Skip to content

model-card.md

Latest commit

 

History

History
72 lines (56 loc) · 2.74 KB

model-card.md

File metadata and controls

72 lines (56 loc) · 2.74 KB
Raw
language
en
es
license apache-2.0
base_model operator-pinned Amazon Bedrock Claude model_id
pipeline_tag text-generation
library_name boto3
model-index
name results
Outcome Receipts optional narrative drafter
task metrics
type name
text-generation
Receipt-constrained narrative drafting
type value name
fail_closed_grounding_rate
1.0
Unbound numeric spans blocked

Model card — optional narrative drafter

Generated by receipts cards. Do not edit by hand.

Intended use

When explicitly enabled, Claude on Amazon Bedrock rewrites a filled funder-report template for clarity. It receives the baseline narrative and an allowlist of receipted display strings. It does not compute metrics, map fields, approve a report, or verify a number.

Policy and enforcement

Cloud drafting is off by default. It requires provider = "bedrock", enabled = true, a pinned model_id, and the per-run --allow-cloud-drafting flag. The deterministic drafter remains the default. Every model draft is grounded against raw receipts, suppression rebuilds the publishable surfaces, and the publishable draft is grounded again before human approval. Any added, altered, rounded, signed, ranged, or written-out number blocks export.

Limitations and evaluation

The model may change meaning without adding a number, produce poor prose, or fail to follow the allowlist. Grounding proves numeric provenance, not narrative quality or fairness. Tests inject invented numbers and require the gate to reject them. A model-based faithfulness judge is not shipped; if added, it must be calibrated against human labels before gating decisions.

The committed synthetic benchmark is split across English and Spanish numeric spans. Both language slices enforce the same absolute requirement: no unbound numeric span survives. This evaluates the mechanical grounding gate, not prose quality, cultural fit, or translation quality.

Out-of-scope use

The drafter must not compute, round, map, approve, or verify a figure. It is not a case-management system, eligibility system, clinical tool, or source-data quality checker. It must not receive client rows or direct identifiers.

Environmental footprint

N/A — this repository trains and fine-tunes no model. Bedrock inference energy and emissions are controlled by the selected provider model and are not exposed to this offline tool as a measurable training run.

Human responsibility

A named human approves the final redacted artifact. Approval is not evidence that the underlying metric definition or source data is correct.

Last verified: 2026-07-12 · Recheck cadence: on any model, prompt, or Bedrock policy change, and at least quarterly.