetry or admission enforcement.

Allowance/discovery check · 21 Oct 2026
OBSERVED

Quota monitoring is separate

Hermes has deterministic account-separated allowance collectors and history. That is monitoring, not DELEGATOR board admission or token-to-quota conversion.

Reviewed runbooks and implementation · 21 Oct 2026
SCOPE & INTERPRETATION

Do not confuse
available with working.

IMPLEMENTED: Experimental Go server, CLI and read-only local UI exist. Private preview v0.1.0-preview.1 is live; isolated install and operation checks passed. Authenticated release download and checksum verification passed for all four curated assets. The UI manually refreshes central reachability, declared capabilities and explicitly registered session references; it does not auto-poll. Registration is not a heartbeat or proof of live activity. Runtime evidence is independently observed, stale or unknown with its age. The temporary board passed bounded private loopback and managed connection-bound SSH checks; production is not deployed.

PROPOSED: Native live control and session import remain proposals; allowance enforcement and background discovery are not implemented. An explicit scoped sessions-list metadata adapter is implemented separately.

UNSUPPORTED: Explicit scoped metadata inventory is opt-in and does not auto-register or upload. Native runtime status remains unknown. Independent background discovery, transcript parsing, native import and control are unsupported.

Public results must be aggregate and sanitized. No private transcripts, session IDs, host details, credentials or account information.

NEXT EXPERIMENT · PROPOSED

Typed contracts and bounded wakeup

Discoverable CLI / MCP

Define typed commands and a result-content schema so clients can inspect capabilities and validate results.

PROPOSED · no savings claim

Separate message references

Represent context_ref (task context) and reply_to (destination for an answer) independently.

PROPOSED

Event-driven coordinator wakeup

Wake on relevant state changes with explicit bounded handling instead of manual orchestration. Any reduction remains unmeasured.

PROPOSED · no savings claim
NEXT EXPERIMENTS

Proofs still needed.

01 · Allowance observation

Replay account/window, freshness, reset, duplicate and aggregation fixtures.

PENDING

02 · Durable hold

Prove threshold and outage behavior; held jobs remain inspectable.

PENDING · policy details open

03 · Override & recovery

Show authorized expiry, one launch per job, safe uncertain-dispatch recovery.

PENDING · no paid/account changes

04 · Native integration

Bounded disposable start, hold and release proof per supported provider.

PENDING · no unlimited workloads