Queue depth
37
▲ 6 vs. yesterday
Avg. time to decision
4.2 hrs
▼ 0.6 hrs
Reviewed today
61
▲ 8 vs. yesterday
Escalated to specialist
5
▲ 2 vs. yesterday
Confidence + recommended action shown here mirror the evidence package a reviewer opens (Workflow 1, node D). Confidence scale: 1 (Not Harmful) – 10 (Harmful).
Pipeline health
INGESTION
Nominal
SCAN LATENCY
3.1 min
PRECISION (7d)
0.83
RECALL (7d)
0.71
MODEL VERSION
v0.4-antisem
Confirmed violations
214
▲ 12% vs. prior period
Reported to Meta
198
92% of confirmed
Networks flagged
9
3 new this period
False-positive rate
17%
▲ 2 pts vs. prior period
Confirmed violations over time 30 days
Flagged by AI
Confirmed by reviewer
By category
Category labels match the client's existing submissions log (Evidence Addendum §1) rather than Figure 1's broader taxonomy.
By platform
Detection funnel 30 days
1,340
AI-flagged
100%
1,340
Reviewed
100%
214
Confirmed violation
16%
198
Reported to Meta
92% of confirmed
Repeat-offender networks
| Network ID | Platform | Linked accounts | Category | Status |
|---|---|---|---|---|
| NW-014 | 6 | Hate Speech | Reported | |
| NW-013 | 4 | Dangerous Orgs | Reported | |
| NW-011 | 9 | Terrorism | Under specialist review |
Sourced from Workflow 2 (Network & Account Analysis) — node I.
By geography (self-reported / inferred)
United States128
United Kingdom31
Germany22
Canada17
Other / unresolved16
Geography inference method not yet specified — placeholder pending data-source confirmation.
Inference spend (24h)
$11.30
under $200/day cap
API calls
1,587
via tool broker
Draft latency (p95)
31 s
target < 45s
Batch throughput
247/hr
cases scored
Token consumption & spend by agent last 24h
| Agent | Runs | Tokens used | Model | Spend |
|---|---|---|---|---|
|
Moderation Triage Agent
Workflow 1 — flag, score & draft evidence
|
1,340 | 8.2M | local-70b | $9.40 |
|
Network & Account Analysis Agent
Workflow 2 — repeat-offender network detection
|
247 | 3.1M | local-8b | $1.90 |
Spend reflects internally-deployed model inference only — no fact content leaves the network.
Spend trend 7 days
Token share by agent
Triage runs on every AI-flagged item; Network Analysis runs only when a linked-account signal is present.
Agents in roster
2
1 group
Active
2
running normally
Paused
0
held, resumable
Stopped
0
not accepting runs
Moderation Triage Agent
Workflow 1 — flags content, scores confidence, drafts the reviewer evidence package
Last run completed Today, 6:15 AM · 1,340 items processed
RUNS TODAY
1,340
ERROR RATE
0.9%
TOKENS (24H)
8.2M
SPEND (24H)
$9.40
SCHEDULE A RUN
Typical run takes ~22 min based on recent throughput — estimated completion shown once you pick a time.
Network & Account Analysis Agent
Workflow 2 — links flagged accounts into networks and surfaces repeat offenders
Last run completed Yesterday, 4:47 PM · 247 items processed
RUNS TODAY
247
ERROR RATE
0.4%
TOKENS (24H)
3.1M
SPEND (24H)
$1.90
SCHEDULE A RUN
Typical run takes ~9 min based on recent throughput — estimated completion shown once you pick a time.
Every start, pause, stop, run, schedule and model change on this screen is recorded here, most recent first. Session only — not persisted.
Console notes
Pausing or stopping an agent here does not affect any run already suspended at a reviewer's approval gate — it only applies to new runs picking that agent going forward. Changes on this screen apply to this session only.
Active entries
—
referenced before every run
Retired entries
—
kept for audit history
Last updated
—
—
Used to attribute anything you add, edit or retire below.
NEW MEMORY ENTRY
| Category | Entry | Detail | Status | Added by | Added / updated | Action |
|---|
Both agents read this knowledge base before each run, in addition to their standard model and classifier stack — it's how the team teaches the agent new patterns without waiting on a model retrain. Retiring an entry keeps it visible for audit but stops the agent from using it.
