---
title: "Model Replication Diff"
method: POST
path: "/v1/tools/replicate_model_outputs"
tags: ["TODO_TAGS_REVIEW", "TODO_AP2_EXPORT_REVIEW"]
---

# Model Replication Diff

`POST /v1/tools/replicate_model_outputs`

## Request

- Base URL: `https://ainumbers.co`
- URL: `https://ainumbers.co/v1/tools/replicate_model_outputs`
- Auth: none declared

Independently recomputes a model's reported outputs from a declared model specification (version, as-of date, transform, intercept, coefficients) and a stated input record set, then diffs the recomputation against the model's own reported outputs within a caller-supplied tolerance. Returns a per-record diff (reported, recomputed, absolute and relative diff, within-tolerance), aggregate tolerance stats, an overall replicated / not-replicated verdict, and the failing segment(s) named when not replicated. If the declared specification cannot be recomputed as stated -- unrecognized transform, no coefficients, no tolerance -- the honest result is not_replicable_as_specified with a reason, not a forced pass or fail. Deterministic recompute-and-diff only: this node never opines on conceptual soundness, assumption reasonableness, or fitness for use -- that judgment belongs in a validator's approval record, not here. Model inputs and specifications are expected to arrive as a workbook export via the shipped WB-BRIDGE-1 workbook-to-OCG artifact bridge (tool 554); this node itself accepts plain JSON records so any upstream source can feed it. Distinct from the shipped model-passport nodes (art-450/451/453), which track a model's inventory tier and revalidation cadence rather than recomputing its numeric output. Inline deterministic transcendental math (no engine Math.exp/log) so the same input always produces the same execution_hash on every surface. NaN-safe. Zero network, zero PII.

**Tool URL:** https://ainumbers.co/chaingraph/art-488-model-replication-diff.html

## Request body

- object
  - `model_spec` unknown, required
  - `records` unknown[], required
    - unknown
  - `tolerance` object, required: type not evidenced by kernel source
    - `rel_tolerance` unknown

## Response `200`

Successful tool execution result

- object
  - `aggregate` object
    - `count_computed` integer
    - `count_out_of_tolerance` integer
    - `count_total` integer
    - `max_abs_diff` number
    - `mean_abs_diff` number
  - `failing_segments` unknown[]
    - unknown
  - `model_spec_as_of_date` string
  - `model_spec_version` string
  - `per_record` object[]
    - `abs_diff` integer
    - `id` string
    - `recomputed_value` integer
    - `rel_diff` integer
    - `reported_value` integer
    - `segment` string
    - `within_tolerance` boolean
  - `reason` string
  - `tolerance_applied` object
    - `abs_tolerance` number
    - `rel_tolerance` string
  - `transform` string
  - `verdict` string

---

[API](https://skmtc.dev/ainumbers/apis/ainumbers-co-fintech-intelligence-suite.md) · [All operations](https://skmtc.dev/ainumbers/apis/ainumbers-co-fintech-intelligence-suite/llms.txt) · [OpenAPI document](https://skmtc.dev/ainumbers/apis/ainumbers-co-fintech-intelligence-suite/revisions/b462e4f5bb28?raw)