Skip to content
Back to skills

S32flashtool Evidence Disciplined Technical Validation

ASecurity

Use for technical investigations where documentation, static analysis, inferred context, package contents, or example assets must not be confused with live verification.

  • 5 stars
  • 0 votes
  • 0 copies
  • 2 views
  • Added October 15, 2026
documentationgorailsdocumentation

Works with

  • mcp

Security analysis

A100/100

Scanned October 15, 2026

npx -y skills add nxp-agentic-ai/s32-ds-ai-support --skill s32flashtool-evidence-disciplined-technical-validation --agent claude-code

Installs into .claude/skills of the current project.

Are you the author of S32flashtool Evidence Disciplined Technical Validation?

Add the live security badge to your README. It updates with every re-scan.

Security grade badge for S32flashtool Evidence Disciplined Technical Validation
[![Security: A — Skills Directory](https://www.skillsdirectory.com/api/skills/nxp-agentic-ai-s32flashtool-evidence-disciplined-technical-valida/badge)](https://www.skillsdirectory.com/skills/nxp-agentic-ai-s32flashtool-evidence-disciplined-technical-valida)

More formats (shields.io, HTML) on the badges page. Keep it an A: scan every change in CI with Pro.

Download with Pro
SKILL.md
---
name: s32flashtool-evidence-disciplined-technical-validation
description: Use for technical investigations where documentation, static analysis, inferred context, package contents, or example assets must not be confused with live verification.
license: LA_OPT_Online Code Hosting NXP_Software_License
metadata:
  author: NXP
  version: "1.0.0"
  product: s32flashtool
  depends_on: '["s32flashtool-agent-rules-minimal", "s32flashtool-workflow-index"]'
  tags: '["s32flashtool", "validation", "methodology", "diagnostic"]'
---
# Evidence-Disciplined Technical Validation

A methodology layer for S32FlashTool answers. It forces every technical
claim to carry an evidence tag so that documentation, package inventory,
inference, and live hardware readings are never mixed into a single
"confirmed" statement.

## When to use

Use this skill when a response would otherwise blur the line between what is
documented, what is available in the installation, what is inferred, and what
is verified on the connected hardware. Apply it as a **methodology layer** on
top of any S32FlashTool workflow that produces technical claims - before
answering questions of the form "is it supported?", "is this the board?",
"will this work?", or "can I use this flash?".

Typical triggers:
- the user asks whether a processor, flash memory, interface, or board is
  **supported**, and only documentation or installation inventory has been
  consulted
- the user asks to **confirm** a board identity, MCU identity, or flash
  identity, and no live tool call has produced that evidence yet
- the answer is about to combine documentation, package contents, and
  hardware behavior into a single statement (e.g. "this board has S32N55
  and the flash is EMMC")
- a support-matrix question was answered by scanning `flash/`, `targets/`,
  or `examples/` rather than by reading `doc/supported_*_devices.txt`
- board-image analysis produced an "expected configuration" and the agent
  is tempted to report it as the actual configuration
- the user asks "will this work?" or "is this compatible?" before any
  live `fread`, `fprogram`, `fverify`, `fcrc`, or `mcuid` has succeeded
- a previous step returned only `inferred` evidence and the user now asks
  for a factual answer

Also apply this skill when reviewing another skill's output before
presenting it to the user, to ensure each claim carries an evidence tag
(`live`, `installed-doc`, `user-supplied`, `inferred`).

## When not to use

Do **not** use this skill as the primary skill for:
- identifying which S32 board the user has - use `s32flashtool-identify-user-board`
- answering "which processors / flashes / interfaces are supported on
  platform X" - use `s32flashtool-supported-devices-per-platform`
- performing any operation on the target (read, program, erase, verify,
  CRC, RCON, SRAM, MCU ID) - use the matching operation skill
- routing a user request to the correct workflow - use `s32flashtool-workflow-index`

This skill is **methodology-only**. It never invokes MCP tools. It sits
alongside the operation and identification skills to keep their outputs
epistemically honest, but it never replaces them.

## Shared references
- Apply shared rules from `s32flashtool-agent-rules-minimal/SKILL.md`.
- Use `s32flashtool-workflow-index/SKILL.md` to select the correct workflow.

## Quickstart

For every technical claim in a response, do this pass:

1. Assign the claim an evidence tag:
   `live` (from a real MCP tool call in this session),
   `installed-doc` (inspected file under `s32flashtool_folder`),
   `user-supplied` (stated by the user), or
   `inferred` (deduced from static context).
2. If the user asked for verified facts and only `inferred` evidence is
   available, refuse and describe what would satisfy verification.
3. Never merge tags in one sentence. Split the claim if needed.

## Core Principle

Always distinguish between:

1. **Documented** - stated in manuals, support files, examples, PDFs, release notes, or package metadata.
2. **Available in installation** - present as files, binaries, examples, templates, or package artifacts.
3. **Inferred** - reasoned from naming, context, patterns, prior steps, or partial evidence.
4. **Verified** - directly confirmed by live tool output, hardware interrogation, successful execution, or observed runtime results.

Do not present documented, available, or inferred conclusions as verified facts.

