fix(ftip): study hybrid-attention cost and discovery tradeoffs [AGENT] - #177
Merged
Merged
Conversation
Add the reported same-SFT Kimi Linear and MLA comparison with separate timing regimes and model identities. Derive conditional active-cache, total attempt-cost, and finite discovery predictions from explicit assumptions informed by directly cited implementation practice. Replace metadata prerequisites in the direct comparison consumers with cost sensitivity and testable transfer arguments. Preserve stable addresses, homogeneous taxon structure, and the existing finite-domain and iso-quality conditions.
utensil
force-pushed
the
kepler/ftip-repair-r4-source-rows-20260908
branch
from
September 8, 2026 10:10
8321bc3 to
5dec9fd
Compare
utensil
changed the title
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
The hybrid-attention comparison previously described measurement prerequisites without showing readers what the reported results and FTIP theory imply. This change adds the published same-SFT Kimi Linear and MLA scores, separates the reported timing regimes, and develops conditional cache, total attempt-cost and discovery predictions.
The study derives the active-state ratio for a 3:1 recurrent/MLA mixture, then shows when lower attempt cost can offset lower per-attempt verified success. An illustrative budget buys 17 hybrid attempts versus 10 baseline attempts; the exact break-even condition and a generation-share sensitivity table explain which workloads could benefit. Assumptions are informed by cited official implementation practice and kept distinct from the paper's measurements. Stable tree addresses and the existing domain and iso-quality conditions are preserved.
Validation: six source cells and timing/source identities; 486 cache controls, 30 cost controls, 480 discovery controls; 22 existing prose tests; lint, full local site/render verification, eight browser routes and visually inspected equations/tables in the 236-page focused PDF. All 773 archived source blobs match the candidate and all saved source/HTML/PDF hashes pass. A fresh independent adversarial reviewer approved exact head
5dec9fd31e821726547052fc71fa1c55dc19da62after checking source claims, mathematical and interpretive boundaries, the complete diff, and final artifact hashes and visuals. Its one citation finding was fixed and rechecked; no unresolved findings remain. Fresh exact-head CI is required before merge.