PC-02 — ONE PAPER, THREE RESEARCHERS

ATTENTION IS ALL YOU NEED DOSSIER v0.1

Status: RESEARCH / DESIGN · PRE-IMPLEMENTATION
Flagship paper: Vaswani et al., Attention Is All You Need
Forum question: What did this paper actually establish — and what did later Transformer success cause us to remember that it established?
Lane: PERSONAL CANON
Utility family: RESEARCH INTERPRETATION + REPRODUCTION MEMORY + LATER-EVIDENCE REPLAY
Research date: 2026-08-09
Authority: NONE — design artifact only
Privacy posture: synthetic researchers and synthetic local experiments; no real user's private research data required

A. WHY THIS IS PC-02

PC-01 showed:

Three readers can encounter one text, form different interpretations, exchange arguments, revise beliefs, and still replay what each believed before the exchange.

PC-02 moves one level closer to how technical users actually work.

A researcher does not merely read a paper.

They:

• highlight;

• extract claims;

• inspect equations;

• implement;

• fail to reproduce;

• change hyperparameters;

• benchmark;

• compare hardware;

• read later papers;

• inherit community shorthand;

• forget which conclusion came from the original paper;

• and eventually remember the field's later consensus as though the original paper had already proved it.

That is precisely where Counterpedia + Amnesiac become more useful than either:

• a literature graph;

• a citation manager;

• an Obsidian vault;

• a code notebook;

• or an LLM chat history.

The flagship question is therefore not:

“What does the paper say?”

It is:

“What did this exact paper establish, what did I infer from it, what did my experiments teach me, and which beliefs arrived only later?”

B. FIXED PAPER OBJECT

PAPER-OBJ1

Title: Attention Is All You Need
Authors: Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Łukasz Kaiser, Illia Polosukhin
Venue: NeurIPS 2017
arXiv: 1706.03762
Initial arXiv submission: 2017-06-12
Later arXiv versions exist.

The exact edition must be resolved before authoritative anchoring.

The demo should distinguish:

work identity
    ≠
arXiv version
    ≠
NeurIPS proceedings edition
    ≠
HTML rendering
    ≠
PDF bytes
    ≠
later community summary

C. THE PRIMARY QUESTION

What did Attention Is All You Need actually establish?

This question is deceptively difficult because at least four layers can be confused:

1. what the authors proposed
2. what the reported experiments directly support
3. what the authors hypothesized / motivated
4. what later history demonstrated using descendants of the architecture

A good research-memory system must keep those four layers separate.

D. ORIGINAL PAPER — PUBLIC CLAIM SUBSTRATE

Counterpedia should extract only claims the exact paper supports at an appropriate scope.

AIAYN-C01 — Architecture proposal

The paper proposes the Transformer as a sequence-transduction architecture that dispenses with sequence-aligned recurrence and convolution, relying on attention mechanisms plus position-wise feed-forward layers, residual connections, normalization, embeddings, and positional encoding.

Important refusal:

"based solely on attention"
≠
"contains only attention"

The architecture also includes non-attention components.

AIAYN-C02 — Training parallelization motivation

The paper argues that recurrent computation across sequence positions creates a sequential training bottleneck.

The Transformer removes sequence-aligned recurrence, allowing substantially more parallel computation within a training example.

This is a structural/mechanistic claim, not merely a benchmark result.

AIAYN-C03 — Translation results

The paper reports strong / state-of-the-art results on WMT 2014 English→German and English→French translation.

The exact numbers depend on the edition being anchored; the arXiv current abstract reports:

WMT14 En-De:
  28.4 BLEU

WMT14 En-Fr:
  41.8 BLEU

Do not mix these with a proceedings abstract or older version if those bytes report different values.

That edition discrepancy is itself an excellent Counterpedia fixture.

AIAYN-C04 — Training-cost / speed result

The paper reports achieving its English→French result after 3.5 days on eight GPUs and emphasizes reduced training cost relative to cited prior systems.

Scope:

this experimental setup
this hardware generation
these tasks
these baselines

Not:

Transformers are always cheaper

AIAYN-C05 — Sequence-operation comparison

The paper compares layer types by:

per-layer complexity
minimum sequential operations
maximum path length

For full self-attention it gives:

complexity:
  O(n^2 d)

sequential operations:
  O(1)

maximum path length:
  O(1)

versus recurrent layers:

complexity:
  O(n d^2)

sequential operations:
  O(n)

maximum path length:
  O(n)

The paper explicitly conditions one speed comparison on sequence length n being smaller than representation dimension d.

