SSlops Discuss calibration

Managed governance

Quality signals should advise routing, not secretly rewrite it.

Slops turns delivery outcomes into decomposed evidence for model-routing proposals. Until customer-specific calibration and approvals exist, the system observes, explains, and proposes only.

No floor-lowering No profile auto-write No autonomous apply Evidence before score

Evidence classes for a public-safe demo

The page shows classes of calibration evidence, not private evaluations, customer payloads, raw logs, or internal operating records.

Durable positives

Work that cleared required gates and stayed clean long enough to be useful as positive evidence.

  • Delivery completed under the required review gates
  • No later hard negative attached to the same routed work
  • Used as a proposal input, never as permission to lower floors

Source-backed negatives

Corrections, reverts, failed gates, or rollback-style outcomes stay visible by source class.

  • Negative evidence remains decomposed and inspectable
  • Operator review separates true failure from noisy context
  • Proposal logic may recommend tighter routing, not silent punishment

Churn texture

Rework and follow-up edits are treated as texture until the customer agrees how to interpret them.

  • Useful for spotting friction and support load
  • Not enough by itself to score a model or a worker
  • Shown beside positives and negatives, never collapsed early

Proposal-only until calibrated

Output quality becomes safe when the operating contract says what the evidence can do. Slops keeps that contract explicit and managed.

01 / Observe

Capture decomposed evidence

Collect durable positives, backed negatives, and churn texture without exposing private payloads or raw customer work.

02 / Calibrate

Agree what each signal means

Define customer-specific thresholds, noise handling, approval rights, and the evidence needed before routing can change.

03 / Propose

Generate above-floor routing suggestions

Suggest profile adjustments only above existing capability, risk, identity, security, and approval floors.

04 / Approve

Apply through a human-controlled change path

No outcome score can auto-write a routing profile, lower a floor, or autonomously apply a policy change.

What a route proposal can say

A proposal can explain evidence and tradeoffs. It cannot pretend a score is settled truth, and it cannot act outside the customer-approved operating boundary.

Allowed

  • Show which evidence class influenced a recommendation
  • Recommend a stronger model for a gated work class
  • Recommend a cheaper model only when all required floors still hold
  • Ask for more calibration data when confidence is thin

Not allowed

  • Lower capability, security, identity, or approval floors
  • Write model-routing profiles from an outcome score
  • Use private payloads or raw journals as demo evidence
  • Let an uncalibrated judge influence production routing

The managed service starts with visibility and review. Automation expands only after the customer agrees the evidence, thresholds, and approval receipts are strong enough.