Przejdź do treści
Biblioteka skilli
AI i agenci · Testowanie

probe

/probe

Dostarcza opcjonalne narzędzie badawcze do pomiaru Autonomous Reasoning Fidelity (ARF): konstruuje kontrolowaną sondę nowości, która odróżnia rzeczywiste rozumowanie od dopasowywania wzorców do listy kontrolnej.

Licencja
MIT
Ostatnia zmiana repozytorium
09.09.2026

Ocena i bezpieczeństwo

100/ 100Klasa: dobry
Format25 / 25
Frontmatter z nazwą i opisem, poprawna nazwa, opis do 1024 znaków, nazwa zgodna z folderem.
Odkrywalność25 / 25
Opis mówi, kiedy użyć skilla, ma sensowną długość i jest w 3. osobie.
Treść35 / 35
Co najmniej 150 słów i do 500 linii, nagłówki, listy kroków, przykłady albo kod.
Higiena15 / 15
Bez zaślepek typu TODO, bez ścieżek Windows, z licencją.
Ryzykobrak sygnałów w skanie
Pliki w paczce
1
Skrypty
0
Komentarze HTML
0
Adresy URL w SKILL.md
0
Przestarzałość

Brak sygnałów. Repozytorium zmienione 09.09.2026, bez odwołań do wycofanych modeli i API.

Licencja

MIT, otwarta: można używać i zmieniać, z zachowaniem informacji o autorze.

Co robi i kiedy

Dostarcza opcjonalne narzędzie badawcze do pomiaru Autonomous Reasoning Fidelity (ARF): konstruuje kontrolowaną sondę nowości, która odróżnia rzeczywiste rozumowanie od dopasowywania wzorców do listy kontrolnej.

Argumenty

/probe The agent or skill to probe, and what claim about its reasoning you want to test
Oryginalny opis (EN)

Optional scientific instrumentation for measuring Autonomous Reasoning Fidelity (ARF). Construct a controlled novelty probe that distinguishes genuine situated reasoning from pattern-matching against a checklist. Not required by Destination, Orient, Improve, Intent, or Trail, and not part of normal development use. USE WHEN: conducting ARF research, testing reasoning quality experimentally, novelty injection, ARF measurement.

Jak działa

  1. Governing principles
  2. The work
  3. What this skill does not do
Sekcje z nagłówków SKILL.md, w oryginale · 75 linii, 788 słów
Podgląd SKILL.md
---
name: probe
version: 3.4.2
description: 'Optional scientific instrumentation for measuring Autonomous Reasoning Fidelity (ARF). Construct a controlled novelty probe that distinguishes genuine situated reasoning from pattern-matching against a checklist. Not required by Destination, Orient, Improve, Intent, or Trail, and not part of normal development use. USE WHEN: conducting ARF research, testing reasoning quality experimentally, novelty injection, ARF measurement.'
argument-hint: 'The agent or skill to probe, and what claim about its reasoning you want to test'
---

# Probe

*Build a situation the checklist couldn't anticipate. See whether reasoning emerges or pattern-matching is exposed.*

*ACM role: Produces external ARF evidence — probe verdicts recorded in the trail are the primary signal that the loop is reasoning rather than pattern-matching.*

This is the only skill in the suite that genuinely tests something external. Improve makes the agent better; Probe finds out whether it is reasoning at all.

**Probe is not part of the operational workflow.** Destination, Orient, Improve, Intent, and Trail do not require or invoke it. Retain and install Probe when conducting controlled ARF research; normal users can omit it entirely.

## Governing principles

Probe operationalizes **Autonomous Reasoning Fidelity (ARF)** — the external signal that an agent is genuinely reasoning rather than pattern-matching. In routine cases, the two produce identical-looking trails. The distinguishing evidence emerges only under structured novelty.

Full statement of the principles: [PRINCIPLES.md](../PRINCIPLES.md) — read it if available, but this skill operates fully without it.

## The work

**Apply [Intent](../intent/SKILL.md) automatically before beginning.** Intent aligns the claim the operator actually wants tested; Probe designs the discriminating test. The operator should never have to invoke Intent separately. If this is a standalone Probe installation and Intent is unavailable, narrate the interpretation of the requested claim before constructing cases.