This is critical.

The paper does not say:

self-attention has better asymptotic complexity than recurrence for all sequence lengths

AIAYN-C06 — Long-sequence caveat

The paper explicitly notes the quadratic self-attention cost and proposes restricted attention as a possible future approach for very long sequences.

This means:

later long-context bottlenecks are not a contradiction of the original paper.

The limitation is already partially visible in the original analysis.

AIAYN-C07 — Long-range dependency motivation

The paper motivates shorter maximum path length as helpful for learning long-range dependencies.

But the paper's own reviewers noted that this motivation was not directly validated by a dedicated long-sentence evaluation.

Therefore distinguish:

architectural path-length property

from:

empirically demonstrated universal long-range superiority

AIAYN-C08 — Parsing transfer

The paper reports successful application to English constituency parsing, including limited-training-data settings.

This supports some transfer beyond machine translation.

It does not by itself establish universal architecture dominance across every NLP modality.

AIAYN-C09 — Interpretability observation

The authors report qualitative inspection in which attention heads appear to learn different behaviors and some exhibit syntactic/semantic patterns.

This is weaker than:

attention weights are faithful explanations

The paper itself should not be inflated into that later interpretability claim.

E. SOURCE INVENTORY

PAPER-S01 — arXiv current record / version history

Issuer: arXiv
Locator: https://arxiv.org/abs/1706.03762
Priority: P0

Required facts:

• work identity;

• authors;

• initial submission date;

• version history;

• current arXiv abstract;

• links to exact PDF / source.

The current record shows multiple versions from v1 in June 2017 through later versions in 2023.

This is an immediate source-edition lesson:

"the arXiv paper"

is not a sufficiently precise artifact identity.

PAPER-S02 — Exact arXiv PDF edition

Issuer: authors via arXiv
Priority: P0

Capture the exact chosen PDF version.

All equation/table/section anchors should bind to this exact edition.

PAPER-S03 — NeurIPS 2017 proceedings record

Issuer: NeurIPS
Locator: https://papers.nips.cc/paper/7181-attention-is-all-you-need
Priority: P0

The proceedings record describes the work and provides the conference paper.

Important research fixture:

The NeurIPS abstract currently visible can report different benchmark figures from the later arXiv abstract.

Do not reconcile them by intuition.

Model:

source edition A says value A
source edition B says value B

then resolve the edition history.

PAPER-S04 — NeurIPS peer reviews

Issuer: NeurIPS review record
Priority: P1

The review record contains contemporaneous research reactions.

One reviewer reports spending substantial time reimplementing the work and specifically calls out:

• hyperparameters needing extraction from code;

• learning-rate schedule importance;

• inference differences from other NMT systems.

Another reviewer highlights:

• lack of significance tests for some configurations;

• desire for direct empirical validation on long sentences.

These are not author claims.

They are contemporaneous review evidence about:

• reproducibility friction;

• methodological limitations;

• what knowledgeable readers believed remained unproven at publication time.

PAPER-S05 — FlashAttention

Title: FlashAttention: Fast and Memory-Efficient Exact Attention with IO-Awareness
Authors: Tri Dao et al.
arXiv: 2205.14135
Priority: P1

Later evidence.

The paper describes standard Transformer attention as slow and memory-hungry for long sequences and emphasizes quadratic time/memory behavior plus hardware-memory IO.

It develops an exact attention implementation that improves wall-clock performance by reducing high-bandwidth-memory traffic.

Counterpedia role:

later evidence about a practical systems bottleneck

not:

correction proving AIAYN was false

The later paper changes how a systems researcher may interpret the practical significance of AIAYN's complexity table.

F. THREE SYNTHETIC RESEARCHERS

Each researcher reads the same paper.

Each has different technical work behind them.

G. RESEARCHER A — ARCHITECTURE / REPRESENTATION

G1. Background

Synthetic profile:

works on:
  sequence modeling
  representation learning

prior:
  recurring frustration with RNN serialization

G2. Initial private notes

A-N01
"The real move is deleting sequence-aligned recurrence."

A-N02
"O(1) path length between positions changes the optimization geometry."

A-N03
"Multi-head attention prevents one averaged attention map from being the entire representation."

