What we shipped.

After #1699 merged, its dispatched run put green backend / frontend / infrastructure check runs on changelog PR #1702's head commit. Once the ruleset required those names,...

Iceland Photo Map (proj_ab5ead538b5947e7) has 3 held drafts with 21, 12 and 4 blocking claims. Almost none are about its own product. They are claims about apps it doesn't com...

The goal is to require backend, frontend and infrastructure on main. Today gh pr merge --auto merges instantly because the "Protect Main" ruleset requires a PR but no...

post_37b7fa16b58641e9 (Emcognito) is still held after #1695 on:

After #1693, Emcognito's replacement draft post_8ee09bd403f4404c was still held. A fresh re-verify plus the owner claim strip ran in prod on 2026-09-28. **8 of its 9 blocking...

Emcognito's held draft post_8ee09bd403f4404c was blocked by 27 claims. Each one is a plan fact that Emcognito's capability file states.

After #1689, Tovin's and FolioInbox's held drafts still blocked on own-product claims that their pricing page, Content Brief and capability file all state (e.g. "Folio has n...

"Posts not generating" (2026-09-28): schedules fired and 152 drafts were created in 7 days, but the truth checker held 37 (24%) in review_required. 34 of those involved `curre...

Sendant post_bd801e12ca4d4052 is parked in publish_failed (redrive_exhausted). The destination 422'd matched_claim: "onion routing" on field: meta_description ("…o...

The outline DeepSeek canary started 2026-09-25 ~19:11 UTC (SSM deepseek_canary_tasks now has outline next to post_metadata). Its guardrail is publish rate + truth first-pa...

No prod writes and no model calls.

> Built on #1679 (fix/truth-own-business-evidence). This branch was cut from it. #1679 was merged into main before this PR opened, so the base is main and the diff con...

Two held drafts, post_57bcd3b58653449a (tovin.io) and post_e2121d6e2be04935 (icelandphotomap.com), stayed held after their flagged sentences were edited. Each re-check flagged t...

Tovin post post_57bcd3b58653449a was held on a cell in its own column, "Tovin — Pricing Model: …". The extractor had not restated that cell. Under #1587 a cell is first-party...

vectraseo-prod-stranded-posts went to ALARM on 2026-09-25 for two held drafts (post_57bcd3b58653449a Tovin, post_e2121d6e2be04935 Iceland Photo Map). A paid manual strip a...

vectraseo-prod-pipeline-loss-rate went into ALARM at 2026-09-25 01:33 UTC (62.5 / 72.2 / 42.9%). Of the 20 posts it counted as lost from 09-22 to 09-25, 15 were published...

The per-day rollback rule (PRD §10) breached 18 of 19 pre-canary days on unchanged operator traffic (~6 truth-checked posts a day). The post_metadata canary was going to rev...

Three EchoThread drafts on one topic (post_44024259, post_dc601cb6, post_0e75f61c) sat at review_required, the only held posts in prod, each at **1 of 3 external sources...

A never-published post that used up its publish_reject_repair_attempts stayed at publish_failed permanently, even after a later repair fix could clear it. `post_0e3fccfcb9f6...

"According to research from X" linked only to X's home page cites nobody. EchoThread post_dc601cb63efd4c0c credited a "chilling effect" finding to cyber.harvard.edu/ and new...

Gemini returned 503 "high demand" after 5 retries, the Anthropic fallback timed out, and OpenAI returned 429 insufficient_quota. During that window the Publish-Readiness ladde...

A draft parked at review_required had no automatic exit unless a later claim-guard change happened to clear it. EchoThread post_44024259deae4f97 shows the full failure. Its...

post_44024259deae4f97 (EchoThread, gap gap_41fef2cf163e4ec8) is parked at review_required because it cites 1 of 3 required sources.

_automatic_recovery_jobs put every recovery-marked post, and every post for which sweep_action returned *any* action including "stranded", into recovery_ids. `process_du...

post_44024259deae4f97 (EchoThread, "Hostile Comment Threads") was held on five product_capability claims such as *"Observed Behavior — Tier 1: Robust Debate: Robust debate c...

EchoThread post_8dc207b034f542ba has been parked at review_required since 2026-09-23. Its fact-check is verified (truth 100, 19/19 claims supported), and metadata, lint and...

The "It is free on every plan ." claims on the EchoThread drafts did not come from inserted text. The body is correct: <a href="…/pricing/">free on every plan</a>. (read-only...

Tempkey's configured /sitemap.xml listed only marketing pages, and its blog sitemap was declared only in robots.txt. As a result, the GSC index check flagged 118 published p...

The held-draft button (#1646) was offered on drafts that also fail content quality. Stripping sentences only removes text, so it can never lift a draft that is already below the...

