Update README.md

This commit is contained in:
PSBigBig × MiniPS 2026-02-10 16:15:18 +08:00 committed by GitHub
parent ae8b9507d7
commit db1b2837a1
No known key found for this signature in database
GPG key ID: B5690EEEBB952194

View file

@ -1,3 +1,137 @@
<!--
Search Anchor:
reasoning global fix map
reasoning failure modes
long chain stability
chain of thought stability
cot variance clamp
reasoning collapse map
hallucination reentry
recursive reasoning loop
logic collapse invariants
symbolic collapse abstraction
proof dead ends unprovable steps
anchoring bridge proofs
context stitching window joins
redundant evidence collapse
entropy overload long plans
llm planning never converge
multi step answer drift
rerun same question different answer
window roll loses anchor
long chat loses earlier facts
plan keeps growing never lands
model rewrites reality in long plans
drift across seeds and paraphrases
auditable reasoning steps
deterministic step count
Typical reasoning bugs:
steps keep growing and answer never lands
model goes in circles and repeats itself
model re asserts false claim after correction
reasoning jumps without citing source step
invariants vanish during abstraction jump
proof chain stalls at unprovable step
model changes interpretation mid chain
long context window drops earlier anchor
window joins flip the claim
different runs choose different evidence
reruns disagree on the same question
minority facts drowned by many similar snippets
model cherry picks evidence to support wrong answer
different seeds give different final conclusion
answer depends on irrelevant snippet ordering
model conflates hypothetical with real data
When to use this folder:
rag already fixed but answers still drift
retrieval stable yet explanation changes every run
you see hallucination only in long plans
cot used but not reproducible
agent loop does many steps without progress
mathematical proof suddenly switches frame
safety or governance chain needs audit trail
researcher wants to compare seeds and paraphrases
benchmark requires stable long chain behavior
legal or medical reasoning must be traceable
Key metrics and targets:
delta s question selected evidence <= 0.45 on all runs
coverage of target section >= 0.70 with cite then explain
lambda observe convergent across three paraphrases and two seeds
e_resonance flat across long windows and window joins
deterministic step count per stage after clamps applied
no hallucination reentry after correction
same conclusion across seeds given same evidence
proof chain explicit about each jump and its source
no hidden frame jumps or abstraction collapse
reasons and evidence stay aligned across windows
Core pages in this folder:
ProblemMap/GlobalFixMap/Reasoning/entropy-overload.md
ProblemMap/GlobalFixMap/Reasoning/recursive-loop.md
ProblemMap/GlobalFixMap/Reasoning/hallucination-reentry.md
ProblemMap/GlobalFixMap/Reasoning/logic-collapse.md
ProblemMap/GlobalFixMap/Reasoning/symbolic-collapse.md
ProblemMap/GlobalFixMap/Reasoning/proof-dead-ends.md
ProblemMap/GlobalFixMap/Reasoning/anchoring-and-bridge-proofs.md
ProblemMap/GlobalFixMap/Reasoning/context-stitching-and-window-joins.md
ProblemMap/GlobalFixMap/Reasoning/chain-of-thought-variance-clamp.md
ProblemMap/GlobalFixMap/Reasoning/redundant-evidence-collapse.md
Related structural fixes:
ProblemMap/rag-architecture-and-recovery.md
ProblemMap/retrieval-playbook.md
ProblemMap/retrieval-traceability.md
ProblemMap/data-contracts.md
ProblemMap/context-drift.md
ProblemMap/entropy-collapse.md
ProblemMap/rerankers.md
ProblemMap/embedding-vs-semantic.md
ProblemMap/Embeddings/metric_mismatch.md
ProblemMap/Embeddings/normalization_and_scaling.md
Reasoning scenarios:
multi step financial plan changes answer across runs
safety policy chain contradicts itself later
math proof switches definition mid argument
system design review flips conclusion with same docs
governance decision tree loops on same node
research summary hallucinate extra papers
cot explanation disagrees with final answer
agent rewrites the task mid execution
long back and forth chat forgets original constraint
model keeps expanding plan without committing
Signals to check:
coverage good but lambda unstable
delta s low on evidence but conclusion different
citations stable but reasoning text drifts
proof steps not numbered or auditable
no explicit reference to step that caused jump
branching factor keeps increasing per step
model keeps asking for more context
seed one and seed two disagree on final verdict
paraphrase question changes the reasoning path
-->
<!--
Cross folder jumps:
ProblemMap/GlobalFixMap/Reasoning/README.md
ProblemMap/rag-architecture-and-recovery.md
ProblemMap/retrieval-playbook.md
ProblemMap/retrieval-traceability.md
ProblemMap/context-drift.md
ProblemMap/entropy-collapse.md
ProblemMap/rerankers.md
ProblemMap/GlobalFixMap/Retrieval/README.md
ProblemMap/GlobalFixMap/Embeddings/README.md
ProblemMap/GlobalFixMap/Chunking/README.md
ProblemMap/SemanticClinicIndex.md
-->
# Reasoning — Global Fix Map
<details>
@ -6,7 +140,7 @@
<br>
> You are in a specialist desk.
> For full triage and doctors on duty, return here:
> For full triage and doctors on duty, return here: a
>
> - [**WFGY Global Fix Map** — main Emergency Room, 300+ structured fixes](https://github.com/onestardao/WFGY/blob/main/ProblemMap/GlobalFixMap/README.md)
> - [**WFGY Problem Map 1.0** — 16 reproducible failure modes](https://github.com/onestardao/WFGY/blob/main/ProblemMap/README.md)
@ -21,6 +155,27 @@ Every page maps symptoms to exact WFGY fixes with measurable targets. No infra c
---
<!--
Anchor Menu:
open: entropy overload guide ProblemMap/GlobalFixMap/Reasoning/entropy-overload.md
open: recursive loop guide ProblemMap/GlobalFixMap/Reasoning/recursive-loop.md
open: hallucination reentry guide ProblemMap/GlobalFixMap/Reasoning/hallucination-reentry.md
open: logic collapse guide ProblemMap/GlobalFixMap/Reasoning/logic-collapse.md
open: symbolic collapse guide ProblemMap/GlobalFixMap/Reasoning/symbolic-collapse.md
open: proof dead ends guide ProblemMap/GlobalFixMap/Reasoning/proof-dead-ends.md
open: anchoring and bridge proofs guide ProblemMap/GlobalFixMap/Reasoning/anchoring-and-bridge-proofs.md
open: context stitching and window joins guide ProblemMap/GlobalFixMap/Reasoning/context-stitching-and-window-joins.md
open: chain of thought variance clamp guide ProblemMap/GlobalFixMap/Reasoning/chain-of-thought-variance-clamp.md
open: redundant evidence collapse guide ProblemMap/GlobalFixMap/Reasoning/redundant-evidence-collapse.md
jump: rag architecture and recovery visual map ProblemMap/rag-architecture-and-recovery.md
jump: retrieval playbook knobs and metrics ProblemMap/retrieval-playbook.md
jump: retrieval traceability and data contracts ProblemMap/retrieval-traceability.md ProblemMap/data-contracts.md
jump: long context drift and entropy collapse ProblemMap/context-drift.md ProblemMap/entropy-collapse.md
jump: rerankers for ordering evidence ProblemMap/rerankers.md
-->
## Orientation: what each page solves
| Page | What it fixes | Typical symptom |