NEUROCOMMENTING Operations Center

AccountsNot loaded

Authorization and runtime eligibility are separate. Server state is canonical.

Filter loaded resources

Identity	Name	Authorization	Global runtime	Proxy	Metadata
Set a workspace ID and refresh.

Resource importPreview is read-only. Apply is explicit and rollback remains available.

0 Telegram mutations during preview

Resource type Duplicate policy Rows JSON Column mapping JSON Provenance JSON

Batch inspectionNothing previewed

Set a workspace and preview rows.

Campaign New campaign

New campaign context

Objective JSON Product facts JSON

State—

Accounts—

Sources—

Queued / waiting—

Blockers—

Select a campaign.

New campaign: Prepare campaign first (DRAFT → STARTING), then explicitly Run (STARTING → RUNNING). Prepare does not start collection or sending.

To stop: Stop blocks new campaign work (RUNNING or PAUSED → STOPPING); then explicitly Finish stop (STOPPING → STOPPED). Already in-flight external operations may finish.

Campaign send budget

Budget action Maximum actions in window Budget window seconds Campaign cooldown seconds

Workspace, account and actor safety limits remain mandatory.

Assign accounts and sources

Account ID Allowed actions Source ID Source admission Policy key Human policy decision Policy evidence JSON Policy action scope Policy evidence TTL seconds Linked discussion source ID

Membership changes are explicit and audited.

This declares the selected channel Source ID and discussion source only. Telegram must confirm the pair; reading and sending still need separate access and policy checks.

Manual ActionJobCanonical Queue → Timing → Router path

Action Source ID Telegram entity Message ID / reply target Message / emoji Idempotency key Constrain to account ID Verified DM recipient ID Job TTL seconds

QueueNot loaded

Controls are fail-closed server decisions; stale snapshots return conflict and reload.

Job	Action	State	Account	Blocker / next	Controls
Select a campaign.

Action attemptsSelect a job

Attempt state Ambiguity Search attempts

—

Approval queueAPPROVAL and escalated reviews can be edited, approved or rejected. Authorized AUTO comments do not require per-message approval.

Review	Stage	State	Run	Created
Select a campaign and load reviews.

DecisionNo review selected

Select a review to inspect its frozen LLM job TTL.

Action plan JSON Candidate JSON Reject reason

Review evidence and model trace appear here.

Provision versioned LLM pipeline

Provider ID Model Campaign-specific prompt instructions Runtime mode Autonomy ceiling LLM data scope LLM requests per day Keyword rules JSON (required in keyword mode)

LIMITED_AUTO / AUTO do not require approval of every comment. A genuine human session, signed rollout/readiness and runtime safety gates are still required. Selecting AUTO is not proof of production readiness and does not start the campaign.

Configured policy has not been read. No production-readiness claim.

No pipeline changes yet.

One SHADOW run — no sendUse existing collected context and a version for the selected campaign. This calls the LLM under existing egress rules; it does not change runtime policy, approve a review or create human labels.

Actual ContextCase ID Actual Snapshot ID Intelligence Version ID

SHADOW only, PUBLIC request scope, LLM enabled. The server still enforces actual context privacy and provider egress authorization. No automatic retry; a lost response can mean the run already exists. Inspect recent decisions before any further action.

No SHADOW run requested.

Recent LLM decisionsLatest 50 runs for the selected campaign. An ACT decision or job ID does not prove delivery. WATCH is a recorded decision, not proof of an active watcher.

Run	Decision	State	AUTO authorization / reason	Action job	Created
Read decisions explicitly; no background polling.

Bulk workflow composerComments, chats and DMs use the same durable ActionJob safety corridor.

Preview → confirm → execute

Campaign ID Action Request key Autonomy Base payload JSON Targets JSON

[{"target_ref":{"entity":"me"},"payload":{"message":"Controlled test message"},"account_constraint":{}}]

Execution reportNo workflow selected

A preview performs validation without Telegram mutations.

Recent workflows

No workflows loaded.

Account readinessPhysical-account hard gate

Account ID

—

Exact Telegram identity lookup

Reads one explicitly selected identity. Does not send, join, grant source access or approve a campaign. The account must be enabled and not owned by a live worker.

Lookup kind Exact Telegram locator Existing source ID

For linked discussion: first resolve the registered channel, then use its source UUID and returned negative peer ID. This reads channel metadata only; configure the returned group pair separately under Campaign sources. NOT_JOINED does not mean comments are inaccessible.

No identity lookup performed.

Profile ManagerDesired → ActionJob → read-back

Campaign IDFirst nameLast nameBio

—

WarmupDedicated/background with explicit safe targets

Campaign IDAccount IDs (comma separated)ModeSteps JSON

—

MaterialsManual/imported/mixed → canonical ActionJob

Campaign IDNameTextTarget entity

—

Select a conversation—

Conversation history appears here.

OwnershipNever silently switches identity

—

Handoff to account Reason

Multi-account SimulationSynthetic provenance and organic-intrusion boundary

Campaign IDName Participants JSON Steps JSON

—

TaxonomyHuman-reviewed versioned schema

Campaign IDSchema JSON

—

Human labelGold/Hard/Regression requires reviewed evidence

ContextCase IDDataset IDTaxonomy versionLabelExpected action (your review)Set

Choose the label and expected action yourself from reviewed evidence. Model output is not a human label. Nothing is saved until you explicitly click Save.

—

Separation / active learningSynthetic evidence excluded

—