Maintain judgment quality

Request an invisible-maintenance recommendation for a specific operator.

Judgment decays. COHESION returns the intervention most likely to arrest that decay without the operator knowing it happened.

What this is

POST /v1/maintenance/recommend takes an operator ID and returns a difficulty-adapted intervention: a calibration injection, a recommendation withholding, an unranked presentation, or a targeted exercise.

When to use

Automated loop: whenever JIS drops below 75 or a specific dimension weakens.
Scheduled cadence: once per shift or once per day per operator.
Triggered: right after a critical error or post-error recalibration event.

Invisibility is non-negotiable

Operators must not be able to distinguish a maintenance interaction from a normal workflow interaction in real time. Post-hoc audit labels only. No visible flags, badges, or notifications. Disclose at onboarding that maintenance is active; do not reveal which specific interactions are maintenance.

Request a recommendation

Python

rec = client.maintenance_recommend(operator_id="analyst-42")

print(rec.intervention_type)
print(rec.target_dimension)
print(rec.difficulty)
for exercise in rec.exercises:
    print(exercise.title, exercise.payload)

TypeScript

const rec = await client.maintenanceRecommend({ operator_id: "analyst-42" });
console.log(rec.intervention_type, rec.target_dimension);

cURL

curl -X POST https://api.cohesionauth.com/v1/maintenance/recommend \
  -H "X-API-Key: $COHESION_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"operator_id": "analyst-42"}'

Intervention types

Type	Typical frequency	Purpose
Calibration injection	2-8% of interactions	Exercise error detection with subtle safety-netted errors
Recommendation withholding	3-10%	Capture independent assessment before showing AI output
Unranked presentation	2-5% of ranked flows	Measure independent evaluation without ordering cues
Difficulty adaptation	continuous	Tune all above to current JIS and weakest dimension

Difficulty adaptation rules

JIS	Calibration difficulty	Frequency	Focus
90-100	High (subtle only)	2-3%	Maintain all
75-89	Moderate-high	3-5%	Lowest dimension
60-74	Moderate	5-7%	Deferral + error detection
40-59	Low-moderate	6-8%	All dimensions
`< 40`	Low (clear errors)	8% max	Rebuild error detection + independent performance

Errors you might see

Status	Meaning	Fix
`404`	Operator not found	Confirm the operator has scored interactions
`422`	Insufficient data for recommendation	Need 50 interactions or 10 days of monitoring

Next step

Feed the recommended intervention back as a normal scored interaction: track AI decisions
Full endpoint: POST /v1/maintenance/recommend