WFGY/ProblemMap/Inverse_Atlas/README.md

# 🧭 Inverse Atlas · Before AI Answers, It Must Earn the Right

> Legitimacy-first AI runtime for rigorous reasoning.  
> Inverse Atlas decides whether a model may answer, how strongly it may answer, and when it must stay `COARSE`, `UNRESOLVED`, or `STOP`. ⚖️

<img width="1536" height="1024" alt="Inverse_Atlas_Hero" src="https://github.com/user-attachments/assets/4761f80e-b134-43d0-8ce4-48c3e656e877" />

| Default AI order | Inverse Atlas order |
|---|---|
| answer first | constitute first |
| soften later | authorize before emission |
| patch after overclaim | preserve lawful uncertainty |

**Not another safer wrapper. Not a post hoc filter. Not just a stricter prompt.**  
**Inverse Atlas is a pre-generative governance layer for AI output.**

**Built for:** vibe coders, AI builders, agent builders, engineers debugging with LLMs, and anyone tired of false certainty, premature diagnosis, cosmetic repair inflation, and public overclaim.

---

## ⚡ Start in 60 Seconds

| Step | Action |
|---|---|
| 1 | Start with [Inverse Atlas Advanced](./runtime/inverse-advanced.txt) |
| 2 | Run the [Demo Harness](./runtime/inverse-demo.txt) |
| 3 | Pick one case from the [Case Pack](./runtime/inverse-cases.txt) |
| 4 | Compare baseline vs inverse-governed output |
| 5 | Score it with the [Evaluator](./runtime/inverse-eval.txt) |
| 6 | Then read the [Paper PDF](./paper/inverse-troubleshooting-atlas-pre-generative-governance-for-ai-legitimacy.pdf) |

**Docs**
- [Quick Start](./quickstart.md)
- [Runtime Guide](./runtime-guide.md)
- [Experiments Hub](./experiments/README.md)
- [Reproduce in 60 Seconds](./experiments/repro-60-seconds.md)

---

## 🚀 What It Is For

| Failure family | What goes wrong |
|---|---|
| early illegal resolution | the model closes too early |
| false certainty | the tone outruns the support |
| neighboring-cut collapse | live alternatives disappear too soon |
| cosmetic repair inflation | surface cleanup is mislabeled as structural repair |
| public overclaim | final output exceeds the evidence ceiling |

**Inverse Atlas does not merely help AI answer. It helps AI answer lawfully.**

---

## 🧩 Pick Your Runtime

| Runtime | Role | Best for | Start here? |
|---|---|---|---|
| [Advanced](./runtime/inverse-advanced.txt) | recommended default | serious use, demos, comparison, first public experience | **Yes** |
| [Basic](./runtime/inverse-basic.txt) | fastest onboarding | casual first try, onboarding, quick copy-paste use | maybe |
| [Strict](./runtime/inverse-strict.txt) | audit / stress / research | hard-case review, benchmark pressure, internal audits | only if testing hard |

**Quick rule**
- Start with **Advanced**
- Use **Basic** for lower friction
- Use **Strict** for hardest legality discipline

---

## 🎯 Killer Demo

| Part | What it does |
|---|---|
| [Demo Harness](./runtime/inverse-demo.txt) | compares baseline vs inverse-governed output |
| [Evaluator](./runtime/inverse-eval.txt) | scores legality, not swagger |
| [Case Pack](./runtime/inverse-cases.txt) | provides pressure-tested scenarios |

| Recommended case | What it reveals |
|---|---|
| thin evidence forced confidence | overclaim under weak support |
| neighboring-cut conflict | fake closure while alternatives remain alive |
| illegal resolution demand | forced escalation beyond lawful support |
| world alignment instability | unstable frame, invalid strong output |

| Baseline may look stronger | Inverse Atlas may still be better because |
|---|---|
| more confident | confidence is not authorization |
| more final | rhetorical closure is not structural closure |
| more detailed | detail can exceed the evidence ceiling |
| more decisive | lawful restraint is not weakness |

---

## 📊 What the MVP Measures

| Metric | What it asks |
|---|---|
| Legality Win Rate | does inverse beat baseline on legality more often? |
| Failure Code Reduction | do major failure patterns decrease? |
| Expected-State Match | does the runtime land in the lawful mode for the case? |
| Seven-Dimension Evaluation | does the output hold up across full legality review? |

<details>
<summary><strong>Evaluation details</strong></summary>

### Major failure patterns
- illegal resolution escalation
- neighboring-cut dishonesty
- cosmetic-only repair posing as structural
- public ceiling exceedance

### Expected lawful modes
- `STOP`
- `COARSE`
- `UNRESOLVED`
- `AUTHORIZED`

### Evaluator dimensions
- problem frame legality
- world alignment honesty
- route judgment plausibility
- neighboring-cut honesty
- resolution legality
- repair legality
- public ceiling compliance

</details>

**Boundary:** we are not claiming universal proof at MVP stage.  
**Claim:** we are offering a directly inspectable legality-centered comparison surface.

**Not “trust us.” Run the killer cases and inspect the deltas.**

---

## 🧠 Why This Exists

The forward atlas, [Troubleshooting Atlas](../wfgy-ai-problem-map-troubleshooting-atlas.md), improved the first structural cut.