Sampled from the client's 2026_Meta_Submissions.xlsx (8 monthly tabs, 2,190 rows total) — this view shows 12 representative rows rather than the full log. AI confidence is a simulated pipeline score for this prototype, shown on a 1 (Not Harmful) – 10 (Harmful) scale. Approving a case here adds it to Human Approved in the exact spreadsheet column format.
Items analyzed
—
by Moderation Triage Agent
Flagged harmful
—
routed to review queues
Cleared, not harmful
—
no policy violation found
Avg. confidence
—
across all analyzed items
This is the agent's full analysis output before human review filters it down to the HITL queue — includes items the agent cleared with no further action. Confidence shown on a 1 (Not Harmful) – 10 (Harmful) scale. Simulated data for this design comp.
Last report sent to Meta: Today, 5:00 AM
SCHEDULE REPORTS TO META
Batches every case currently in this queue and sends it to Meta on the interval you choose, starting from now.
No cases reported to Meta yet.
Mark a case Harmful from its View Details panel to see it appear here.
Mark a case Harmful from its View Details panel to see it appear here.
In a live deployment this queue feeds the signed report / platform API call described in Workflow step 9.
No cases escalated yet.
Choose Escalate from a case's View Details panel to queue it here.
Choose Escalate from a case's View Details panel to queue it here.
Escalating a case sends a notification email to the specialist team — see Alerts for an example.
Email alerts 4 samples
These are static mockups to illustrate tone and content — no emails are actually sent from this design comp.
SMS alerts 2 samples
⚠️ Fluid Spear: Queue depth alert — 37 open cases, 1 overdue SLA (FS-1013). Avg time to decision 4.2 hrs. Reply STOP to opt out.
✅ Fluid Spear: Moderation Triage Agent run complete — 1,340 items processed, 42 flagged for review. Reply STOP to opt out.
These are static mockups to illustrate SMS tone and length — no texts are actually sent from this design comp.
No cases approved yet.
Approve a case from All open cases or step 8 of the Workflow walkthrough to see it appear here.
Approve a case from All open cases or step 8 of the Workflow walkthrough to see it appear here.
Case ID, Date, Type, Platform, Agent recommendation, AI Conf and Status are shown in the same column order used across every case table in this app.
1. Pick a sample content item 6 of 2,190 logged posts
Terrorism
ISIS watermark repost (repeat account)
Terrorism
ISIS reel with explicit threat caption
Hate Speech
Samidoun reference — sole FB hate-speech case
Extremism
WWII SS recruitment footage, reposted
Dangerous Orgs
Samidoun organizational branding
Hate Speech
Coded antisemitic hashtag cluster
All six are real rows from the client's 2026_Meta_Submissions.xlsx (Facebook & Instagram only). Pipeline scores are simulated for this walkthrough — no live model calls or platform connections are made.
2. Follow the case through the pipeline step 1 of 8
1
Start
2
Ingest & Preserve
3
Analyze
4
AI Recommendation
5
Human Review
6
Human Decision
7
Outcome
8
Results
Automated
AI decision
Human action
Outcome
Learning loop
Click a step you've already reached to jump back, use Back / Next, or press Play — Play stops at Human Decision and waits for you to click Confirm, Reject or Escalate before it continues.
STEP DETAIL
Step 1 — Start
SYSTEMS INVOLVED
STEP SIGNAL
CASE & POLICY
CASE IDWF-1
ACCOUNT
PLATFORM
CATEGORY
DATE FOUND
KEYWORDS / NOTES
LEARNING LOOP
Not yet engaged — reached once a human decision is made at Step 6.
*A small share of "do not report" cases are still audit-sampled to Human Review.
Profile
NAMEHaneeth
ROLETrust & Safety Lead
EMAILhaneeth@fluidspear.ai
MOBILE NUMBER
TEAMContent Moderation — Project Spear
LAST LOGINToday, 9:02 AM
Account settings
Settings shown for this design comp only — nothing here is persisted or sent.
Access & permissions
| Capability | Access level |
|---|---|
| Review & decide HITL / All open cases | Full access |
| Change agent verdict (Harmful / Not Harmful / Escalate) | Full access |
| Start, pause, stop or schedule agents | Full access |
| Edit Agent memory (patterns, hashtags, exceptions) | Full access |
| Schedule & export Meta reports | Full access |
| User & access management | Admin |
Role-based access shown for illustration — this prototype does not enforce permissions.