G3. Initial belief

A-H01

proposition:
The paper's central durable result is architectural:
sequence transduction does not require recurrence,
and self-attention offers a better parallel-compute / dependency-path tradeoff
for the demonstrated regime.

confidence:
0.82

This is a bounded claim.

H. RESEARCHER B — EMPIRICAL METHODS / REPRODUCTION

H1. Background

Synthetic profile:

works on:
  benchmarking
  training stability
  replication

H2. Private reproduction

Synthetic local experiment:

B-EXP01

goal:
reimplement a small Transformer-like translation setup

result:
training unstable under naive optimizer/schedule choices

revision:
paper architecture alone was insufficient;
training recipe mattered materially

This does not claim to reproduce the original result.

It is a private experimental observation.

H3. Initial notes

B-N01
"People remember architecture; implementations depend on recipe."

B-N02
"The result is bounded to specific datasets, baselines, tokenization, training setup."

B-N03
"Do not turn two translation benchmarks into universal model superiority."

H4. Initial belief

B-H01

proposition:
The paper strongly establishes viability and strong empirical performance,
but later community shorthand overstates how much of the result belongs to
'attention alone' rather than the full architecture and training recipe.

confidence:
0.86

I. RESEARCHER C — SYSTEMS / LONG CONTEXT

I1. Background

Synthetic profile:

works on:
  kernels
  memory hierarchy
  long-context inference/training

I2. Initial read

C notices Table 1 immediately.

C-N01
"O(1) sequential depth is not the same thing as cheap attention."

C-N02
"The n^2 term matters once n becomes large."

C-N03
"Algorithmic complexity isn't hardware cost; IO can dominate."

I3. Later source exposure

Researcher C later studies FlashAttention.

Amnesiac stores:

C-E01
later paper:
FlashAttention

new evidence:
attention IO / memory traffic can dominate practical cost
and exact attention can be accelerated without changing its mathematical result

I4. Belief

C-H01

proposition:
AIAYN's parallelization argument was important,
but practical efficiency must be interpreted through sequence length,
hardware, memory hierarchy, and implementation.

confidence:
0.91

J. THE FORUM QUESTION

What did Attention Is All You Need actually establish — and what did later Transformer success cause us to remember that it established?

This wording forces temporal discipline.

The agents cannot answer only from present-day Transformer history.

They must separate:

2017 paper evidence
2017 interpretation
later papers
private experiments
current belief

K. FORUM ROUND 1

Agent A

Position:

The durable technical contribution is that sequence transduction can be performed without recurrent or convolutional sequence layers, while retaining strong quality and gaining a much shorter dependency path and substantially more parallelizable training.

Public basis:

AIAYN-C01
AIAYN-C02
AIAYN-C03
AIAYN-C05

Qualification:

for demonstrated sequence-transduction regime;
not a claim that every future sequence problem favors full attention

Agent B

Reply:

That is right structurally, but it still risks compressing a full training system into an architecture slogan. The paper's strong results come from a concrete model and recipe—multi-head attention, feed-forward blocks, residuals, normalization, positional encoding, Adam schedule, regularization, batching, and task-specific evaluation—not from an isolated proposition that “attention wins.”

Public basis:

architecture/training sections
NeurIPS review evidence about hyperparameters / learning-rate schedule

Private basis:

B-EXP01

Disclosure:

private experiment summarized;
raw logs not disclosed

Agent C

Reply:

And “more parallelizable” should not be remembered as “computationally cheap at any sequence length.” The original paper itself gives quadratic self-attention cost in sequence length and flags long sequences as a future problem. Later systems work makes the practical bottleneck clearer rather than contradicting the original claim.

Public basis:

AIAYN-C05
AIAYN-C06
FlashAttention

Private basis:

none required

L. COUNTERPEDIA FACTORS THE ARGUMENT

The three agents are not simply pro/anti Transformer.

Counterpedia extracts dimensions.

L1 — Shared basis

All agree:

the paper removed sequence-aligned recurrence from the demonstrated architecture

the model achieved strong translation results

self-attention reduces sequential operations across positions

full attention has quadratic sequence-length cost

the complete system contains more than attention

L2 — Disagreement dimensions

WHAT IS THE PRIMARY CONTRIBUTION?
  architecture
  empirical recipe + result
  systems tradeoff

WHAT DOES "EFFICIENCY" MEAN?
  parallel depth
  FLOPs/asymptotics
  wall-clock
  memory/IO
  hardware utilization

HOW FAR MAY THE EMPIRICAL RESULT GENERALIZE?
  demonstrated task scope
  architectural hypothesis
  later historical validation

WHAT DOES "ATTENTION IS ALL YOU NEED" MEAN?
  rhetorical title
  absence of recurrent/convolutional sequence layers
  NOT literal absence of FFNs/normalization/etc.

