xvoon

documenting how AI changes the way I work

[xvoon @xvoon0]

https://t.co/YDDIbtGfXq

[xvoon @xvoon0]

Claude can invent a fact without hallucinating a single word...

Simon Willison, creator of datasette, mapped the context failure that lets ten citations hide a single unsupported claim

the sources look independent until you trace them backward...

one article copies another, five newsletters repeat it and a report cites the entire pile. claude retrieves ten matching pages and mistakes repetition for evidence, even though every claim started with the same unverified sentence

the origin ledger traces every claim back to the earliest source that actually made it

the citation graph groups sources that copied or referenced one another into a single evidence family

the independence gate keeps a claim out of durable memory until separate primary sources support it

retrieval can find ten matching pages... context engineering checks whether they all came from one rumor

[xvoon @xvoon0]

@diamai_ that's what makes it dangerous. the citations lower your guard before you check where they all lead

[xvoon @xvoon0]

Claude can invent a fact without hallucinating a single word...

Simon Willison, creator of datasette, mapped the context failure that lets ten citations hide a single unsupported claim

the sources look independent until you trace them backward...

one article copies another, five newsletters repeat it and a report cites the entire pile. claude retrieves ten matching pages and mistakes repetition for evidence, even though every claim started with the same unverified sentence

the origin ledger traces every claim back to the earliest source that actually made it

the citation graph groups sources that copied or referenced one another into a single evidence family

the independence gate keeps a claim out of durable memory until separate primary sources support it

retrieval can find ten matching pages... context engineering checks whether they all came from one rumor

[Diam @diamai_]

https://t.co/Mi1HjoYVN4

[xvoon @xvoon0]

one file can secretly control everything Claude tells you...

Tom Brown, co-founder of anthropic, found out why one long document can overpower every better source and almost everyone missed it

length starts masquerading as authority...

one report fills most of the context and repeats the same claim across dozens of pages. claude sees it more often than the shorter primary sources, so volume starts looking like confidence

the source budget limits how much of the context any single document can occupy

the repetition filter collapses duplicated claims before claude sees them

the coverage reserve protects space for primary, current and contradictory sources

context windows count tokens... context engineering decides who gets the microphone

[xvoon @xvoon0]

one file can secretly control everything Claude tells you...

Tom Brown, co-founder of anthropic, found out why one long document can overpower every better source and almost everyone missed it

length starts masquerading as authority...

one report fills most of the context and repeats the same claim across dozens of pages. claude sees it more often than the shorter primary sources, so volume starts looking like confidence

the source budget limits how much of the context any single document can occupy

the repetition filter collapses duplicated claims before claude sees them

the coverage reserve protects space for primary, current and contradictory sources

context windows count tokens... context engineering decides who gets the microphone

[Neheart @neheart]

https://t.co/3YjYHyOJCa

[xvoon @xvoon0]

claude can answer the wrong question perfectly..

Daniela Amodei, co-founder of anthropic, traced it to a hidden context routing flaw that sends claude into the wrong memory before it writes a single word

the mistake happens before retrieval...

most systems search the entire memory with the raw prompt. one ambiguous phrase opens the wrong project, pulls relevant documents from the wrong history and gives claude everything it needs to solve a task you never meant to ask

the intent frame turns the raw message into an explicit task, domain, time horizon and output before retrieval begins

the namespace gate opens only the memory and source collections allowed for that intent

the fallback ladder expands the search one layer at a time when evidence is weak instead of flooding the context with everything

memory gives claude more of the past... routing decides which past belongs to the question

i broke down the full context routing system in the article below

[xvoon @xvoon0]

@0xkkai the dated contradictions matter more than the summaries. that's where the actual research starts

[xvoon @xvoon0]

claude can answer the wrong question perfectly..

Daniela Amodei, co-founder of anthropic, traced it to a hidden context routing flaw that sends claude into the wrong memory before it writes a single word

the mistake happens before retrieval...

most systems search the entire memory with the raw prompt. one ambiguous phrase opens the wrong project, pulls relevant documents from the wrong history and gives claude everything it needs to solve a task you never meant to ask

the intent frame turns the raw message into an explicit task, domain, time horizon and output before retrieval begins

the namespace gate opens only the memory and source collections allowed for that intent

the fallback ladder expands the search one layer at a time when evidence is weak instead of flooding the context with everything

memory gives claude more of the past... routing decides which past belongs to the question

i broke down the full context routing system in the article below

[xvoon @xvoon0]

https://t.co/YDDIbtGfXq

[xvoon @xvoon0]

Anthropic found something inside Claude's memory that was supposed to be gone...

Dario Amodei, co-founder of Anthropic, traced it to a context flaw that brings rejected decisions back to life and almost everyone missed it

deleting the note isn't enough...

a decision gets rejected, the file disappears and everyone moves on. but its embedding, summary and old chat still exist, so claude retrieves the dead idea weeks later and treats it like approved context

the tombstone registry marks every rejected fact, instruction and decision as blocked from retrieval

the supersession link connects the dead version to its replacement, so claude receives the current rule and why the old one was rejected

the purge sweep removes every embedding, summary and cached reference still pointing to the outdated version

memory isn't only what you keep... it's what you prevent from returning

[xvoon @xvoon0]

@sindikitil the nasty part is it looks like memory working perfectly until the dead decision starts steering new answers

[xvoon @xvoon0]

Anthropic found something inside Claude's memory that was supposed to be gone...

