code/cl8y-research
2
0
Fork 0
Code Issues 10 Pull requests Projects Releases Packages Wiki Activity

Editor passes: SEO/Voice, cycle-until-pass, and grounding against ingested sources #6

New issue
Open
opened 2026-08-31 05:33:12 +00:00 by PlasticDigits · 4 comments
PlasticDigits commented 2026-08-31 05:33:12 +00:00 (Migrated from gitlab.com)
Copy link

Summary

Editor quality: add missing SEO and Voice passes from #1; cycle editor steps until mechanical + grounding checks pass (budget-capped); persist one final document (intermediate pass files only in debug). Replace the incorrect "every H2 maps to a citation or is marked commentary" rule with an internal grounding check: every substantive claim in the post must match ingested data (or be dropped / labeled commentary that does not invent facts).

Gap: gaps/GAP_1788152435.md §2 (editor, outline citation map).

Current codebase

Issue #1 asked for 11 separate creates in order: SEO, grammar, voice, AI-tell, sentence-length, active/passive, clarity, structure, show-don't-tell, conciseness, open+close.

src/pipeline.rs EDITOR_PASSES is: grammar, punctuation, sentence-length, paragraph-rhythm, active-voice, clarity, structure, show-dont-tell, conciseness, open-close, ai-tell.

Missing: SEO, Voice. Extra: punctuation, paragraph-rhythm (keep as additional criteria — "11+").

Implementation does one editor create_once, then writes the same blob to all 11 passes/*.md. Mechanical lint_markdown runs once at emit.

Numeric grounding: unsourced() flags numbers not in NumericClaims (years 20xx and ≤2-digit numbers exempt). There is no check that prose matches ingested repo events, indexer stats, telegram (non-authoritative), tokenlist, or Tavily snippets. Outline H2↔citation is not enforced (and should not be the product rule).

Happy-week replicate.json pads the same paragraph to hit 2000 words — does not prove live editor quality.

Why this is needed

SEO/voice were explicit in #1. Dumping 11 identical files wastes disk and review. Cycling until lint+grounding pass avoids publishing dirty or hallucinated recaps. Posts must not invent shipping, volumes, or prices that were never ingested.

Constraints / guardrails

  1. Create-once per attempt still applies: each cycle iteration is a new create, never a retry of the same create. Prefer one create per cycle with a rubric listing all criteria, then mechanical+grounding fail → new create with the failed report (counts against weekly budget). Do not revive 11 creates per week unless budget is raised explicitly.
  2. Weekly budget remains a hard cap (REPLICATE_WEEKLY_CREATE_BUDGET). Exhaustion → fail the week (or emit short announcement if --short / thin-week), no silent dirty publish.
  3. Debug mode (CL8Y_RESEARCH_DEBUG=1 or --debug): write intermediate cycle artifacts under runs/<week>/passes/. Default: one post.mdx (+ report).
  4. Grounding: claims about volumes, prices, fees, addresses, "we shipped/merged X", dates must appear in ingested sources (with citation). Commentary/opinion allowed without a number but must not fabricate product behavior. Telegram/Tavily are not onchain proof.
  5. Mechanical lint still fail-closed (dashes, curly quotes, banned CTAs, unofficial addresses, fee-tier invention).
  6. Voice: CL8Y-web/blog_gen/SKILL.md. SEO: unique title/description bounds already in invariants.rs; slug; no keyword stuffing; href allowlist including first-party hosts (yieldomega.com if we add it).
  7. Do not pad short weeks with generic DeFi filler (--short / 800 word cap).

Relevant files

Path Role
src/pipeline.rs Editor bundle, emit, word counts
src/lint.rs Mechanical lint
src/numeric.rs Number vs claims
src/mdx.rs Sanitize, frontmatter bounds
src/invariants.rs Title/description bounds, banned phrases
docs/invariants.md Prose rules
CL8Y-web/blog_gen/SKILL.md Voice
tests/issue_plan.rs Lint / injection

Recommended direction

  1. Typed rubric list: #1 eleven + punctuation + paragraph-rhythm (document the set in invariants.rs).
  2. ground_post(body, &Collection) -> Vec<Violation>: numbers (existing), plus checks (indexer figures, known repo event strings, listed token names vs unofficial tickers). Fail closed on unsourced quantities and on sentences that assert "merged/shipped" when repo sources are empty.
  3. Loop: editor create → lint + ground → if fail and budget remains, next create with violation list; else emit or error.
  4. Default artifacts: post.mdx, report.md (cycle count, prediction ids, remaining violations if any).