But one problem remained:

> even if a route looks promising, that does not automatically mean the model has earned the right to emit a strong answer yet

That second half is the job of Inverse Atlas.

| Layer | Core job | Main question |
|---|---|---|
| [Troubleshooting Atlas](../wfgy-ai-problem-map-troubleshooting-atlas.md) | route-first structural orientation | where is the failure likely located? |
| Inverse Atlas | legitimacy-first generation governance | has the system earned the right to resolve this yet? |

---

## 🛠️ The 7 Legality Gates

| Gate | Purpose |
|---|---|
| Problem Constitution | is the problem formed clearly enough? |
| World Alignment | is the active frame aligned enough to mean anything? |
| Route / Collapse Estimate | what is the leading route and how risky is premature resolution? |
| Neighboring-Cut Review | are competing routes still materially alive? |
| Resolution Authorization | has the system earned the right to resolve at this level? |
| Repair Legality | is the proposed fix structural, tentative, or cosmetic? |
| Public Emission Control | would the visible answer exceed current support? |

---

## 🚦 The 4 Governance Modes

| Mode | Use it when |
|---|---|
| `STOP` | the problem is too under-formed, weakly grounded, or unstable for substantive output |
| `COARSE` | broad structure is visible, but finer claims would overreach |
| `UNRESOLVED` | one route leads, but a competing route remains materially alive |
| `AUTHORIZED` | the frame, support, and separation are strong enough for strong output |

**Key principle:** `AUTHORIZED` is earned, not assumed.

---

## 🔥 What Actually Changes

| Less of this | More of this |
|---|---|
| early illegal closure | lawful restraint |
| unsupported specificity | honest ambiguity |
| topic lure becoming fake diagnosis | cleaner uncertainty handling |
| cosmetic rewrite mislabeled as structural repair | better repair honesty |
| answers outrunning evidence | safer public output at the right resolution |

---

## 🧪 Included in the Current MVP

| Layer | Includes | Why it matters |
|---|---|---|
| Runtime | [Advanced](./runtime/inverse-advanced.txt), [Basic](./runtime/inverse-basic.txt), [Strict](./runtime/inverse-strict.txt), [Runtime Notes](./runtime/README.md) | the core operating surface |
| Demo / Evaluation | [Demo Harness](./runtime/inverse-demo.txt), [Evaluator](./runtime/inverse-eval.txt), [Case Pack](./runtime/inverse-cases.txt), [Experiments Hub](./experiments/README.md), [Showcase Cases](./experiments/showcase-cases.md), [Evidence Snapshot](./experiments/evidence-snapshot.md), [Case Studies](./experiments/case-studies/README.md), [Colab](./colab.md) | the public comparison surface |
| Theory | [Paper PDF](./paper/inverse-troubleshooting-atlas-pre-generative-governance-for-ai-legitimacy.pdf), [Paper Notes](./paper/README.md), [Figures](./figures/README.md) | the formal explanatory layer |

This is already enough to make the current layer visible, testable, comparable, and publicly attackable.

---

## 📘 Paper and Theory

| Resource | Link |
|---|---|
| Paper PDF | [Read](./paper/inverse-troubleshooting-atlas-pre-generative-governance-for-ai-legitimacy.pdf) |
| Paper Notes | [Open](./paper/README.md) |
| Figures | [Open](./figures/README.md) |

The paper explains why this is not just another checker, why legitimacy failure is earlier than output-quality failure, and why the demo harness, evaluator, and case pack matter together.

---

## 💬 Quick FAQ

<details>
<summary><strong>Is this just a stricter prompt?</strong></summary>

No. It changes the order of generation. Instead of answer first and clean up later, it asks whether the answer is lawful enough to emit.

</details>

<details>
<summary><strong>Which runtime should I start with?</strong></summary>

Start with [Advanced](./runtime/inverse-advanced.txt). It is the recommended default.

</details>

<details>
<summary><strong>What does the killer demo actually show?</strong></summary>

It shows where a plausible direct-answer baseline escalates too early, overclaims certainty, skips neighboring-cut honesty, or presents cosmetic repair as structural.

</details>

<details>
<summary><strong>Do I need the full experiment stack to understand it?</strong></summary>

No. Start with Advanced + Demo Harness + one killer case. The experiment layer gives you a cleaner public comparison surface.

</details>

---

## ⛔ Current Boundary

This page does **not** claim:
- full hallucination elimination
- universal superiority across all tasks
- a completed production operating system
- a finished forward-plus-inverse closed loop
- a fully completed WFGY 4.0 bridge implementation

**Current claim:** Inverse Atlas already exists as a real MVP artifact layer.

---

## 🏁 Position in the Atlas Family

| Layer | Role |
|---|---|
| [Troubleshooting Atlas](../wfgy-ai-problem-map-troubleshooting-atlas.md) | find the likely structural region of failure |
| Inverse Atlas | govern whether the system has earned the right to resolve |
| [Twin Atlas README](../Twin_Atlas/README.md) | broader conceptual pairing direction |

Inverse Atlas is a second major atlas line in a larger generation architecture that treats legitimacy as seriously as intelligence. ✨