跳至正文
OpenAgentSkill
浏览已开放 · 目录记录日期 2026-11-21 · 自动更新和账户服务暂时关闭。
浏览器与自动化

ach

Analysis of Competing Hypotheses — structured technique for evaluating multiple explanations against evidence. Use when facing ambiguous attribution or multiple plausible scenarios. Prefers graded evidence items from /quality-of-information-check and runs that skill first when handed raw URLs or text. Runs on ungraded evidence when asked, labelled as such, unweighted, and capped at Moderate confidence.

Liberty91LTD · 已有静态检查记录

安装与使用

运行命令前先阅读来源与依赖说明。复制命令不会自动安装或执行。

npx skills add Liberty91LTD/cti-skills --skill ach

技能说明


name: ach description: Analysis of Competing Hypotheses — structured technique for evaluating multiple explanations against evidence. Use when facing ambiguous attribution or multiple plausible scenarios. Prefers graded evidence items from /quality-of-information-check and runs that skill first when handed raw URLs or text. Runs on ungraded evidence when asked, labelled as such, unweighted, and capped at Moderate confidence. user-invocable: true metadata: version: 2.0.0

Analysis of Competing Hypotheses (ACH)

ACH is the most important structured analytic technique for CTI. It forces you to evaluate ALL plausible hypotheses against ALL significant evidence, reducing the impact of cognitive biases — especially confirmation bias.

When to Use ACH

  • Attribution questions: "Who is behind this campaign?"
  • Ambiguous situations: Multiple plausible explanations exist
  • High-stakes assessments: Getting it wrong has significant consequences
  • Contested analysis: Analysts disagree on the conclusion

Evidence: Graded Preferred, Ungraded Allowed

ACH works best on graded evidence items, one per claim, as produced by /quality-of-information-check (the QoI JSON or its claim table). The schema is in skills/quality-of-information-check/references/evidence-item-schema.md. Grading is the default. It is not a condition for running.

An item counts as graded when it has all of: claim_id, claim, claim_type, anchor, grading.access_level, grading.claim_support, corroboration and provenance_basis.

What you were handedWhat to do
QoI JSON or graded claim tableProceed. If the JSON is a file, confirm it with python3 skills/quality-of-information-check/scripts/validate_evidence.py <qoi.json> (exit 0 = valid). Items that fail are not graded evidence: rerun the check on them or carry them as ungraded.
Raw URLs, articles, pasted report textComplete Step 1, then invoke /quality-of-information-check on the material, passing the question. Use its output as the evidence. This is the default and needs no permission.
Evidence with a bare source rating but no claim table (for example "Mandiant report, Admiralty B2")Run /quality-of-information-check on the underlying document if it is available. If not, carry the item as ungraded and record the rating as the analyst's own.
Evidence the user asserts from memory, with no document behind itCannot be graded. Carry it as ungraded, marked "asserted, no document".
A mix of graded and ungradedGrade what can be graded. Carry the rest as ungraded. Graded rows keep their grades and weights.
The user asks to skip grading, or /quality-of-information-check cannot be runRun in ungraded mode, below. Say once what that costs, then proceed. Do not ask again.

Ungraded mode

Any run with at least one ungraded row in the matrix follows these rules:

  • Label it. The report header carries Evidence basis: ungraded when no row is graded and Evidence basis: mixed when some are. A fully graded run carries Evidence basis: graded.
  • Do not invent grades. Do not grade items yourself inside this skill. An ungraded row shows ungraded in the Access level and Claim support columns. Where the analyst supplied a rating through /source-assessment, show it as Admiralty B2 (analyst). It is displayed and does not change the weight. An Admiralty rating applies to a lookup result or a single item, the evidence grade applies to a claim from a document, and neither is converted into the other.
  • Ungraded rows weigh 1. One ungraded item cannot eliminate a hypothesis alone. A hypothesis is rejected by ungraded evidence only through accumulation.
  • Confidence is capped at Moderate when any load-bearing item is ungraded. High is reserved for conclusions whose load-bearing items are all graded and independently confirmed.
  • Give each row an id. Ungraded rows are numbered U01, U02 and so on, with the claim in one sentence and its source named.
  • Say it in Caveats. State that the evidence was not graded, which items that applies to, and that running /quality-of-information-check on them may change the ranking.

First-party observations from the organisation's own telemetry are graded direct, established by rule R12 and need no check. They are graded rows in any mode.

Step-by-Step Procedure

The order is strict: hypotheses first, then evidence. Step 1 is completed and written down before any evidence item is read, graded or fetched.

Step 1: Generate Hypotheses (before reading any evidence)

List ALL reasonable hypotheses. Include unlikely ones — the point is to avoid premature narrowing.

Rules:

  • Minimum 3 hypotheses (if you only have 2, you're doing binary thinking)
  • Include at least one that challenges your initial instinct
  • Hypotheses should be mutually exclusive where possible
  • Include "unknown actor" or "coincidence" as hypotheses when appropriate

Ordering rules:

  • Generate hypotheses from the question alone. Do not open the evidence, run lookups or invoke /quality-of-information-check until the list is written out.
  • If the evidence is already in the conversation, generate the hypotheses from the question as worded and say in the output that the evidence was visible beforehand.
  • Once evidence has been read, hypotheses may be added but never removed or reworded. Mark any late addition "added after evidence" in the report. A hypothesis is only rejected through the matrix.

Step 2: Load the Evidence

Apply the evidence table above. Then list every evidence item relevant to the hypotheses.

For each graded item, carry over from the QoI output without changing it:

  • claim_id, claim and claim_type
  • The grade (access_level and claim_support, e.g. limited, firm)
  • load_bearing, flags, and corroboration.independent_primaries with its basis
  • provenance_basis

For each ungraded item, record an id (U01), the claim in one sentence, the sourc

阅读完整来源 ↗