Skip to content

Cycle 7 Stage A: collapse-resistant durable transforms - #95

Merged
cursor[bot] merged 3 commits into
mainfrom
research/cycle7-durable-transforms
Aug 25, 2026
Merged

cursor[bot] merged 3 commits into
mainfrom
research/cycle7-durable-transforms

Conversation

@byte271

@byte271 byte271 commented Aug 25, 2026

Copy link
Copy Markdown
Owner

Summary

  • Cycle 6 remains frozen at NONZERO_RESIDUAL 7/192. This PR does not retune it or reuse 720000/730000/760000/770000/780000.
  • Adds a Cycle 7 seed ledger, versioned whitespace-collapse-v1, a durable contraction/orthography family (including bounded it's copulas), quote-interior durable policy, and geometry instrumentation.
  • Stage A on fresh seed 810000 (4 watermarked + 4 unwatermarked): Cycle 6 spacing is 0/4 raw and 4/4 after whitespace collapse; the durable family is 4/4 on both arms. Decision: INSUFFICIENT_EVIDENCE as a Cycle 6 replacement. The collapse-survival channel works on contraction-rich fixtures.

Test plan

  • Unit tests for ledger, collapse sanitizer, durable rules, quote policy, instrumentation, Cycle 6 sanitizer boundary
  • Cycle 6 quote_safe_zrd ruleset hash unchanged
  • Fresh 810000 detector compare (raw + ws_collapse) vs Cycle 6 spacing on the same texts
  • CI green on this PR
  • CodeRabbit / review threads: fix real findings only

… evidence.

Cycle 6 remains frozen at NONZERO_RESIDUAL 7/192. Fresh seed 810000 shows Cycle 6 spacing dies under whitespace collapse, while the first durable family lacks density on GPT-2 TinyDev text.

Co-authored-by: Cursor <cursoragent@cursor.com>
@chatgpt-codex-connector

Copy link
Copy Markdown

You have reached your Codex usage limits for code reviews. You can see your limits in the Codex usage dashboard.
To continue using code reviews, you can upgrade your account or add credits to your account and enable them for code reviews in your settings.

@coderabbitai

coderabbitai Bot commented Aug 25, 2026 •

Copy link
Copy Markdown

Important

  • 🔍 Trigger review

This repository does not receive automatic reviews because it has fewer than 10 stars.

⚙️ Run configuration

Configuration used: defaults

Review profile: CHILL

Plan: Pro Plus

Run ID: ff35147c-edf7-44da-bb5c-94283d8dd6b8


Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

Comment @coderabbitai help to get the list of available commands.

byte271 and others added 2 commits August 25, 2026 15:46
… pass.

Co-authored-by: Cursor <cursoragent@cursor.com>
Stage A on the same seed 810000 texts remains INSUFFICIENT_EVIDENCE: durable detections stay 4/4 after whitespace collapse.

Co-authored-by: Cursor <cursoragent@cursor.com>
@cursor
cursor Bot merged commit af34188 into main Aug 25, 2026
41 checks passed
cursor Bot pushed a commit that referenced this pull request Aug 25, 2026
Retargeting #96 to main does not emit pull_request synchronize, so
the rebase SHA had only CodeRabbit. This push starts full CI.

Co-authored-by: Byte271 <byte271@users.noreply.github.com>
byte271 added a commit that referenced this pull request Aug 25, 2026
* Add Cycle 7 Stage B high-density durable catalog v3.

Extend the seed ledger with exploratory seed 860000 before any
Stage B generation. Add collapse-resistant families for sentence
newlines, bounded optional that, discourse commas, prenominal
hyphens, parenthetical adverbs, and coordinating commas.

Co-authored-by: Byte271 <byte271@users.noreply.github.com>

* Record Cycle 7 Stage B1 PROMISING_DEVELOPMENT on seed 860000.

Catalog v3 durable edits survive whitespace collapse and cut
collapse-surviving watermarked detections to 2/4 on the new
exploratory corpus. Freeze validation seed 820000 topic before
any disjoint look.

Co-authored-by: Byte271 <byte271@users.noreply.github.com>

* Record Stage B3 validation failure on seed 820000.

Density and collapse-surviving geometry replicate catalog v3, but
the seed-860000 durable 2/4 detector reduction does not. Overall
Stage B remains INSUFFICIENT_EVIDENCE. Do not retune on 820000.

Co-authored-by: Byte271 <byte271@users.noreply.github.com>

* Skip GPT-2 tokenizer Stage B tests when transformers is missing.

Default CI pytest does not install transformers; the Stage B channel
tokenization check now importorsksips instead of failing collection.

Co-authored-by: Byte271 <byte271@users.noreply.github.com>

* Add Cycle 7 Stage C durable catalog v4 and freeze seed 870000.

Clause-punctuation newlines and optional quantifier-of-determiner
pairs target higher natural site density. Ledger v3 reserves 880000
as unseen validation and leaves 830000/840000/850000 untouched.

Co-authored-by: Byte271 <byte271@users.noreply.github.com>

* Record Cycle 7 Stage C1 INSUFFICIENT_EVIDENCE on seed 870000.

Catalog v4 raised mean durable density only to 4.75 sites/sample.
Clause newlines fired; quantifier-of did not. Collapse-surviving
intact windows stayed ~71.5%. Seed 880000 remains unseen.

Co-authored-by: Byte271 <byte271@users.noreply.github.com>

* Record PR #95 squash-merge and retrigger Actions on main.

Retargeting #96 to main does not emit pull_request synchronize, so
the rebase SHA had only CodeRabbit. This push starts full CI.

Co-authored-by: Byte271 <byte271@users.noreply.github.com>

---------

Co-authored-by: Cursor Agent <cursoragent@cursor.com>
Co-authored-by: Byte271 <byte271@users.noreply.github.com>
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant