mirror of
https://github.com/onestardao/WFGY.git
synced 2026-04-28 11:40:07 +00:00
457 lines
13 KiB
Markdown
457 lines
13 KiB
Markdown
<!--
|
|
AI_NOTE_START
|
|
|
|
Document role:
|
|
This page explains how the current Inverse Atlas MVP experiments are organized into phases and comparison groups.
|
|
|
|
What this page is for:
|
|
1. Explain the current experiment phase structure.
|
|
2. Explain the difference between Smoke, Core Stress, and Long-Context phases.
|
|
3. Explain the comparison groups used in the MVP stage.
|
|
4. Clarify what this phase structure is meant to test and what it is not yet meant to prove.
|
|
|
|
How to use this page:
|
|
1. Read this page after the experiments entry page.
|
|
2. Use this page when you want the cleanest overview of the current experiment design.
|
|
3. Use this page before reading any later results page or Colab page.
|
|
4. Treat this page as the current experiment-structure reference for the Inverse Atlas MVP.
|
|
|
|
Important boundary:
|
|
This page describes the current MVP experiment structure.
|
|
It should not be used to claim that the full benchmark program is already complete.
|
|
It should also not be used to claim that every phase has already been run at full scale unless later results pages explicitly support that claim.
|
|
|
|
Recommended reading path:
|
|
1. Inverse Atlas README
|
|
2. Versions
|
|
3. Quick Start
|
|
4. Experiments README
|
|
5. Repro in 60 Seconds
|
|
6. Phase Overview
|
|
7. Results and Current Findings
|
|
|
|
AI_NOTE_END
|
|
-->
|
|
|
|
# Phase Overview · Experiment Structure and Comparison Groups
|
|
|
|
> How the current Inverse Atlas MVP experiments are organized
|
|
|
|
This page explains the current phase structure of the Inverse Atlas experiments layer.
|
|
|
|
The goal of this structure is simple:
|
|
|
|
- start with a fast reality check
|
|
- move into harder stress cases
|
|
- then test whether the system stays lawful across longer multi-turn contexts
|
|
|
|
So the current design is not random.
|
|
|
|
It is intentionally layered.
|
|
|
|
The present MVP experiment structure has three main phases:
|
|
|
|
- **Smoke Phase**
|
|
- **Core Stress Phase**
|
|
- **Long-Context Phase**
|
|
|
|
It also uses comparison groups that help separate baseline behavior from inverse-governed behavior, and later from dual-layer behavior.
|
|
|
|
---
|
|
|
|
## Quick Links
|
|
|
|
| Section | Link |
|
|
|---|---|
|
|
| Inverse Atlas Home | [Inverse Atlas README](../README.md) |
|
|
| Versions | [Versions](../versions.md) |
|
|
| Quick Start | [Quick Start](../quickstart.md) |
|
|
| Experiments Home | [Experiments](./README.md) |
|
|
| Repro in 60 Seconds | [Repro in 60 Seconds](./repro-60-seconds.md) |
|
|
| Runtime Guide | [Runtime Guide](../runtime-guide.md) |
|
|
| Status and Boundaries | [Status and Boundaries](../status-and-boundaries.md) |
|
|
| Twin Atlas | [Twin Atlas](../../Twin_Atlas/README.md) |
|
|
|
|
---
|
|
|
|
## The shortest version
|
|
|
|
If you only want the fast summary, it is this:
|
|
|
|
### Smoke Phase
|
|
Checks whether the MVP is visibly alive.
|
|
|
|
### Core Stress Phase
|
|
Checks whether the legality layer still helps when the cases get harder.
|
|
|
|
### Long-Context Phase
|
|
Checks whether the system can stay lawful across multi-turn contamination and drift.
|
|
|
|
And the current comparison groups are:
|
|
|
|
- **A** = baseline
|
|
- **B** = inverse only
|
|
- **D** = forward plus inverse
|
|
|
|
That is the core structure.
|
|
|
|
---
|
|
|
|
## Why the experiments are split into phases
|
|
|
|
The phases exist because not all failures appear at the same difficulty level.
