Transformer Systems Lab

Capstone: one systems claim you can defend

Turn the earlier steps into one claim. Predict which assumption fails first when the wording is broadened, see what each earlier step allows, then save the narrower version.

Step 8 of 8Predict, then checkSaved in this browser

What this step asks

The lab ends with one claim you can use, not a summary.

Each earlier step limits what one sentence may say. If the sentence leaves out its workload, baseline, threshold, data provenance, uncertainty or the result that would refute it, it has to narrow.

PredictionWhich assumption from the earlier steps breaks the broader wording first?
From the last stepThe evaluation step's scoped claim, counterexample and uncertainty.
What you checkOne row each for context use, KV cache and serving, speculative decoding, evaluation, and whether the claim holds outside this lab.
NextUse the scoped claim as a paper-reading checklist, an experiment or a request for a fix.

Step 8 of 8

Capstone systems claim

Checking saved workNo prediction yet
  1. 01Text to one update
  2. AtlasTokens and position
  3. AtlasAttention routing
  4. 02RoPE phase
  5. 03KV memory
  6. 04Long-context pressure
  7. 05Serving and decoding
  8. 06Speculative decoding
  9. 07Evaluation and falsification
  10. 08Capstone systems claim

From the evaluation step

This step works on one sentence: a claim about the whole system.

Checking this browser for the claim saved in the evaluation step.

Evaluation stepChecking this browser for saved work.
Claim from the evaluation stepChecking this browser for saved work.
Earlier steps savedChecking this browser for saved work.

The claim

A scoped claim about serving long contexts

A claim you can defend names the mechanism, the workload, the baseline, the metric and its threshold, where the test data came from, the uncertainty, and the result that would refute it.

Predict first

If the claim is worded more broadly, which assumption fails first?

Try it

Widen the claim one step, then see what the earlier steps allow.

Change how broad the claim is, the workload, the intervention, the evidence required, the uncertainty you accept, the dependency in focus and the failure to test. A fixed rule on this page then says whether the sentence holds, must narrow or is unsupported. The scores are illustrative.

Choose the assumption you expect to fail first.
Outline of this lesson
Claim draftSpeculative serving improves long-context QA.mechanism + workload + baseline + threshold + provenance + uncertainty + what would refute it
Broad wordingimproves QA
Limits from earlier stepswhich dependency narrows it first?
Scoped claimmay help this workload, with its limits stated
01Example

one claim sentence

02Question

a summary is not evidence

03Prediction

which assumption fails first

04Controls

scope, workload, standard, limits

05Result

what each earlier step lets the claim say

06What stays true

a claim keeps its assumptions attached

07Then

a paper checklist, an experiment or a request for a fix

Where this step sits

Every earlier step becomes a limit on one sentence.

Serving, speculative decoding and evaluation each add a condition the claim has to state, or a reason it has to narrow.

  1. 05Serving and decoding

    latency, cache memory, sampling

    previous
  2. 06Speculative decoding

    speedup separated from correctness

    previous
  3. 07Evaluation and falsification

    what a claim covers, and a counterexample

    previous
  4. 08Capstone systems claim

    one defensible scoped claim

    current