Forecast AI capability atrophy loss
Learn how aggregate fallback capability decays with AI reliance and is preserved by unaided practice using a Bayesian right-censored transition model, then forecast ready/degraded/unavailable capacity and correlated provider-outage economic VaR/CVaR.
What it's for
Quantifies the hidden downside of successful AI adoption: whether unaided operating capability is decaying, how much fallback capacity will remain, and what a correlated outage could cost.
What you give it
Inputs split into evidence read from your connected systems, calibration your team owns, and numerical controls that affect precision but never the result's meaning.
| Field | Type | Role | Required |
|---|---|---|---|
| coefficient_grid_size | integer ≥ 11, ≤ 101 | Your calibration | Optional |
| coefficient_prior_sd | number > 0, ≤ 20 | Your calibration | Optional |
| current_exposures | array of objects (19 fields) | Evidence | Yes |
| horizon_months | number > 0 | Your calibration | Optional |
| max_detail_rows | integer ≥ 1, ≤ 500 | Numerical control | Optional |
| maximum_coefficient | number > 0, ≤ 20 | Your calibration | Optional |
| minimum_history_rows | integer ≥ 1, ≤ 1000000 | Your calibration | Optional |
| minimum_transitions_per_state | integer ≥ 0, ≤ 1000000 | Your calibration | Optional |
| random_seed | integer ≥ 0, ≤ 4294967295 | Your calibration | Optional |
| scenarios | array of objects (8 fields) | Evidence | Yes |
| simulation_count | integer ≥ 1000, ≤ 1000000 | Your calibration | Optional |
| tail_probability | number > 0, < 1 | Your calibration | Optional |
| transition_history | array of objects (8 fields) | Evidence | Yes |
Each current_exposures
record
| Field | Type | Required |
|---|---|---|
| ai_provider_group_id | string (non-empty) | Yes |
| capability_class | string (non-empty) | Yes |
| current_degraded_units | integer (≥ 0) | Yes |
| current_ready_units | integer (≥ 0) | Yes |
| current_unavailable_units | integer (≥ 0) | Yes |
| degraded_effectiveness_fraction | number (≥ 0, ≤ 1) | Yes |
| degraded_unavailable_prior_rate_months | number (> 0) | Yes |
| degraded_unavailable_prior_shape | number (> 0) | Yes |
| evidence_verified | boolean | Yes |
| fixed_response_cost | number (≥ 0) | Yes |
| forecast_ai_reliance_fraction | number (≥ 0, ≤ 1) | Yes |
| id | string (non-empty) | Yes |
| planned_unaided_practice_events_per_unit_month | number (≥ 0) | Yes |
| ready_degradation_prior_rate_months | number (> 0) | Yes |
| ready_degradation_prior_shape | number (> 0) | Yes |
| recovery_time_log_mean_hours | number | Yes |
| recovery_time_log_sd | number (> 0, ≤ 5) | Yes |
| required_fallback_units | integer (≥ 1) | Yes |
| value_at_risk_per_hour | number (≥ 0) | Yes |
{
"coefficient_grid_size": 21,
"current_exposures": [
{
"ai_provider_group_id": "agent-provider",
"capability_class": "deployment",
"current_degraded_units": 2,
"current_ready_units": 8,
"current_unavailable_units": 0,
"degraded_effectiveness_fraction": 0.5,
"degraded_unavailable_prior_rate_months": 20,
"degraded_unavailable_prior_shape": 1,
"evidence_verified": true,
"fixed_response_cost": 5000,
"forecast_ai_reliance_fraction": 0.7,
"id": "deployment-current",
"planned_unaided_practice_events_per_unit_month": 0.5,
"ready_degradation_prior_rate_months": 20,
"ready_degradation_prior_shape": 1,
"recovery_time_log_mean_hours": 2.0794415416798357,
"recovery_time_log_sd": 0.4,
"required_fallback_units": 6,
"value_at_risk_per_hour": 1000
}
],
"random_seed": 61,
"scenarios": [
{
"ai_reliance_multiplier": 1,
"id": "base",
"loss_multiplier": 1,
"outage_duration_log_mean_hours": 1.3862943611198906,
"outage_duration_log_sd": 0.3,
"probability": 0.8,
"provider_outage_probability": 0.05,
"unaided_practice_multiplier": 1
},
{
"ai_reliance_multiplier": 1.2,
"id": "stress",
"loss_multiplier": 2,
"outage_duration_log_mean_hours": 2.4849066497880004,
"outage_duration_log_sd": 0.5,
"probability": 0.2,Truncated for display — the full payload is 292 lines.