Dario Amodei, co-founder of Anthropic, traced it to a context flaw that brings rejected decisions back to life and almost everyone missed it

deleting the note isn't enough...

a decision gets rejected, the file disappears and everyone moves on. but its embedding, summary and old chat still exist, so claude retrieves the dead idea weeks later and treats it like approved context

the tombstone registry marks every rejected fact, instruction and decision as blocked from retrieval

the supersession link connects the dead version to its replacement, so claude receives the current rule and why the old one was rejected

the purge sweep removes every embedding, summary and cached reference still pointing to the outdated version

memory isn't only what you keep... it's what you prevent from returning

[Aspect0 @Aspect0F]

https://t.co/69kZWa9yxI

[xvoon @xvoon0]

you've been testing Claude backwards...

Jack Clark, co-founder of anthropic, found the evaluation mistake hiding inside almost every ai workflow and almost everyone missed it

the answer is the last place the failure appears...

most teams grade the final response. when it's wrong, they rewrite the prompt or switch models. but claude may have received an outdated source, missed the current document or been buried in noise before generation even started

the retrieval trace records every source considered, rejected and selected - including why it entered the active context

the context diff compares passing and failing runs to show what disappeared, changed priority or arrived stale

the attribution check forces every claim in claude's answer to point to an active source and rejects anything built from missing or lower-priority evidence

answer evals tell you that claude failed... context evals tell you where the failure entered

[xvoon @xvoon0]

you've been testing Claude backwards...

Jack Clark, co-founder of anthropic, found the evaluation mistake hiding inside almost every ai workflow and almost everyone missed it

the answer is the last place the failure appears...

most teams grade the final response. when it's wrong, they rewrite the prompt or switch models. but claude may have received an outdated source, missed the current document or been buried in noise before generation even started

the retrieval trace records every source considered, rejected and selected - including why it entered the active context

the context diff compares passing and failing runs to show what disappeared, changed priority or arrived stale

the attribution check forces every claim in claude's answer to point to an active source and rejects anything built from missing or lower-priority evidence

answer evals tell you that claude failed... context evals tell you where the failure entered

[xvoon @xvoon0]

https://t.co/YDDIbtGfXq

[xvoon @xvoon0]

@codymclain it should move with the task. otherwise you're just replacing one rigid rule with another

[xvoon @xvoon0]

stop putting your most important instruction first...

Dario Amodei, co-founder of anthropic, found the hidden placement flaw making claude ignore the rule that matters most and almost everyone missed it

the instruction isn't missing... it's buried

most people put permanent rules first, then add pages of files, notes and examples. by the time claude reaches the actual task, the rule is still inside the context - but too far from the decision it needs to control

the execution capsule repeats the active goal, non-negotiable constraints and output contract directly beside the request

the evidence stack orders sources by how much they can change the decision instead of when they entered the conversation

the context checksum makes claude name the rules controlling its answer and flag any conflict before generating it

context length decides what fits... context placement decides what gets used

[cesar ai @caesar_aii]

https://t.co/HgjxpCnFck

[xvoon @xvoon0]

stop putting your most important instruction first...

Dario Amodei, co-founder of anthropic, found the hidden placement flaw making claude ignore the rule that matters most and almost everyone missed it

the instruction isn't missing... it's buried

most people put permanent rules first, then add pages of files, notes and examples. by the time claude reaches the actual task, the rule is still inside the context - but too far from the decision it needs to control

the execution capsule repeats the active goal, non-negotiable constraints and output contract directly beside the request

the evidence stack orders sources by how much they can change the decision instead of when they entered the conversation

the context checksum makes claude name the rules controlling its answer and flag any conflict before generating it

context length decides what fits... context placement decides what gets used

[xvoon @xvoon0]

@0xhhoi u know what

[xvoon @xvoon0]

you're still letting ai choose which source is true???

Jensen Huang, co-founder of Nvidia, exposed the retrieval flaw feeding models outdated information and almost everyone missed it

the failure happens before claude reads a single word...

most rag systems rank documents by similarity. an outdated draft using the same words as your question can beat the current policy written differently, so claude answers from the nearest chunk instead of the source allowed to decide

the authority filter removes drafts, notes and model outputs before similarity search begins

the version gate retrieves only documents whose status and effective date are valid for the current task

the conflict bundle gives claude the winning source, the rejected version and the reason one overrides the other

similarity finds what sounds right... source priority decides what is right

i broke down the full retrieval system in the article below

[xvoon @xvoon0]

you're still letting ai choose which source is true???

Jensen Huang, co-founder of Nvidia, exposed the retrieval flaw feeding models outdated information and almost everyone missed it

the failure happens before claude reads a single word...

most rag systems rank documents by similarity. an outdated draft using the same words as your question can beat the current policy written differently, so claude answers from the nearest chunk instead of the source allowed to decide

the authority filter removes drafts, notes and model outputs before similarity search begins

the version gate retrieves only documents whose status and effective date are valid for the current task

the conflict bundle gives claude the winning source, the rejected version and the reason one overrides the other

similarity finds what sounds right... source priority decides what is right

i broke down the full retrieval system in the article below

[Mikadzyki🌙 @Mikadzyki_NFT]

https://t.co/PgwTlyjqP8

[xvoon @xvoon0]

Claude can turn one hallucination into permanent memory...

Demis Hassabis, co-founder of deepmind, found the hidden context loop making ai cite its own mistakes as facts and almost nobody noticed

the loop starts after the answer...

most teams save every useful claude response back into memory. one invented detail gets stored beside real sources, returns in the next session as retrieved context, and suddenly claude is no longer hallucinating - it's "remembering"

the evidence boundary labels every item as an external source, human decision or model output - only the first two can become trusted memory

the writeback gate refuses to save any model claim without a source link or explicit human approval

the expiry check forces old facts to be verified again before retrieval can promote them into active context

memory makes an error persistent... provenance makes knowledge trustworthy

i broke down the full memory writeback system in the article below