Repository navigation
Restore exact user-visible invariance and start Cycle 8 carrier research - #97
Conversation
The public CLI no longer applies contractions or other visible edits. release_transform_registry is now the empty product-visible-invariance registry and fail-closes to the original text. Historical visible-edit catalogs remain replayable under explicit names. Cycle 8 starts invisible-carrier research; U+034F is a durable-track sanitizer candidate, not a release promotion. Co-authored-by: Byte271 <byte271@users.noreply.github.com>
|
Important
This repository does not receive automatic reviews because it has fewer than 10 stars. ⚙️ Run configurationConfiguration used: defaults Review profile: CHILL Plan: Pro Plus Run ID: 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. Comment |
The installed-CLI GitHub check still expected contraction output after release-cli-v4 fail-closed. Align that assertion with unchanged text, and instrument detector-blind U+034F space-carrier arms without promoting any carrier to the product registry. Co-authored-by: Byte271 <byte271@users.noreply.github.com>
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 2ef9f8f891
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
Keep the product registry empty. Detector-blind space-carrier arms pass visible projection on generated ASCII pairs. U+034F reduces watermarked detections without unwatermarked inflation on these tiny corpora; U+200C is undone by Cf-strip and fails Chromium pixel equality. Co-authored-by: Byte271 <byte271@users.noreply.github.com>
load_product_contract uses the embedded v1 payload when specs/ is absent from an installed wheel. Cycle 8 writers require a GPT-2 encoder before replacing verified tokenizer artifacts. Chromium HTML evidence encodes less-than and greater-than so an ASCII script terminator cannot close the inline script. Co-authored-by: Byte271 <byte271@users.noreply.github.com>
Pinned specs/cycle8 writers still fail closed without a GPT-2 encoder. The detector harness writes fixture-compare with encoder unavailable so GitHub Cycle8 Carrier Detector can score seed 890000 without tiktoken. Co-authored-by: Byte271 <byte271@users.noreply.github.com>
Why
The public product path had drifted from the original contract.
release_transform_registry()was the default contraction catalog, soFuckMarkvisibly rewrote user text (I do not agree.→I don't agree.). Semantic equivalence is not the product. Exact user-visible projection is Priority Zero.Live GitHub at start of this work:
maind7eeb0f(PR #96 merged), no open PRs, tagv0.2.0.Contract repair
specs/fuckmark-user-visible-invariance-v1.contract.jsonanddocs/product-contract.md.USER_VISIBLE_TEXT_CHANGED. Detector score cannot override it.release_transform_registry()is the empty product registry. CLIrelease-cli-v4fail-closes to the original text.GitHub CI's installed-CLI end-to-end still expected contraction output after the product path was emptied. That leftover assertion now expects unchanged visible text.
Historical reclassification (not rewritten)
NONZERO_RESIDUAL7/192. PRODUCT_DISQUALIFIED (visible U+0020 runs).REJECTED.Cycle 8 research (not a release promotion)
U+034F is not product-authorized. The public CLI still authorizes zero carriers.
890000, replication900000, validation910000, secondary920000. Confirmation reserves830000/840000/850000untouched. Cycle 7880000still unseen.prescreenshots: U+034F and U+FE00 VERIFIED pixel-equal; U+200C REJECTED (PNG bytes differ).0.5570987654320988, 4 watermarked + 4 unwatermarked ASCII pairs per seed:U+200C also drops raw detections but Cf-strip restores identity detections (3/4, 4/4, 3/4). U+034F x8 can overflow GPT-2's 1024-token context; prefer x1.
This is
PROMISING_DEVELOPMENT/HYPOTHESISon tiny corpora. It is not Cycle 6-scale confirmation and is not a CLI promotion.Codex P2 follow-up
Addressed the three Codex review comments on
12445a5(originally reviewed2ef9f8f):load_product_contract()returns the embedded v1 payload whenspecs/is absent from an installed wheel. Checkout copies of the contract file must still match the embedded payload.specs/cycle8tokenizer JSON callrequire_gpt2_encoder()and fail closed. The detector harness (cycle8_hf) still scores when tiktoken is absent, writingencoder: unavailableinto artifact fixture-compare instead of crashing GitHub Cycle8 Carrier Detector.Tests
Full
pytestpassed locally after the Cycle8 CI fix. Cycle 8 evidence tests still pin the committed artifacts. The public CLI still leavesI do not agree.andWe cannot continue.unchanged.