## Required behavior

| # | Rule | Canonical phrasing |
|---|---|---|
| 1 | Label the epistemic status of every meaningful technical claim (`Documented` / `Available in installation` / `Inferred` / `Verified` / `Unknown`). | *"The MCU identity is **verified** by live `read_mcuid` output."* |
| 2 | Never upgrade evidence: documentation is not a hardware fact; package contents are not a populated board; example availability is not compatibility; algorithm presence is not board support; image resemblance is not identification. | *"The algorithm is **available in the installation**; it does not by itself prove eMMC is populated on this board."* |
| 3 | Keep five concepts separate: **documented support**, **package availability**, **board/platform expectation**, **connected hardware identity**, **operational validation**. Never merge them in one sentence. | *"The support file documents S32N55; the connected board is not yet verified."* |
| 4 | For support-matrix questions (supported processors / flashes / interfaces / limitations), read `doc/supported_*_devices.txt` first. Use `flash/`, `targets/`, examples, release notes, or semantic search only for cross-checking. | *"Per `supported_s32n5_devices.txt`, S32N55 is documented as supported."* |
| 5 | Presence of a file in the install tree means only "available in the installed package". It does not imply documented support, board population, current-setup validity, or hardware verification. | *"`EMMC.bin` is available in the installation; the support file must confirm eMMC for the board."* |
| 6 | If a live tool exists that would materially raise certainty, name it before making a strong claim. | *"Documented support only; live confirmation would require `s32flashtool-read-mcuid`."* |
| 7 | Cite the evidence source by file or tool name when giving important conclusions. | *"Based on `supported_s32n5_devices.txt` and the `X-S32NZ55-EVB.pdf` example."* |
| 8 | Prefer precise wording over convenient wording. Split "it is supported" into what layer of evidence supports it. | *"**Documented as supported** by the installed package, but **not yet verified on the connected hardware**."* |
| 9 | When user intent is ambiguous ("Is it supported?", "Will this work?"), clarify or answer at every evidence layer explicitly. | *"Documentation-level: listed. Hardware-level: not yet."* |
| 10 | For hardware and flashing workflows, use the 5-level reporting model: Documented / Available / Expected / Verified / Validated. | See `## Default answer template`. |
| 11 | Board-image analysis is a vision tool for observed physical state only. Ignore any "expected/reference/template" configuration it returns; official setup comes from S32FlashTool PDFs and support files. | *"Image shows a USB connector; official boot-mode setup must come from the board PDF."* |

### Preferred vs. avoided wording

**Prefer:** *"documented as supported"*, *"available in the installation"*, *"expected to be"*, *"appears consistent with"*, *"suggests"*, *"not yet confirmed"*, *"would need live verification"*, *"cannot be concluded from documentation alone"*.

**Avoid:** *"is supported"* (when only docs consulted), *"confirmed"* (without direct evidence), *"the board is X"* (when only inferred), *"this flash is on the board"* (when only the algorithm exists in the package), *"this will work"* (without validation).

## Evidence-source cheat sheet

Use this to decide what a source can and cannot justify.

| Source | Justifies | Does NOT justify |
|---|---|---|
| `doc/supported_*_devices.txt`, board PDFs, examples, release notes, semantic-search results, image-analysis template expectations | *"documented as supported"*, *"expected to be available"*, *"example exists for this board"* | *"confirmed on this physical board"*, *"proven to work in the current setup"* |
| Files in `flash/`, `targets/`, `examples/`, package-contained PDFs and templates | *"available in the installation"*, *"present in the package"*, *"candidate artifact exists"* | *"available on the board"*, *"supported for this board"* (unless docs also say so) |
| Successful `s32flashtool-read-mcuid`, successful `fread`/`fprogram`/`fverify`/`fcrc`, explicit board-image match with confidence, port discovery, runtime identification output | *"verified"*, *"confirmed by tool output"*, *"operationally validated"* | anything beyond what the tool actually returned |

## Default answer template

When a user asks for confirmation, use this structure:

```
Here is the precise status:

- Documented:              ...
- Available in installation: ...
- Expected / inferred:     ...
- Verified:                ...
- Not yet verified:        ...

Conclusion: ...
```

## Guardrails

**Scope**
- Apply only to technical claims where documentation, static analysis, package inventory, or
  inferred context could be mistaken for live verification (board support, supported
  processors/flash, flashing workflows, board identification).
- Never merge documentation-based conclusions with live-hardware conclusions in the same statement.
  Recovery: split the claim into two labeled statements.

**Destructive actions**
- This skill is a methodology, not an operation. It never triggers writes. If verification demands
  live interaction, hand off to the matching operation skill and require explicit confirmation.

**Refuse-and-escalate**
- If a claim cannot be sourced to (a) live tool output, (b) inspected installed documentation, or
  (c) explicit user statement: refuse to assert it as verified. Recovery: label it "unverified",
  cite what would be needed to verify, and ask the user how to proceed.
- If installation folder is not provided, refuse "installed-example evidence" claims.

**Output contract**
- Every assertion carries an evidence tag: `live`, `installed-doc`, `user-supplied`, or
  `inferred`. Recovery: if you cannot tag a statement, drop it.

## Validation loop

1. Every technical assertion emitted during the workflow carries one of the
   evidence tags `live`, `installed-doc`, `user-supplied`, or `inferred`.
   Fail on any unlabeled assertion.
2. No `inferred` claim is presented as if it were `live` or `installed-doc`.
   Fail if the transcript merges categories.
3. When the user asks a support/verification question that requires live data
   and no live data exists, the response is a refusal plus a description of
   what would satisfy verification. Fail if the response fabricates a live-style
   answer from static sources.
4. No MCP tool that writes to a target was invoked. This skill is
   methodology-only. Fail on any destructive tool call.

## Out of scope

- Executing any MCP tool. This skill is a methodology, not an operation.
- Board identification. Use `s32flashtool-identify-user-board`.
- Support-matrix answers. Use `s32flashtool-supported-devices-per-platform`.
- Producing conclusions from `inferred` evidence when the user asked for
  verified facts. Refuse and describe what would satisfy verification.

Attribution

Is this your skill, or is something wrong with this listing? Request removal or report an issue. Author removals are honored within 72 hours.

Comments

Loading comments…

Newsletter

Get the weekly list of safe new skills, and the ones we pulled.