What you get back
This is the actual output of running the example above — computed by the same function the platform calls, not an illustration.
{
"assumptions": [
"Tenant-local aggregate capability transitions are right-censored exposure intervals; AI reliance raises and unaided practice lowers both adjacent-state hazards multiplicatively.",
"Ready, degraded and unavailable form a progressive continuous-time Markov chain over the forecast horizon; provider outage events and durations are shared within provider groups.",
"Fallback units, degraded effectiveness, recovery time and financial loss are defined and calibrated at capability-class grain rather than inferred from individual activity."
],
"capability_forecasts": [
{
"ai_provider_group_id": "agent-provider",
"capability_class": "deployment",
"degraded_unavailable_transitions": 4,
"expected_degraded_units": 1.491,
"expected_fallback_unit_shortfall": 4.5877,
"expected_loss": 1868.5103,
"expected_ready_units": 0.6735,
"exposure_id": "deployment-current",
"history_rows": 24,
"probability_fallback_shortfall": 0.9895,
"ready_degradation_transitions": 4,
"support_gate_pass": true
}
],
"counts": {
"capability_classes": 1,
"history_rows": 24,
"provider_groups": 1,
"scenarios": 2,
"simulations": 2000,
"supported_exposures": 1
},
"decision": "review_ai_capability_atrophy_tail_exposure",
"forecast": {
"conditional_value_at_risk": 32836.5127,
"expected_loss": 1868.5103,
"median_loss": 0,
"tail_probability": 0.05,
"value_at_risk": 14413.1178
},
"limitations": [
"Unmeasured workflow redesign, selective exercise assignment, changing state definitions or dependence among capability units can invalidate transition and shortfall estimates.",
"The model cannot infer any individual's skill, intent, effort, loyalty, nationality, employment suitability or hidden employment and must not drive person-level action."
],
"method": "bayesian_piecewise_exponential_ai_capability_atrophy_ctmc_loss_v1",
"posterior": {Truncated for display — the full payload is 59 lines.
How it works
Sequential Bayesian & bandits — Learn while deciding — update beliefs as evidence arrives and choose where the next unit of effort is worth spending.
- 1 Fit shared nonnegative AI-reliance and unaided-practice log-hazard coefficients on a governed grid while integrating class/state baseline hazards under Gamma priors and right censoring.
- 2 Draw posterior ready-to-degraded and degraded-to-unavailable rates, propagate the progressive continuous-time Markov chain, and convert degraded capacity to effective fallback units.
- 3 Reuse provider-group outage states and durations across dependent capability classes, price aggregate shortfall plus recovery and abstain when local transitions or evidence are insufficient.
Before you trust it
Every tool in the catalog ships with the conditions under which its answer is meaningful — and the conditions under which it should abstain instead of guessing.
Assumptions & guardrails
- The likelihood or reward model, prior support, action logging, delayed outcomes, and any stationarity assumptions match the deployment process.
- State definitions, censoring, AI-reliance/practice measurement and capability classes remain stable; progressive Markov hazards and represented provider scenarios transport to the forecast horizon.
- Posterior probability and adaptive selection are model-conditional; they are not substitutes for randomized propensities or guaranteed safety.
- This is aggregate recoverability economics under represented evidence—not a measure of an individual's skill, productivity, intent, loyalty or employability.
