Abstract

An AI system scores a risk at 0.503. The rule says to act only above 0.500. The model is approved, the quantized runtime is approved, and the agent is authenticated. But the low-precision path the result actually took carries an error of up to 0.006, so the true value could be anywhere from 0.497 to 0.509. An approved model, an approved runtime, and a signed result do not prove that this particular number is reliable enough for this particular action. Who decides whether the arithmetic is good enough to act on?

Based on Indian Provisional Filing: 202631136555, filed 15 November 2026, and disclosed with the Mothership (PCT/IB2026/055615) on 16 July 2026.

This patent-pending technical concept proposes a complementary mechanism for mixed-precision, quantized, and heterogeneous AI systems. Its core principle separates three questions: whether the computation was approved, whether the numerical result is reliable enough for the specific consequence, and whether the action is authorized. Only the second is new here, and it is checked before the action takes effect.

The mechanism works in ten steps:

  1. A proposed action derived from a numerical result is held as non-effective.
  2. The system reconstructs the action that would actually take place.
  3. It identifies the effective precision path the result took: formats, quantization scales, accumulation precision, rounding, fused operations, kernels, and runtime version.
  4. A protected numerical acceptance contract states the input domain, the permitted error, the required evidence class, the decision margin, and the validity epoch.
  5. A validator collects numerical error evidence and classifies its strength as deterministic bound, reference-verified, empirical, or unresolved. An empirical accuracy figure is never relabeled as a worst-case guarantee.
  6. Errors across stages are composed conservatively, with correlation and sensitivity taken into account.
  7. A decision-stability check tests whether the whole error interval stays on the permitted side of the threshold.
  8. Authority, approved decision logic, purpose, and target scope are checked independently.
  9. If the interval is ambiguous, an approved higher-precision or reference recomputation is allowed, and the fallback result must itself be validated.
  10. A result-bound authorization is issued, and the final enforcement point admits only the covered action.

In the example, 0.503 with an error of 0.006 gives the interval [0.497, 0.509], which crosses 0.500, so the action is held. The higher-precision fallback returns 0.497, so the action is denied. A score of 0.540 with the same error has a lower bound of 0.534, so the numerical check passes without any recomputation. Other authorization checks still apply. Cheap low-precision execution is kept for ordinary work, and expensive recomputation happens only for unresolved, consequential decisions. A change to the quantization, kernel, or runtime after validation invalidates the earlier evidence.

Prospective industry relevance (AI-generated; see notice below). The mapping shows where a numerical-consequence check could sit alongside documented products. Each placement is an engineering hypothesis, and none of these vendors is represented as incorporating this concept.

  • OpenAI (Responses API, function calling, Agents SDK): the check sits in a customer-controlled tool dispatcher before a consequential call. It receives the numerical result, the precision-path record, the action parameters, and the epoch. The destination independently authorizes the effect, and the call is rejected if the interval crosses the threshold, because a model output is not a numerical proof.
  • Anthropic (Claude tool use, MCP, Agent SDK): the check sits in a customer-controlled MCP proxy or tool runner that stages the tool-use request. It verifies the numerical contract and allows only a bounded target capability, rejecting a request whose computed threshold is unresolved. A generic tool-success report is not evidence that the effect was committed.
  • Google (Gemini function calling, ADK, Vertex AI, Cloud TPU): the check sits in the application’s function handler and customer-operated configuration, database, or control APIs. It tests certified score bounds and destination scope before a real network or industrial resource is changed.
  • Microsoft (Foundry Agent Service, function tools, Azure AI): the check sits in the customer-controlled function executor. It evaluates the numerical predicate and target scope before the downstream operation. A vendor-hosted prediction is not a trusted numerical certificate.
  • NVIDIA (TensorRT, Model Optimizer): documented reduced-precision and calibration data can feed the precision-path record. A protected service ties the result margin to an action-specific output or commit gate.
  • Intel (OpenVINO, NNCF), AMD (ROCm), and Huawei (Ascend, CANN, MindSpore Lite): their quantization and runtime data can supply the precision-path record. A reference monitor checks the interval before a high-risk tool call or resource change. Operator observability and a trusted enforcement point depend on the deployed stack.
  • AWS SageMaker, Qualcomm Snapdragon AI, MediaTek NeuroPilot, and Apple Core ML: pipeline or on-device evidence feeds a separately managed contract, enforced at a constrained gateway at the actual target. Only supported API and firmware interfaces can claim enforcement.
  • Telecom and industrial operators: the check sits at the network-configuration or physical-command boundary, only for numerically sensitive planned changes. Mandatory real-time and safety loops are left alone.

These are possible complementary integration points, not claims of deficiency in any named product. Mixed-precision arithmetic, quantization, calibration, interval arithmetic, selective recomputation, runtime monitoring, and output gating are established and not claimed as individually new. The proposed contribution is the specific combination: verifying the effective precision path, the consequence-specific error and decision-stability check, and result-bound enforcement at the actual effect boundary. The mechanism does not guarantee factual, semantic, or physical correctness. A signature protects evidence from alteration, but it does not prove the math. The error evidence must come from a trusted source, and a software gateway governs only the actions routed through it. No latency benefit is asserted without measurement.

Where to read the full filing. Search “Sangam Das” on WIPO PATENTSCOPE, or search PCT/IB2026/055615 (pctib202655615). The full PCT file is also available in Technical Disclosure Commons: https://www.tdcommons.org/dpubs_series/12542/

Basis and notice. The underlying invention is human-originated and based on the Mothership application PCT/IB2026/055615, filed 4 June 2026 and publicly disclosed 16 July 2026. This abstract was prepared with AI-assisted tools, including the industry-relevance section, and may contain errors, omissions, or prospective interpretations. Readers must cross-check it against the original PCT filings and the companies’ own public documentation before drawing any conclusion. Company and product references are for explanatory, complementary architectural alignment only. They are not an endorsement, validation, or infringement statement, and imply no adoption, collaboration, licensing, or finding on patent validity or third-party rights. The original specifications and claims are authoritative.

Part 8 of 39. This is Part 8 of a 39-part technical disclosure series: 38 numbered embodiments plus one separately designated Embodiment 12A. Each part presents one embodiment, its problem, the proposed solution, practical implementation mechanisms, and prospective complementary industry relevance.

License: CC BY-NC 4.0.

KEYWORDS

AI Safety; Numerical Reliability; Mixed-Precision Inference; Quantization; Quantization Robustness; Floating-Point Error Analysis; Interval Arithmetic; Error Propagation; Decision Stability; Threshold Decisions; Numerical Uncertainty; High-Precision Fallback; Model Runtime Drift; Compiler and Kernel Provenance; Trusted Inference; Runtime Authorization; Consequence-Bound Authorization; AI Tool-Call Security; Agentic AI Security; Model Context Protocol; Function Calling; Critical Infrastructure; Industrial Control; Safety-Critical AI; AI Accelerators; Edge AI

Creative Commons License

Creative Commons License
This work is licensed under a Creative Commons Attribution-Noncommercial 4.0 License

Share

COinS