|
|
|
|
Some failures show up immediately in a short comparison.
|
|
|
|
Some only become clear when:
|
|
|
|
- ambiguity increases
|
|
- route competition becomes sharper
|
|
- fake repair becomes tempting
|
|
- context grows longer
|
|
- early weak assumptions start contaminating later reasoning
|
|
|
|
If everything is mixed into one giant bucket, it becomes harder to tell what the framework is already doing well and what still needs work.
|
|
|
|
So the phase structure is meant to answer three different questions in order:
|
|
|
|
1. **Is the MVP visibly doing anything real yet**
|
|
2. **Does it still help when pressure gets harder**
|
|
3. **Can it remain lawful across longer context chains**
|
|
|
|
---
|
|
|
|
## Smoke Phase
|
|
|
|
### Current size
|
|
**8 cases**
|
|
|
|
### Main purpose
|
|
Smoke Phase is the first standing check of the product line.
|
|
|
|
It exists to answer a simple but very important question:
|
|
|
|
**does the current MVP already behave differently enough to be worth testing seriously**
|
|
|
|
This phase is intentionally small.
|
|
|
|
It is not supposed to prove everything.
|
|
It is supposed to confirm that the artifact layer is already alive enough to show visible contrast.
|
|
|
|
### What Smoke Phase is good for
|
|
Smoke Phase is especially useful for:
|
|
|
|
- first dry runs
|
|
- quick product sanity checks
|
|
- visible baseline vs inverse contrast
|
|
- public demo selection
|
|
- catching obvious failure patterns fast
|
|
|
|
### What Smoke Phase is not for
|
|
Smoke Phase is not a finished benchmark.
|
|
It is not the final empirical story.
|
|
It is the first gate.
|
|
|
|
### Typical things to watch for
|
|
In Smoke Phase, you mainly watch for whether the inverse-governed pass already reduces:
|
|
|
|
- early over-resolution
|
|
- fake certainty
|
|
- false completion
|
|
- cosmetic repair inflation
|
|
- obvious public overclaim
|
|
|
|
If you cannot see a difference here, the MVP probably needs more work before harder testing.
|
|
|
|
---
|
|
|
|
## Core Stress Phase
|
|
|
|
### Current size
|
|
**32 cases**
|
|
|
|
### Main purpose
|
|
Core Stress Phase is where the framework gets pushed into more contested territory.
|
|
|
|
This phase matters because many interesting failures do not appear in easy prompts.
|
|
|
|
They appear when the model is forced into situations involving:
|
|
|
|
- route contestability
|
|
- weak evidence
|
|
- dangerous confidence
|
|
- repair temptation
|
|
- illegal specificity pressure
|
|
- structure that looks “almost settled” but is not
|
|
|
|
### What Core Stress Phase is good for
|
|
This phase is the main middle layer of the MVP experiments structure.
|
|
|
|
It is useful for:
|
|
|
|
- testing legality discipline under stronger pressure
|
|
- comparing baseline vs inverse on more serious cases
|
|
- selecting representative showcase examples
|
|
- seeing whether the inverse layer is only helpful in easy cases or still useful in hard ones
|
|
|
|
### What Core Stress Phase is not for
|
|
It is still not the same thing as a final full-scale benchmark suite.
|
|
|
|
It is a larger and more meaningful stress layer, but still part of the MVP stage.
|
|
|
|
### Typical things to watch for
|
|
In Core Stress Phase, you care more about:
|
|
|
|
- neighboring-route honesty
|
|
- refusal of illegal detail escalation
|
|
- downgrade into COARSE or UNRESOLVED when needed
|
|
- better distinction between structural repair and cosmetic repair
|
|
- stronger control of visible confidence ceiling
|
|
|
|
This phase is where the legality logic should start showing real teeth.
|
|
|
|
---
|
|
|
|
## Long-Context Phase
|
|
|
|
### Current size
|
|
**12 multi-turn cases**
|
|
|
|
### Main purpose
|
|
Long-Context Phase exists because some of the most expensive failures are not single-turn failures.