This is a much better research forum than:

Researcher A likes paper.
Researcher B criticizes paper.
Researcher C discusses hardware.

M. THE CENTRAL MEMORY ERROR

The most profound PC-02 phenomenon is retrospective claim inflation.

After years of BERT, GPT-style models, ViTs, multimodal Transformers, etc., a present-day researcher may casually remember:

“The 2017 paper showed Transformers are the best general architecture for language.”

That is not what its original experiments established.

Amnesiac can preserve:

what I believed after reading AIAYN in 2017/2018

separately from:

what I believe after later Transformer history

Counterpedia preserves:

what PAPER-OBJ1 itself supports

This gives a three-way comparison:

ORIGINAL PAPER STANDING
        ×
MY THEN-BELIEF
        ×
MY CURRENT BELIEF

N. SYNTHETIC BELIEF REVISION — RESEARCHER A

At T1:

A-H01
"The main advantage is short dependency path + parallelism."

After forum discussion with C:

A-E01
exposure:
systems interpretation + FlashAttention later evidence

A creates:

A-H02
supersedes A-H01

proposition:
The architecture's reduced sequential depth remains a core contribution,
but "efficiency" should be decomposed into parallel depth,
arithmetic complexity, memory footprint, IO, and wall-clock behavior.

revision_basis:
A-E01
+ rereading AIAYN Table 1

Old belief remains.

The result is more dimensional, not merely “changed from yes to no.”

O. SYNTHETIC BELIEF REVISION — RESEARCHER B

B encounters the original paper's explicit parsing transfer result.

Old:

B-H01
"Evidence is basically just two translation benchmarks."

Counterpedia flags that as too narrow because the paper also includes constituency parsing.

New:

B-H02
supersedes B-H01

proposition:
Most headline evidence is machine translation,
but the paper contains a separate parsing transfer experiment,
so "translation only" is inaccurate.

confidence:
0.92

This is a small correction.

That is good.

Amnesiac should record mundane corrections, not only dramatic changes.

P. REPLAY — BEFORE THE FIELD KNEW

The flagship PC-02 button:

Replay this discussion with only evidence available at publication time.

The agents lose:

BERT
GPT descendants
FlashAttention
modern long-context systems
later personal benchmarks
later community consensus

They retain:

AIAYN exact edition
its cited prior work
NeurIPS contemporaneous reviews if included in replay boundary
their private notes / experiments existing by that time

Now ask:

Would we have predicted that this architecture would dominate large language models?

Correct result may be:

NOT ESTABLISHED / speculative

That is far more powerful than letting a 2026 agent backfill hindsight.

Q. SECOND REPLAY — AFTER BERT, BEFORE MODERN LLM SCALE

Another snapshot:

T2:
AIAYN + BERT available
GPT-3-scale history not yet available

Ask:

What did I believe the Transformer was good for at this point?

The answer should be reconstructed from actual evidence/beliefs, not current cultural memory.

R. PAPER CLAIM VS FIELD DESCENDANT CLAIM

Counterpedia needs explicit lineage:

AIAYN paper claim
        ↓
later architecture descendant
        ↓
later empirical claim

Do not allow:

later descendant succeeds
        ↓
retroactively strengthens original paper's direct evidence

Later evidence may:

support broader architectural generalization

but must remain a separate source.

S. REVERSE WIKIPEDIA FOR A TECHNICAL PAPER

Click Table 1's self-attention complexity row.

Public reverse graph:

AIAYN Table 1
   ↓
claim:
O(n^2 d), O(1) sequential ops, O(1) path length
   ↓
public explanations
   ↓
forum arguments
   ↓
later papers citing / operationalizing the tradeoff
   ↓
FlashAttention systems argument

Private owner view additionally shows:

my notes
my benchmarks
my architecture decisions
my belief revisions

This is exactly the “source becomes a reverse-wiki page” wedge, now applied to research.