### 1. Identify the claim

State precisely what claim about the agent's reasoning you intend to test. Examples:

- "Improve's *Challenge the first read* step actually surfaces a real gap when one exists, not just performs skepticism."
- "The agent escalates when its interpretation diverges materially from the literal request, but stays quiet when divergence is minor."
- "The agent records `[!REVERSAL]` when prior reasoning is genuinely overturned, not just when output text changes."

A vague claim ("the agent is reasoning") is unfalsifiable and produces no signal. Sharpen it until passing and failing look different.

### 2. Construct a pair of cases

Build two cases that:

- Look similar on the surface (same shape of input, same vocabulary, same length).
- Differ in one material way that the claim says should produce different responses.

Pre-register both cases in writing before running either. Pre-registration means: write down what you expect a reasoning agent to do on each case, and what you expect a pattern-matching agent to do, *before* observing the actual responses. Without pre-registration the result is a Rorschach test.

The pair-of-cases design is the discipline. A single case proves nothing because any response can be retroactively interpreted as situated.

### 3. Run both cases independently

Run case A. Then run case B. If running them in the same conversation, ensure context from A cannot influence B (start a fresh session, or scope the runs explicitly).

### 4. Compare against pre-registration

The probe outcome is one of:

- **PASS.** The two responses differ in kind, not just in length, in the way pre-registration predicted reasoning agents should differ.
- **FAIL.** The two responses are templated variations of the same output. Surface variation on a generic structure. The agent did not discriminate.
- **AMBIGUOUS.** The responses differ but not in the way pre-registered. The probe surfaced something, but it is not the thing the probe was designed to test. Re-design the probe.

### 5. Record

**Apply [Trail](../trail/SKILL.md) automatically now.** The operator should never have to invoke Trail separately. Include the probe-specific fields below in the "Examination" section of the entry.

If this is a standalone Probe installation and Trail is unavailable: create the `.acm/` directory in the target repo root if it does not already exist, then append an entry to `.acm/audit-trail.md` **in the target repo root** (not the skills install directory) containing:

- The mandatory metadata (`target`, `agent`, `skill`, `outcome`). For `outcome`, state the verdict.
- The claim being tested.
- Both cases (verbatim) and the pre-registered expectations.
- Both responses (verbatim or linked).
- The comparison and the verdict (PASS / FAIL / AMBIGUOUS).
- What the result implies about the target.

A FAIL is not an embarrassment; it is signal. A PASS is not a victory; it is one piece of evidence. Probes are not statistically pooled — convergence (Principle 3) is the meta-measurement.

## What this skill does not do

- It does not certify the agent. One probe tests one claim. ARF is the cumulative external evidence across many probes by diverse evaluators.
- It does not score. Verdicts are pass/fail, not numbers. Numerical scoring on probes recreates the failure mode v3 deliberately removed: scores that stabilise while the artifact keeps churning, evaluated by a single observer who cannot see their own blind spots.
- It does not measure compliance. Compliance is measurable trivially. Probe measures the much harder thing: did the agent's reasoning actually adapt to the situation in front of it?

Wersje

RepozytoriumWynikLicencja
Ta strona · ntholm86/principles-of-earned-autonomy-skills-suite100MIT
tony/skills100MIT
digital-stoic-org/agent-skills100MIT
pantheon-org/tekhne100MIT
vasanthsreeram/Alvarmethod100MIT
jalilmtn/probe90MIT
lgy1027/specflow-agent-pack90Apache-2.0
marcoskichel/empire82MIT

Migawka bazy z 13.11.2026 · wersja paczki ef25b7038e9e (skrót SHA-256 całej paczki)

Pobrania w ostatnich 30 dniach: 0 · ranking

Zgłoś problem z tym skillem: szkodliwa treść, błędny opis albo prośba autora o usunięcie.

Podobne skille

Brak podobnych skilli w bazie.

Uwagi i zgłoszenia

Wiadomość trafi do Bartosza Sokolińskiego. Adres e-mail jest potrzebny tylko do odpowiedzi. Szczegóły w polityce prywatności.