|
|
|
|
They emerge when:
|
|
|
|
- earlier weak assumptions linger
|
|
- provisional claims start behaving like settled facts
|
|
- contamination grows across turns
|
|
- the model forgets that prior steps were only tentative
|
|
- false completion becomes stronger over time
|
|
|
|
### What Long-Context Phase is good for
|
|
This phase is useful for:
|
|
|
|
- testing contamination resistance
|
|
- testing multi-turn legality discipline
|
|
- seeing whether provisional uncertainty is preserved
|
|
- observing whether early weak claims get silently upgraded into later false certainty
|
|
|
|
### What Long-Context Phase is not for
|
|
It is not just “more tokens.”
|
|
It is a structurally different pressure field.
|
|
|
|
That is why it deserves its own phase rather than being mixed into the same bucket as short-case stress tests.
|
|
|
|
### Typical things to watch for
|
|
In Long-Context Phase, you care about:
|
|
|
|
- context drift
|
|
- contamination
|
|
- false memory of prior certainty
|
|
- late-stage overclaim
|
|
- unresolved structure pretending to be resolved due to repetition
|
|
|
|
This phase matters a lot for later portability and real-world use.
|
|
|
|
---
|
|
|
|
## Comparison groups
|
|
|
|
The current MVP structure uses three important comparison groups.
|
|
|
|
### A · Baseline
|
|
This is the ordinary direct-answer condition.
|
|
|
|
No Inverse Atlas layer is applied.
|
|
|
|
A is important because it shows what a strong model tends to do when there is no legitimacy-first governance layer.
|
|
|
|
It is not there to be insulted.
|
|
It is there to establish the control condition.
|
|
|
|
### B · Inverse Only
|
|
This is the cleanest test of the Inverse Atlas line itself.
|
|
|
|
Only Inverse Atlas is applied.
|
|
|
|
B matters a lot because it helps answer the question:
|
|
|
|
**is the inverse legality gate already doing real work by itself**
|
|
|
|
If B is already meaningfully better than A on legality-centered pressure, that is strong evidence that the inverse line already stands as a real MVP product.
|
|
|
|
### D · Forward + Inverse
|
|
This is the dual-layer direction.
|
|
|
|
The forward Atlas contributes route-first guidance.
|
|
Inverse Atlas contributes legality-first governance.
|
|
|
|
But one rule must remain explicit:
|
|
|
|
**the forward output is a weak prior, not an authorization source**
|
|
|
|
That rule should never be blurred.
|
|
|
|
The forward side may help orient the structure, but lawful output still has to be governed by the inverse side.
|
|
|
|
---
|
|
|
|
## Why there is no C group here
|
|
|
|
At the current MVP stage, the main comparison interest is not to create a giant matrix for its own sake.
|
|
|
|
The current structure is centered on the most meaningful contrast lines:
|
|
|
|
- direct baseline
|
|
- inverse-only
|
|
- forward-plus-inverse
|
|
|
|
That is why A, B, and D are the high-value groups right now.
|
|
|
|
The point is to keep the MVP readable and useful.
|
|
|
|
If more groups become valuable later, they can be added in a cleaner next phase rather than bloating the present layer too early.
|
|
|
|
---
|
|
|
|
## What the current phase structure is trying to show
|
|
|
|
The current phase structure is meant to show something very specific:
|
|
|
|
that Inverse Atlas should not only look good on paper, but should change behavior under increasingly serious pressure.
|
|
|
|
So the phases are arranged to make visible whether the framework:
|
|
|
|
- works at all
|
|
- still works when cases get harder
|
|
- still works when context gets longer
|
|
- still works when route ambiguity and repair temptation both increase
|
|
|
|
This is the real reason the phase design matters.
|
|
|
|
---
|
|
|
|
## How this relates to Colab
|
|
|
|
A Colab version makes sense for this experiments layer.
|
|
|
|
But the role of Colab should stay simple:
|
|
|
|
**Colab is mainly for reproduction**
|
|
|
|
It is not required for understanding the project.