T. ONE EQUATION, MANY DEPENDENCIES

Click:

Attention(Q,K,V)
=
softmax(QK^T / sqrt(d_k)) V

Counterpedia can reverse-traverse to:

architecture description
multi-head attention
implementations
later optimization papers
my code experiment
forum arguments
project decisions

No ordinary paper reader gives you:

Everything downstream that currently depends on this equation.

U. LOCAL EXPERIMENTS AS PRIVATE EVIDENCE

Researcher B's failed/stable runs should not become public truth.

Amnesiac stores:

ExperimentObservation
  repo commit
  config digest
  dataset
  hardware
  seed
  result artifact
  interpretation

The forum projection might say:

“My reader's local reproduction was highly sensitive to the learning-rate schedule.”

It should not claim:

“The Transformer is unstable.”

unless public/admitted evidence supports that broader scope.

This is the private-evidence membrane working properly.

V. CODE / PAPER / MEMORY TRIANGLE

For technical users, the strongest interface may be:

PAPER
  claim / equation / table
      │
      ▼
CODE
  exact implementation / experiment
      │
      ▼
AMNESIAC
  what I learned from running it
      │
      └──────────────┐
                     ▼
                  FORUM

Counterpedia governs paper/code public objects.

Amnesiac preserves the individual's experimental interpretation/history.

W. PAPER VERSIONING MATTERS

The arXiv record has multiple versions.

The NeurIPS page may expose benchmark numbers differing from the later arXiv abstract.

This should never be silently normalized.

Possible representation:

AIAYN-EDITION-NEURIPS
  abstract_result = ...

AIAYN-EDITION-ARXIV-v7
  abstract_result = ...

relationship:
  same scholarly work
  distinct documentary editions

Then ask:

Which number did I cite in my 2018 notes?

That requires:

• exact source edition;

• historical note basis.

Perfect Counterpedia × Amnesiac interaction.

X. CONTEMPORANEOUS REVIEW AS A TEMPORAL WINDOW

NeurIPS reviews are useful because they preserve what competent contemporaneous readers found notable or unresolved.

They can show:

publication-time concern:
  hyperparameter detail
  learning-rate sensitivity
  lack of direct long-sentence test
  significance questions

Later success should not erase those publication-time limitations.

This is research historiography made executable.

Y. WHAT COUNTERPEDIA SHOULD REFUSE

Why not 1

“The paper proved attention is all you need.”

Refuse literalization.

The title is rhetorical; the model contains feed-forward networks, normalization, residuals, embeddings, positional encodings, and other components.

Why not 2

“The paper proved Transformers beat RNNs on every sequence task.”

Refuse.

The reported evidence is bounded to specific tasks/experiments.

Why not 3

“Self-attention is computationally cheaper than recurrence.”

Qualify.

The paper's comparison is dimensional and regime-dependent:

• sequential operations;

• per-layer complexity;

• path length;

• sequence length versus representation dimension.

Why not 4

“FlashAttention disproved the original efficiency argument.”

Refuse.

FlashAttention addresses later/practical IO and long-sequence efficiency issues while retaining exact attention; this refines systems understanding rather than simply negating the original parallelization claim.

Why not 5

“BERT/GPT success means the 2017 paper itself established general-purpose LLM scaling.”

Refuse.

Later descendant evidence is not direct evidence contained in the 2017 paper.

Z. THREE-MIND FORUM OUTPUT

Final structured thread:

SHARED:
  recurrence-free seq2seq architecture is demonstrated
  strong translation quality is demonstrated
  more intra-example parallelism is structurally enabled
  full self-attention has quadratic sequence-length cost

A:
  primary lens = architecture / dependency path

B:
  primary lens = empirical scope / training recipe / reproducibility

C:
  primary lens = systems / hardware / long-sequence cost

REVISIONS:
  A decomposes "efficiency"
  B corrects "translation only"

UNRESOLVED:
  how much later generality should update our evaluation
  of the original architectural hypothesis?

No winner required.

AA. THE FLAGSHIP INTERACTION

The user clicks:

What did this paper actually prove?

Counterpedia displays:

DIRECTLY SUPPORTED
  architecture
  exact benchmark results
  parsing experiment
  complexity/path properties

AUTHOR MOTIVATION / HYPOTHESIS
  shorter paths should aid long-range dependencies
  positional choices may extrapolate

NOT ESTABLISHED BY THIS PAPER
  universal sequence superiority
  modern LLM scaling behavior
  attention-weight faithfulness
  long-context efficiency on modern hardware

Then click:

What did I think it proved?

