beyhan/reverse-skill
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

feat(ops): analysis decision framework R1-R51 (#77)

Browse Source 
Add analysis-decision-framework cookbook (P0 full / P1 short / P2 index).
R4* validated sufficiency (compatible >=1 bind). Thin hooks in evidence-
finding-path, re-agent-workflow Synthesis, RULES(+zh), verify anchors.
Not a second master workflow; keep #65 gates and A-T/U-AV superior.
This commit is contained in:
yhc
2026-08-14 20:26:03 +08:00
parent 2c95829db5
commit ec61a6f951
6 changed files with 188 additions and 1 deletions
Download Patch File Download Diff File Expand all files Collapse all files
RULES.md
+2
View File
@@ -20,6 +20,8 @@ After reading this file, you MUST NOW (immediately) execute:
4. NEXT: Initialize case scope (`skills/scripts/case-init.ps1` / `skills/ops/scope-contract.md`). MUST NOT ACT against targets until auth.status=granted and network_profile set
5. ACT: Open PRIMARY skill SKILL.md; use roles (`ops/role-map.md`), timeline/workitems, Evidence→Finding→Path (`ops/evidence-finding-path.md`). Identity: `ops/IDENTITY.md` (not a Z3r0 platform clone)
> **Decision quality (Issue #77):** Follow skills/ops/analysis-decision-framework.md for hypothesis exits, validated sufficiency (R4*), grounded conclusions, and deadlock replan. Do **not** paste the full R1-R51 list into this file.
IMPORTANT — Shared installation:
- tool-index.md is the SINGLE SOURCE OF TRUTH for tool availability
- If another CLI already installed tools (tool-index shows "yes"), DO NOT reinstall
RULES_zh.md
+2
View File
@@ -144,6 +144,8 @@
## 执行原则
> **决策质量（Issue #77）：** 假设退出、validated 充分性（R4*）、结论锚定与死锁重规划见 skills/ops/analysis-decision-framework.md。**不要**把 R1-R51 全文塞进本文件。
### 工具使用
- **永远不要猜工具路径**，先读 `tool-index.md`
- 缺少工具时先调用平台对应的 bootstrap 脚本自动补齐，不要直接报错：
skills/ops/analysis-decision-framework.md
+146
View File
@@ -0,0 +1,146 @@
# Analysis Decision Framework (Issue #77)
> **SSoT role**: Decision-quality / evidence-sufficiency / agent-bias cookbook for reverse-skill agents.
> **Not** a second master analysis workflow. Obey `re-agent-workflow.md`, feasibility gate (#73), IAT iron rule (#72), A-T / U-AV cookbooks, and `evidence-finding-path.md` first.
> Rule IDs **R1-R51** keep the reporter numbering (**no R15**; includes **R50/R51**). Do not renumber.
## 0. How to use
| When | Action |
|------|--------|
| Finding promotion / Synthesis | **P0** (R4*, R1, R2, R41) |
| Stage change or stuck loop | R2, R31, R43 |
| Multi-module / anti-analysis | R50, R51 -> anti-analysis + A-T |
| Already covered by ops/CI/skills | **P2** pointer only |
**Evidence IDs** (record even on failure):
`E-confidence-low` · `E-hypothesis-confirmed` · `E-hypothesis-rejected` · `E-insufficient-evidence` · `E-negative-evidence` · `E-scope-boundary` · `E-runtime-only` · `E-bias-detected` · `E-over-trust-bias` · `E-module-decoupled` · `E-anti-adversarial` · `ungrounded`
---
## 1. P0 — Full recipes (trigger -> action -> Evidence)
### R1 — Decompile / static confidence band
| | |
|--|--|
| **Trigger** | Static decompile/CFG/types as primary reasoning |
| **Action** | Band `high`/`medium`/`low`. If `low`: MUST schedule dynamic before `validated` |
| **Evidence** | `E-confidence-low` |
### R2 — Hypothesis-driven stage exit
| | |
|--|--|
| **Trigger** | End of triage/static/dynamic or long tool loops |
| **Action** | State hypothesis; **continue / switch / stop** |
| **Evidence** | `E-hypothesis-confirmed` or `E-hypothesis-rejected` |
### R3 — Least surprise
Unverified unusual claims MUST be tagged `speculative`.
