arcline
Illustrative product concept. Data shown is simulated to demonstrate the attribution model.
Rules Engine

Your allocation policy, running on our intelligence

Bring your own rules for how tokens get allocated — by team, model, consequence, or budget. Arcline evaluates every work item against them in real time, proposes better rules from your data, and self-tunes inside approval gates you control.

Active policies4
Evaluations fired · 30d608
Monthly savings$38K
Value protected$142K

Bring your integration

Point your gateway, proxy, or agent framework at the policy API. Rules evaluate inline on every request.

Rules as code

Policies are a versioned condition → action DSL. Manage them in the UI or sync from your repo via CI.

Approval gates

Set autonomy per category. The engine tunes freely inside bounds, and escalates anything outside them.

Policy impact simulation
Projected 12 weeks with the current active rule set vs. no rules engine — spend compounds down while attributed value compounds up
Value · no rules ($K)Value · rules active ($K)Spend · no rules ($K)Spend · rules active ($K)
Allocation policies
Your org's rules, expressed as condition → action. Toggle to enable; Arcline evaluates every work item against the active set in real time.

Downgrade low-consequence work

Org-defined

Route work items scored below complexity 3 and consequence 2 to a cheaper model tier.

complexity < 3 AND consequence < 2route_model(tier = 'efficient')
Fired 342× · 30dSaves $18K/moSuccess 96%Scope: All teamsLast fired 12 minutes ago

Protect high-consequence work

Org-defined

Force frontier-tier models for payments, auth, and incident-response work regardless of budget pressure.

repo IN ('payments', 'auth') OR label = 'incident'route_model(tier = 'frontier', override_budget = true)
Fired 87× · 30dProtects $142K/moSuccess 99%Scope: Infra · PaymentsLast fired 1 hour ago

Retry budget circuit breaker

Arcline intelligence

Cut token allocation 50% for any person-model pairing whose rolling 14-day ship rate drops below 40%.

pairing.ship_rate(14d) < 0.40throttle_allocation(0.5) THEN notify(lead)
Fired 23× · 30dSaves $10K/moSuccess 91%Scope: All teamsLast fired 3 hours ago

Verbosity guard

Self-healing

Flag sessions exceeding 2.5x median token volume for equivalent complexity; suggest context pruning.

session.tokens > 2.5 * median(tokens | complexity)flag_session() THEN suggest(playbook = 'context-pruning')
Fired 156× · 30dSaves $6K/moSuccess 88%Scope: All teamsLast fired 26 minutes ago

Weekend batch deprioritization

Org-defined

Shift non-urgent bulk analysis jobs to off-peak windows with cheaper batch pricing.

job.type = 'batch' AND priority < 3defer_to(window = 'off-peak')
Fired 0× · 30dSaves $3K/moSuccess 94%Scope: Data ScienceLast fired 14 days ago
Proposed rules
Arcline watches your spend-to-value data and drafts new policies. Approve to activate — every rule stays inside your approval gates.

Route Legal contract review to Sonnet tier

high93% confidence

Claude Sonnet matches Opus ship rate (94% vs 95%) on contract-review work at 42% of the cost. Quality delta is inside noise for consequence ≤ 3.

Evidence: 212 work items · 8 weeks · Legal team

team = 'Legal' AND task = 'contract-review' AND consequence <= 3route_model('claude-sonnet-4.5')
+$11,200/mo savings, no measured value loss

Raise allocation for top-decile pairings

high89% confidence

The top 6 person-model pairings are hitting allocation ceilings weekly while returning 19x. Every throttled week costs ~$34K in unrealized value.

Evidence: 6 pairings · 4 consecutive weeks at ceiling

pairing.roi(30d) > 15 AND pairing.utilization > 0.95raise_allocation(1.5x, review_after = '30d')
+$34,000/mo unrealized value recovered

Anomaly: Gemini retry loop in CI pipeline

high99% confidence

A CI automation account is burning 4.2M tokens/day in a retry loop against a failing test harness — 96% of attempts rejected since Tuesday.

Evidence: 1 account · 4.2M tokens/day · 96% rejection

account = 'ci-bot-7' AND rejection_rate(24h) > 0.9suspend_allocation() THEN alert(oncall)
-$2,900/day waste stopped immediately

Drift: Support deflection quality declining

medium84% confidence

Support's assisted-resolution value per token has drifted -18% over 6 weeks as ticket complexity rose. Current model tier no longer fits the workload.

Evidence: 1,840 tickets · 6-week trend

team = 'Support' AND value_per_token.trend(6w) < -0.15escalate_model_tier() AND refresh(playbook = 'support-triage')
Restore ~$8,400/mo in deflection value
Self-healing loop
The engine tunes thresholds, rolls back mistakes, and extracts new patterns — inside bounds you approve.
Autonomy level72%
x

72% of tuning actions execute without human review. Raise or lower this per policy category.

  1. Tightened verbosity guard threshold 2.8x → 2.5x

    False-negative rate dropped after last week's loosening; re-tightened within approved bounds. Savings up 11% with no new false positives.

    Today · 09:41 · Auto-tuned
  2. New pattern extracted from top pairing

    Dana Okafor × Opus context-staging pattern codified into playbook PB-31; candidates identified in 3 teams for propagation.

    Today · 07:15 · Learned
  3. Rolled back aggressive batch deferral

    Deferral window pushed a priority-3 job past SLA. Policy reverted to previous window; incident logged and bounds narrowed.

    Yesterday · 22:03 · Rolled back
  4. Circuit breaker fired 3x for same pairing

    Repeated throttling of the same pairing suggests a workload mismatch, not a usage problem. Escalated to team lead with model-fit analysis attached.

    Yesterday · 14:37 · Escalated