Minimum evidence
- transition_history: required and organization-defined
- current_exposures: required and organization-defined
- scenarios: required and organization-defined
How to validate it
Use chronological train/calibration/test windows, compare proper scores and decision value with a simple baseline, and recalibrate only from outcomes resolved after prediction time.
Calibrating it to your org
Same for everyone
The mathematical kernel, validation rules, method version, and JSON output semantics are organization-independent; no tenant-trained coefficients or company benchmark is embedded in the function.
Specific to you
- versioned aggregate capability-state panel joined to AI-workflow reliance, qualifying unaided practice/exercises, mature transitions and provider incident history using only facts available in each interval
- capability class/state/censoring definitions, reliance and practice measurement, prior transport, support floor, forecast horizon, provider groups/scenarios, degraded effectiveness, recovery/loss units, tail appetite and risk owner
Calibration workflow
- 1 Define the management decision, target outcome, aggregate unit, privacy boundary, cadence, and prediction/intervention horizon for this organization.
- 2 Build a tenant-scoped historical cohort using only information available before each prediction or decision; preserve zero periods, censoring, assignment probabilities, and unresolved outcomes when the method requires them.
- 3 Estimate statistical parameters on training history, but obtain costs, utilities, risk tolerance, practical-effect thresholds, capacity, and policy constraints from accountable decision owners.
- 4 Validate on later time windows or held-out aggregate units at the deployment grain, against a simple baseline and the function-specific validation strategy.
- 5 Deploy only if the returned decision clears evidence, overlap, calibration, robustness, and guardrail checks; warning, unsupported, schema-gap, and fallback decisions are abstentions.
- 6 Monitor realized outcomes, data drift, coverage, and decision regret; recalibrate at a governed cadence or after a detected regime/definition change, never merely because a stakeholder dislikes the result.
Call it from your AI
You don't wire up 388 tools in your MCP client. The GitRevio MCP server exposes 18 tools, three of which let an agent search the catalog, read a tool's schema, and run it — so the assistant finds this one on its own.
gitrevio_capabilities_search
{ "q": "learn how aggregate fallback capability decays" }
→ finds "forecast_ai_capability_atrophy_loss"
gitrevio_capability_describe
{ "capability_id": "forecast_ai_capability_atrophy_loss" }
→ returns the input schema and agent guidance shown on this page
gitrevio_capability_run
{ "capability_id": "forecast_ai_capability_atrophy_loss", "arguments": { ... } }
→ returns the result shown aboveWorks in Claude Desktop, Claude Code, Cursor, Cline, Continue.dev, Goose and Aider. See the MCP server.
Related tools
Audit AI capability fallback integrity
Prove that every aggregate capability required when AI is unavailable has a current approved runbook and a sufficiently large, timely, successful, independently observed exercise conducted with AI actually disabled.
Optimize AI capability resilience portfolio
Choose unaided drills, work rotations, cross-training, dual running, fallback redesign or monitoring per aggregate capability class using exact binomial shortfall, common-provider unique loss, hard readiness/control/capacity gates and a CVaR Pareto frontier.
Audit AI code change evidence integrity
Prove that aggregate AI-assisted coding evidence comes from prospectively registered, nonoverlapping treatment/control studies with immutable assignment, configuration, trace and mature-outcome denominators before anyone estimates an effect.
Audit AI inference cost allocation integrity
Reconcile provider AI invoices bottom-up to workload and route usage, price terms, cached requests, retries, fixed charges and credits without combining currencies or silently allocating unexplained spend.
Audit AI knowledge grounding integrity
Audit the complete AI knowledge supply chain from immutable source versions through indexed chunks and effective access policy to retrieved evidence, claim-level citations and honestly mature grounding outcomes, without treating unresolved answers as failures.
Audit AI model routing evidence integrity
Audit every live AI-model route against current version-matched local evaluation, uncontaminated temporal holdout, pricing freshness, residency, retention, reliability and genuinely independent provider fallback evidence, counting each workload's value at risk once.