### R4* — Finding sufficiency (compatible rewrite)
Does **not** replace "Finding binds >=1 Evidence". Tightens **validated** only:
| status | Evidence bar |
|--------|----------------|
| preliminary / candidate | >=1 (unchanged) |
| **validated** | **SHOULD >=2 independent** Evidence (best 1 static + 1 dynamic). Single Evidence MUST NOT silently promote — residual + human, or stay candidate |
| blocked promotion | `E-insufficient-evidence` |
### R6 — Negative evidence
Checked-absent branch -> `E-negative-evidence`.
### R7 — Analysis boundary
Declare scope limits -> `E-scope-boundary` (align scope-contract).
### R8 — Suspicious != malicious
Default `flavor=null` unless `explicit_malware` / user asks (#71).
### R30 — Runtime back-annotation
Dynamic without static anchor: try relocate; else `E-runtime-only`, Finding <= candidate.
### R31 — Stage bias self-check
Tool/stage fixation -> `E-bias-detected`.
### R41 — Grounded conclusions
Claims MUST map Finding -> Evidence; else `ungrounded`.
### R43 — Plan deadlock -> replan
3 actions with no new Evidence, or 2 stage switches without Evidence -> replan under **feasibility gate**. Aligns RULES Self-Supervision.
### R44 — Single-source high confidence
Cross-check before validated; else `E-over-trust-bias`.
### R50 — Multi-module decoupling
Separate work items -> `E-module-decoupled`.
### R51 — Adversarial effort
Effort band + A-T pointers -> `E-anti-adversarial` (**no** A-T table copy).
---
## 2. P1 — Short rows
| ID | Landing |
|----|---------|
| R5 | content_hash; evidence-finding-path + review_case --verify-hashes |
| R12 | parallel hypotheses via R2 (no heavy case-branch product) |
| R22 | docs-generator executive summary MUST |
| R23 | IOC dual-channel **only** explicit_malware / user IOC; forbidden default ordinary RE |
| R28 | authorized lab; no full weaponized exploit chain in-repo |
| R34 | SHOULD journal stale note; **no** 90-day auto engine |
| R36 | archive feedback one-liner |
| R38-R40 | pointer -> llm-security only |
| R42 | YARA/detections experimental until benign validation |
| R45-R49 | route cloud-k8s / firmware / pentest-pwn / code-audit; limit confidence if missing context |
### Downgrades (not Agent runtime MUST)
| ID | Why |
|----|-----|
| R16/R19 auto-close PR | CI / human maintainer |
| R37 MCP gateway product | no in-repo gateway; tool-index + human confirm |
| R23 unconditional block | conflicts flavor=null |
| R34 automated stale | no metrics infra |
---
## 3. P2 — Covered elsewhere (index only)
| IDs | SSoT |
|-----|------|
| R9-R11 | timeline-workitem, case-init, case-review |
| R13-R14 + IAT | re-agent-workflow (#67/#72/#73) |
| R16-R19 (sans auto-close) | smoke, verify, CI, case-review |
| R20-R21 | evidence-finding-path, append-evidence |
| R24 | field-journal/anonymization |
| R25-R26 | skill-supply-chain, bootstrap pin (#76) |
| R27/R29 | scope-contract, RULES security |
| R32-R33 | MASTER-ROUTING, role-map |
| R35 | field-journal |
| A-T PE anti-analysis | anti-analysis.md |
| U-AV non-PE | nonpe-format-cookbook.md |
---
## 4. Synthesis checklist
1. R41 grounded claims
2. R4* validated bar
3. R1 low confidence -> dynamic
4. R43 deadlock -> replan under feasibility gate
5. R8/R23 no default malice/IOC
skills/ops/evidence-finding-path.md
+15
View File
@@ -114,3 +114,18 @@ The review is read-only and checks scope fields, Evidence records, work item and
| PG 不可变行 + API | Markdown 文件 + hash 字段 |
| UI 审阅队列 | 报告 + next-step 菜单 + journal |
| ATT&CK 深度绑定 | 可选标签，不强制 UI |
## Validated sufficiency (Issue #77 / R4*)
Global bind rule remains: every Finding references **>=1** Evidence.