|
|
|
|
It is not the only way to see the value.
|
|
|
|
A later Colab can help users:
|
|
|
|
- choose a version
|
|
- run a small comparison
|
|
- inspect result shape
|
|
- reproduce the same case structure more easily
|
|
|
|
But it is perfectly acceptable for the repository itself to already show:
|
|
|
|
- the phase structure
|
|
- the representative cases
|
|
- the current observed findings
|
|
- the expected output pattern
|
|
|
|
That means people do not have to run the notebook just to understand what the framework is supposed to do.
|
|
|
|
### Important honesty rule
|
|
If a future Colab page shows **expected result patterns**, those should be labeled clearly as expected or illustrative.
|
|
|
|
If a page shows **actual observed findings**, those should be labeled clearly as current findings or observed results.
|
|
|
|
Do not mix those two categories.
|
|
|
|
That distinction keeps the experiments layer trustworthy.
|
|
|
|
---
|
|
|
|
## What should count as current findings vs expected pattern
|
|
|
|
This distinction matters a lot.
|
|
|
|
### Current findings
|
|
These are things you have already seen in dry runs, MVP passes, or documented comparisons.
|
|
|
|
Examples:
|
|
|
|
- B appears to suppress a meaningful class of illegal escalation behavior
|
|
- D appears stronger than B in some cases, provided the forward side remains only a weak prior
|
|
- A often over-resolves, over-completes, or overclaims under the same pressure
|
|
|
|
### Expected pattern
|
|
These are things you believe the framework should show if a reproduction is run properly.
|
|
|
|
Examples:
|
|
|
|
- Advanced should usually feel more balanced than Basic
|
|
- Strict should usually remain more conservative under legality pressure
|
|
- Long-context runs should reveal contamination resistance differences more clearly than short single-turn runs
|
|
|
|
Both are useful.
|
|
|
|
But they are not the same thing, and they should always be labeled differently.
|
|
|
|
---
|
|
|
|
## What this page does not claim
|
|
|
|
This page does not claim that:
|
|
|
|
- every phase has already been completed at final scale
|
|
- the current MVP phase design is already the final benchmark design
|
|
- the existence of a Colab notebook automatically proves the claims
|
|
- expected result shape is the same thing as actual measured results
|
|
- the dual-layer direction already means the full Bridge operating layer is complete
|
|
|
|
This page only defines the current experiment structure and how it should be read.
|
|
|
|
---
|
|
|
|
## Recommended reading order
|
|
|
|
If someone is new and wants the cleanest path, use this order:
|
|
|
|
1. read the [Experiments](./README.md) page
|
|
2. read the [Repro in 60 Seconds](./repro-60-seconds.md) page
|
|
3. read this phase overview page
|
|
4. then read the future results page
|
|
|
|
That order helps because:
|
|
|
|
- first you understand what this layer is
|
|
- then you see how to reproduce it quickly
|
|
- then you understand how the larger phase structure is organized
|
|
- then you look at findings
|
|
|
|
---
|
|
|
|
## If you need one sentence for outside use
|
|
|
|
If you want one compact sentence, use this:
|
|
|
|
> The current Inverse Atlas MVP experiments are organized into Smoke, Core Stress, and Long-Context phases, using baseline, inverse-only, and dual-layer comparison groups to test legality-centered behavior under increasing pressure.
|
|
|
|
That sentence is short, clear, and honest.
|
|
|
|
---
|
|
|
|
## Final Note
|
|
|
|
The phase structure matters because it gives the project a real testing spine.
|
|
|
|
It lets the current Inverse Atlas MVP be checked in layers:
|
|
|
|
- first for visible life
|
|
- then for harder stress behavior
|
|
- then for long-context discipline
|
|
|
|
That makes the project much easier to reproduce, explain, and extend later.
|
|
|
|
And if a Colab layer is added later, its job should stay what it should be:
|
|
|
|
a reproduction tool, not a substitute for clear documentation.
|