Repository navigation
Conversation
Freeze exploratory seed 890000 and topic document structure before generation. Catalog v5 adds Family 12, the collapse-resistant analogue of Cycle 6 general word spacing, instead of another lexical whitelist. Cover-greedy v4 is unchanged. Seed 880000 stays unseen. Co-authored-by: Byte271 <byte271@users.noreply.github.com>
Family 12 word-boundary newlines reach mean density 38.0 and collapsed intact/root 0.164. Durable detections stay 0/4 after whitespace collapse on this n=4 snapshot. Catalog v5 is not retuned. Seed 880000 remains the disjoint validation split. Co-authored-by: Byte271 <byte271@users.noreply.github.com>
Disjoint seed 880000 replicates D1: wrap density 37.875, collapsed intact/root 0.080, durable 0/4 after whitespace collapse, Cycle 6 spacing 4/4 after collapse. Do not retune v5 on 890000 or 880000. Confirmation seeds remain unseen. Co-authored-by: Byte271 <byte271@users.noreply.github.com>
Fresh generation still requires an active rule-construction seed. --samples-from rescoring uses exploratory-development admission so the frozen 890000 evidence can be replayed after catalog v5 is locked. Co-authored-by: Byte271 <byte271@users.noreply.github.com>
Cycle 8 independently used 890000 with a disjoint document-structure versus invisible-carrier corpus. Stage D evidence stays on the document-structure texts. 880000 is Cycle 7 Stage D validation and remains blocked from Cycle 8. Catalog v5 was not retuned. 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 |
![chatgpt-codex-connector[bot]](https://avatars.githubusercontent.com/in/1144995?s=60&v=4){kind=small}
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 0eacecac56
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".
| CYCLE7_STAGE_B_DURABLE_RULE_CATALOG_VERSION = "cycle7-durable-rule-catalog-v3" | ||
| CYCLE7_DURABLE_RULE_CATALOG_VERSION = "cycle7-durable-rule-catalog-v4" | ||
| CYCLE7_STAGE_C_DURABLE_RULE_CATALOG_VERSION = "cycle7-durable-rule-catalog-v4" | ||
| CYCLE7_DURABLE_RULE_CATALOG_VERSION = "cycle7-durable-rule-catalog-v5" |
There was a problem hiding this comment.
Advancing this alias to v5 also changes fuckmark/cycle7_stage_c_hf.py, which still imports CYCLE7_DURABLE_RULE_CATALOG_VERSION and evaluates through the current durable registry. Replaying the committed Stage C samples now includes the new word-boundary rules and reports a mean candidate count of 37.625 instead of the frozen v4 result of 4.75, so the Stage C runner can no longer reproduce its evidence. Update that runner to use a genuinely v4-pinned rule set/registry before advancing the current alias.
Useful? React with 👍 / 👎.
The live durable alias is catalog v5. Stage C --samples-from replay must keep enumerating v4 so frozen 870000 density stays 4.75 instead of picking up Family 12 word-boundary newlines. Stage B replay is pinned to v3 for the same reason. Catalog v5 evidence is unchanged. Co-authored-by: Byte271 <byte271@users.noreply.github.com>
2 participants
Summary
Rebases closed #98 onto current
mainafter PR #97 (Cycle 8 / product-visible invariance) landed.Cycle 7 Stages A–C showed that local semantically conservative English surface equivalence plus punctuation-newline layout stays in a four-site regime (means 4.0–4.75). Catalog v4 is not a Cycle 6 whitespace-collapse replacement.
This PR adds catalog v5 (Family 12: word-boundary newline). Cover-greedy v4 is unchanged. No new lexical whitelist. Catalog v5 remains PRODUCT_DISQUALIFIED for
release_transform_registry().Family 12 replaces a single ASCII space between two alphabetic words of length ≥2 with LF. That is the collapse-resistant analogue of Cycle 6 general word spacing.
whitespace-collapse-v1preserves LF. HYPOTHESIS: a wrap/reflow sanitizer would erase it.Seeds
document structureindependent checkCycle 8 independently assigned
890000with topicinvisible carrier development. Same pair seeds, disjoint text hashes. Stage D evidence is thedocument structurecorpus only. Do not mix corpora. Do not use Cycle 8 seeds900000/910000/920000for Cycle 7.Do not retune on 810000, 860000, 820000, or 870000.
Results (
PROMISING_DEVELOPMENT, not formal confirmation)Writeup:
docs/cycle7/stage-d-decision.mdws_collapseWMws_collapseWM890000880000Transformed UW: 0 detections on every arm. Combined-after-collapse remains worse than durable-only. Unlike Stage B, the detector reduction replicated on the held-out seed.
n=4+4 is not 7/192 or 0/192. Do not open Cycle 7 formal confirmation from this PR.
Review follow-up
CYCLE7_ACTIVE_EXPLORATORY_SEED_BASESto(),run_stage_d1could not replay frozen D1 evidence. Fresh generation still requiresassert_rule_construction_seed.--samples-fromuses exploratory-development admission.--samples-fromcount Family 12 (37.625 vs frozen 4.75). Stage C replay is now pinned to catalog v4 / registry v4. Stage B replay is pinned to v3. Live v5 and the Stage D artifacts are unchanged.Test plan
--samples-fromreplay of 890000 is admittedmain