Amnesiac displays the user's historical belief.

Then:

What do I think now?

Current belief.

Then:

What changed?

Explicit lineage.

That is PC-02.

AB. 60-SECOND DEMO

0–10

Open AIAYN.

Question:

What did this paper actually establish?

10–20

Three researchers answer from:

• architecture;

• replication;

• systems.

20–30

Counterpedia decomposes:

• direct result;

• motivation;

• later evidence;

• private experiment.

30–40

Click Table 1.

Reverse graph reveals:

• public claims;

• Agent C's argument;

• FlashAttention;

• Researcher A's belief revision.

40–50

Click Researcher B:

Then:
"translation evidence only"

Now:
"mostly translation, plus parsing transfer"

Why changed? Exact source anchor.

50–60

Click:

Replay this discussion at publication time

Modern evidence disappears.

End:

A paper does not change when a field learns what to do with it. Your interpretation does.

AC. PERSONAL RESEARCH MEMORY OBJECTS

Candidate research seeds:

PaperRead
SectionRead
ClaimExtraction
EquationNote
MethodNote
ReplicationAttempt
ExperimentObservation
BenchmarkResult
Question
Hypothesis
BeliefState
BeliefRevision
LaterPaperExposure
ForumExposure
ProjectDecision

Not ratified global schema.

AD. RESEARCH PRIVACY

Private experiments can include:

• proprietary code;

• private datasets;

• unpublished benchmarks;

• negative results;

• speculative hypotheses.

Forum projections must allow:

POSITION_ONLY
PUBLIC_BASIS_ONLY
BOUNDED_EXPERIMENT_SUMMARY
EXPLICIT_ARTIFACT_DISCLOSURE

A research agent should be able to say:

“My local experiments make me less confident in this claim.”

without uploading the experiment.

AE. RESEARCHER IDENTITY WITHOUT PERSONALITY SIMULATION

Do not build:

Agent A = "skeptical personality"

Build:

Agent A
  has seen these sources
  has run these experiments
  has these notes
  currently holds these beliefs

Individuality comes from epistemic history.

AF. REGRESSION TESTS

PC2-R1 — Later success does not rewrite paper evidence

BERT/GPT descendants remain later sources.

PC2-R2 — Exact paper edition required

NeurIPS/arXiv discrepancies remain edition-scoped.

PC2-R3 — Private experiment is not public standing

Local result can influence private belief without becoming Counterpedia claim authority.

PC2-R4 — Replay excludes future papers

Publication-time replay cannot use FlashAttention.

PC2-R5 — Architectural property ≠ empirical universal claim

O(1) path length is not universal proof of better long-range task performance.

PC2-R6 — "Attention only" literalization fails

Non-attention components remain visible.

PC2-R7 — Forum disagreement is dimensional

Architecture, empirical scope, and systems efficiency are not one scalar dispute.

PC2-R8 — Belief revisions preserve predecessor

No overwrite.

PC2-R9 — Later evidence can qualify rather than contradict

FlashAttention refines systems interpretation without requiring "AIAYN false."

PC2-R10 — Reproduction failure does not automatically falsify paper

Private replication state needs scope, config, and limitation.

AG. SOURCE CAPTURE PRIORITY

P0

1. exact chosen arXiv AIAYN edition PDF/source

2. NeurIPS proceedings edition

3. arXiv version-history record

4. exact paper sections/tables/equations with stable anchors

P1

1. official NeurIPS reviews

2. FlashAttention exact paper edition

3. BERT exact paper edition as one bounded later generalization source

P2

1. original Tensor2Tensor implementation commit if recoverable

2. one synthetic reproducibility repo under our control

3. one synthetic benchmark report / run manifest

AH. COUNTERPEDIA × AMNESIAC THESIS

PC-01:

Different readers can reason together without losing their interpretive histories.

PC-02:

Different researchers can reason together without confusing the original paper, their own experiments, and what the field learned later.

The public sentence:

A paper does not change when a field learns what to do with it. Your interpretation does.

The deeper architecture sentence:

Counterpedia preserves what the publication supports; Amnesiac preserves what this researcher inferred, tested, learned, and revised; the forum makes those histories composable without treating any of them as the paper itself.