Acceptance criteria

  • SEO and Voice are in the rubric and covered by tests (SEO: title/description bounds, no stuffed keyword blob; Voice: banned promotional/home copy).
  • Default run writes one post file, not 11 duplicate passes/*.md.
  • --debug / env writes intermediates.
  • Dirty lint fails; cycle with fixture predictor that first returns dirty then clean → one published clean body, create_count ≥ 2, under budget.
  • Body asserting volume 99999 when sources only have 12000 fails even if lint-clean.
  • Body asserting "we shipped malware-repo" when only yieldomega events exist fails.
  • Commentary without fake numbers can pass.
  • Budget exhaustion does not publish dirty MDX.

Test plan — functional paths

  1. Clean fixture — one emit, no passes/ unless debug.
  2. Debug — passes/ present.
  3. SEO fail — title 3 chars → fail or cycle to valid title (fixture sequence).
  4. Voice fail — "Buy CL8Y" / "the future of DeFi" → fail.
  5. Cycle dirty→clean — fixture map by create index.
  6. Grounding numbers — existing unsourced test.
  7. Grounding shipping — empty repos + "we shipped the bridge indexer API" → fail.
  8. Short week — --short not padded past 800.
  9. Budget — max 1 remaining create, first editor dirty → fail closed, no publish.

Test plan — attack, hack, and abuse

  1. Injection in editor output — script / lookalike href / unofficial address still fail after cycles.
  2. Infinite cycle — budget/max-iterations cap.
  3. Grounding bypass — "as the README describes 99999 volume" still unsourced.
  4. Telegram as proof — "indexer volume is 99999 because chat said so" fails (telegram not numeric source).
  5. Tavily as proof (when that issue lands) — web snippet cannot invent DEX volume.
  6. Secret in cycle artifacts — debug files also run assert_clean_artifact.