Promotion to status=validated is stricter (decision cookbook):
| status | Evidence bar |
|--------|----------------|
| preliminary / candidate | >=1 (unchanged) |
| **validated** | **SHOULD >=2 independent** Evidence (best: 1 static + 1 dynamic). A single Evidence item alone MUST NOT silently promote to validated — keep candidate/preliminary, or record residual_risk + human confirm. |
| blocked promotion | record Evidence E-insufficient-evidence |
Full recipes: [nalysis-decision-framework.md](analysis-decision-framework.md) (R4*, R1, R41, R44).
skills/reverse-engineering/references/re-agent-workflow.md
+5
View File
@@ -193,6 +193,11 @@
## 4. Synthesis（IOC / 攻击链 / 报告）
### Decision quality overlay (Issue #77)
Before closing Synthesis, apply [nalysis-decision-framework.md](../../ops/analysis-decision-framework.md) **P0 checklist**: R41 grounded claims, R4* validated sufficiency, R1 confidence->dynamic, R2 hypothesis exit, R43 deadlock replan (under feasibility gate), R8/R23 no default malice/IOC. Multi-module -> R50; anti-analysis effort -> R51 + A-T cookbook.
```text
□ Finding：算法/校验逻辑/可利用点 / 行为结论
□ Path：callflow 或 solve 步骤挂 E-*
skills/scripts/verify-routing-coherence.ps1
+18 -1
View File
@@ -449,6 +449,23 @@ if ($fail.Count -gt 0) {
$fail | Set-Content (Join-Path $ScratchDir 'failures.txt') -Encoding UTF8
exit 1
}
Write-Host 'ALL ROUTING COHERENCE CHECKS PASSED' -ForegroundColor Green
Write-Host '
# Issue #77 — analysis decision framework anchors
$adf = Join-Path $PackageRoot "skills\ops\analysis-decision-framework.md"
Assert-FileExists $adf "analysis-decision-framework.md (issue #77)"
Assert-FileContains $adf "R4*" "ADF R4* validated sufficiency"
Assert-FileContains $adf "E-insufficient-evidence" "ADF E-insufficient-evidence"
Assert-FileContains $adf "E-hypothesis-confirmed" "ADF hypothesis evidence"
Assert-FileContains $adf "ungrounded" "ADF ungrounded flag"
Assert-FileContains $adf "Not** a second master" "ADF not second master workflow"
$efp77 = Join-Path $PackageRoot "skills\ops\evidence-finding-path.md"
Assert-FileContains $efp77 "analysis-decision-framework" "evidence-finding-path hooks ADF"
Assert-FileContains $efp77 "E-insufficient-evidence" "evidence-finding-path R4* id"
$wf77 = Join-Path $PackageRoot "skills\reverse-engineering\references\re-agent-workflow.md"
Assert-FileContains $wf77 "analysis-decision-framework" "re-agent-workflow hooks ADF"
$rules77 = Join-Path $PackageRoot "RULES.md"
Assert-FileContains $rules77 "analysis-decision-framework" "RULES.md hooks ADF"
ALL ROUTING COHERENCE CHECKS PASSED' -ForegroundColor Green
'ALL ROUTING COHERENCE CHECKS PASSED' | Set-Content (Join-Path $ScratchDir 'verify.txt') -Encoding UTF8
exit 0