Verification criteria

  1. Fixture week: single post.mdx, report lists rubric + cycle count.
  2. Reviewer reads invariants and sees SEO + Voice + grounding rule (not H2-citation fiction).
  3. Injection and unsourced tests still fail closed.
## Summary Editor quality: add missing **SEO** and **Voice** passes from [#1](https://gitlab.com/PlasticDigits/cl8y-research/-/issues/1); **cycle** editor steps until mechanical + grounding checks pass (budget-capped); persist **one** final document (intermediate pass files **only in debug**). Replace the incorrect "every H2 maps to a citation or is marked commentary" rule with an **internal grounding check**: every substantive claim in the post must match **ingested** data (or be dropped / labeled commentary that does not invent facts). Gap: [`gaps/GAP_1788152435.md`](https://gitlab.com/PlasticDigits/cl8y-research/-/blob/main/gaps/GAP_1788152435.md) §2 (editor, outline citation map). --- ## Current codebase Issue #1 asked for **11 separate creates** in order: SEO, grammar, voice, AI-tell, sentence-length, active/passive, clarity, structure, show-don't-tell, conciseness, open+close. `src/pipeline.rs` `EDITOR_PASSES` is: `grammar`, `punctuation`, `sentence-length`, `paragraph-rhythm`, `active-voice`, `clarity`, `structure`, `show-dont-tell`, `conciseness`, `open-close`, `ai-tell`. **Missing:** SEO, Voice. **Extra:** punctuation, paragraph-rhythm (keep as additional criteria — "11+"). Implementation does **one** editor `create_once`, then writes the **same blob** to all 11 `passes/*.md`. Mechanical `lint_markdown` runs once at emit. Numeric grounding: `unsourced()` flags numbers not in `NumericClaim`s (years `20xx` and ≤2-digit numbers exempt). There is **no** check that prose matches ingested repo events, indexer stats, telegram (non-authoritative), tokenlist, or Tavily snippets. Outline H2↔citation is **not** enforced (and should not be the product rule). Happy-week `replicate.json` pads the same paragraph to hit 2000 words — does not prove live editor quality. --- ## Why this is needed SEO/voice were explicit in #1. Dumping 11 identical files wastes disk and review. Cycling until lint+grounding pass avoids publishing dirty or hallucinated recaps. Posts must not invent shipping, volumes, or prices that were never ingested. --- ## Constraints / guardrails 1. **Create-once per attempt** still applies: each cycle iteration is a new create, never a retry of the same create. Prefer **one create per cycle** with a rubric listing all criteria, then mechanical+grounding fail → **new** create with the failed report (counts against weekly budget). Do not revive 11 creates per week unless budget is raised explicitly. 2. Weekly budget remains a hard cap (`REPLICATE_WEEKLY_CREATE_BUDGET`). Exhaustion → fail the week (or emit short announcement if `--short` / thin-week), no silent dirty publish. 3. **Debug mode** (`CL8Y_RESEARCH_DEBUG=1` or `--debug`): write intermediate cycle artifacts under `runs/<week>/passes/`. Default: **one** `post.mdx` (+ report). 4. Grounding: claims about volumes, prices, fees, addresses, "we shipped/merged X", dates must appear in ingested `sources` (with citation). Commentary/opinion allowed without a number but **must not** fabricate product behavior. Telegram/Tavily are not onchain proof. 5. Mechanical lint still fail-closed (dashes, curly quotes, banned CTAs, unofficial addresses, fee-tier invention). 6. Voice: `CL8Y-web/blog_gen/SKILL.md`. SEO: unique title/description bounds already in `invariants.rs`; slug; no keyword stuffing; href allowlist including first-party hosts (`yieldomega.com` if we add it). 7. Do not pad short weeks with generic DeFi filler (`--short` / 800 word cap). --- ## Relevant files | Path | Role | | --- | --- | | `src/pipeline.rs` | Editor bundle, emit, word counts | | `src/lint.rs` | Mechanical lint | | `src/numeric.rs` | Number vs claims | | `src/mdx.rs` | Sanitize, frontmatter bounds | | `src/invariants.rs` | Title/description bounds, banned phrases | | `docs/invariants.md` | Prose rules | | `CL8Y-web/blog_gen/SKILL.md` | Voice | | `tests/issue_plan.rs` | Lint / injection | --- ## Recommended direction 1. Typed rubric list: #1 eleven + punctuation + paragraph-rhythm (document the set in `invariants.rs`). 2. `ground_post(body, &Collection) -> Vec<Violation>`: numbers (existing), plus checks (indexer figures, known repo event strings, listed token names vs unofficial tickers). Fail closed on unsourced quantities and on sentences that assert "merged/shipped" when repo sources are empty. 3. Loop: editor create → lint + ground → if fail and budget remains, next create with violation list; else emit or error. 4. Default artifacts: `post.mdx`, `report.md` (cycle count, prediction ids, remaining violations if any). --- ## Acceptance criteria - [ ] SEO and Voice are in the rubric and covered by tests (SEO: title/description bounds, no stuffed keyword blob; Voice: banned promotional/home copy). - [ ] Default run writes **one** post file, not 11 duplicate `passes/*.md`. - [ ] `--debug` / env writes intermediates. - [ ] Dirty lint fails; cycle with fixture predictor that first returns dirty then clean → one published clean body, create_count ≥ 2, under budget. - [ ] Body asserting volume `99999` when sources only have `12000` fails even if lint-clean. - [ ] Body asserting "we shipped malware-repo" when only yieldomega events exist fails. - [ ] Commentary without fake numbers can pass. - [ ] Budget exhaustion does not publish dirty MDX. --- ## Test plan — functional paths 1. **Clean fixture** — one emit, no `passes/` unless debug. 2. **Debug** — `passes/` present. 3. **SEO fail** — title 3 chars → fail or cycle to valid title (fixture sequence). 4. **Voice fail** — "Buy CL8Y" / "the future of DeFi" → fail. 5. **Cycle dirty→clean** — fixture map by create index. 6. **Grounding numbers** — existing unsourced test. 7. **Grounding shipping** — empty repos + "we shipped the bridge indexer API" → fail. 8. **Short week** — `--short` not padded past 800. 9. **Budget** — max 1 remaining create, first editor dirty → fail closed, no publish. --- ## Test plan — attack, hack, and abuse 1. **Injection in editor output** — script / lookalike href / unofficial address still fail after cycles. 2. **Infinite cycle** — budget/max-iterations cap. 3. **Grounding bypass** — "as the README describes 99999 volume" still unsourced. 4. **Telegram as proof** — "indexer volume is 99999 because chat said so" fails (telegram not numeric source). 5. **Tavily as proof** (when that issue lands) — web snippet cannot invent DEX volume. 6. **Secret in cycle artifacts** — debug files also run `assert_clean_artifact`. --- ## Verification criteria 1. Fixture week: single `post.mdx`, report lists rubric + cycle count. 2. Reviewer reads `invariants` and sees SEO + Voice + grounding rule (not H2-citation fiction). 3. Injection and unsourced tests still fail closed.
PlasticDigits commented 2026-08-31 05:35:05 +00:00 (Migrated from gitlab.com)
Copy link

marked as related to #5

marked as related to #5
PlasticDigits commented 2026-08-31 05:35:06 +00:00 (Migrated from gitlab.com)
Copy link

marked as related to #7

marked as related to #7
PlasticDigits commented 2026-08-31 05:35:06 +00:00 (Migrated from gitlab.com)
Copy link

marked as related to #8

marked as related to #8
PlasticDigits commented 2026-09-02 02:18:49 +00:00 (Migrated from gitlab.com)
Copy link

mentioned in issue #12

mentioned in issue #12
Sign in to join this conversation.
No Branch/Tag specified
Branches Tags
main
No results found.
Labels
Clear labels
No items
No labels
Milestone
Clear milestone
No items
No milestone
Projects
Clear projects
No items
No project
Assignees
Clear assignees
No assignees
1 participant
Notifications
Due date
The due date is invalid or out of range. Please use the format "yyyy-mm-dd".

No due date set.

Dependencies

No dependencies set.

Reference
code/cl8y-research#6
No description provided.