Content quality's uncited-claim check (_sentences_with_unsourced_claims, backend/app/services/blog_quality.py; trigger _HIGH_RISK_CLAIM_RE at line 55) flags any sentence c...

The truth layer merged an FAQ question heading into the claim that follows it. The merged claim can't be matched to any source, so it comes back unverifiable and holds the post.

On 2026-09-23 UTC, Tovin, DNSCove and Tempkey were each flagged Scheduled post overdue even though each had published a post that UTC day. Their scheduled full_seo_run job...

A draft held by the truth gate reached review_required and then had nowhere to go. The detail page hid every action and said "No manual flag review is needed", but automatic r...

test_typesafe_jev::test_sc_004 failed intermittently in full-suite runs (passed alone every time).

1. Before stripping, it runs partition_protected over claim_strip.strip_target_claims. That is the stripper's own verdict filter (unsupported, plus unverifiable unde...

Tempkey's operator post keeps getting held on claims that are true and about Tempkey itself: "public REST API covering grants, extension, revocation…", "administrative sign-in i...

A capability contract's preferred_wording put a writer-directed sentence into Tempkey post bodies, and publish lint let it through:

The automatic-recovery strip step from #1638 deletes every sentence that carries an unconfirmed claim. That includes the business's own promotion. In a prod re-drive it was abou...

Automatic recovery publishes only when every claim in a fresh truth report is supported. Drafts held by a few claims no source could confirm (Tempkey, Marry Math) stayed held...

Commits QA reports that were sitting uncommitted in the main checkout. Docs only, no code or config changes.

The DeepSeek canary now polices itself (PRD section 10, new REQ-canary-auto).

PostRepository.conditional_advance_phase builds its own update_item call and never converts floats; BaseRepository.update does. Any payload carrying a fresh verify_post...

DeepSeek phase 3 (docs/prd/cost-optimized-content-pipeline.md, DL-003 "all ungrounded tasks"). The scheduled writer path can now be routed. Before this, routing saved almost not...

Corrects two numbers in the docs and a docstring that #1630 merged. They were found when the terminal-outcome rows were recounted by archive reason.

The current-capability truth gate shipped in 084ac618 on 2026-09-22. Since then, a post held on an unconfirmed third-party plan or file-format claim had no repair:

Production caps recovery at one dispatch per 15-minute sweep (automatic_recovery_batch_size=1). After #1623 shipped, FolioInbox post_1d1c98f23ad24610 was a release candidate...

Follow-up to #1624.

Phase 2 of docs/prd/cost-optimized-content-pipeline.md (phase 1: #1621). Nothing changes in production at merge: no task is routed, the canary lists are empty, off-peak de...

Once deployed, the recovery sweep releases the stuck post on its first run. Unresolved backlog posts become eligible for retry #1, dispatched at automatic_recovery_batch_size...

Scheduled posts are parked on truth-gate holds with 1-3 residual claims. One source: the product-capability backstop read a moderation row labelled Tier 1: Robust Debate as a...

P0 core of the cost-optimized content pipeline (docs/prd/cost-optimized-content-pipeline.md, tdd/cost-optimized-content-pipeline.tdd.yaml).

Clarified spec and red-phase TDD artifact for moving model spend to the cheapest provider that passes each task's quality bar.

Held-post recovery could hide the strongest evidence when reserving space for supplemental sources, and its grounded-response formatter could discard source URLs. Rank passages...

Correction hints and fresh verification currently spend the same eight-request Jev allowance. Two recovery canaries exhausted it and left 14 later source-backed claim records ea...

Held-post recovery skipped repeated current assertions, missed fresh product-contract evidence, and could introduce failed citation links. It now binds repeated prose to individ...

A truth report can have legacy status verified while still containing generic unverifiable claims. Recovery requires supported evidence for every claim, so checking only that...

Held posts could skip AI repair despite unresolved claims, exhaust Gemini's old allowance, or spend extra resolution calls solely because Jev abstained. This change fixes that r...

A production scheduler recheck read Tempkey before automatic recovery claimed it, then moved the already-running recovery back to gate_check. Recovery correctly refused to ove...

Held posts could not verify their own product claims when a legacy project stored a bare domain: first-party fetching rejected it and internal links were classified as external....

Held drafts could spend the six-call recovery allowance re-investigating saved findings, leaving no room to correct them. Recovery now corrects exact, unique flagged sentences f...

Held drafts could lose an extracted citation when the literal capability backstop replaced the claim; the first automatic recovery trial also spent its six Gemini calls before r...

Held posts currently leave the owner with manual flag reviews and no automatic next step. This adds one bounded Gemini/Jev investigation per post and policy: retrieve relevant s...

Saving post review feedback returned HTTP 500 because the API execution role lacked dynamodb:ConditionCheckItem. The transaction checks both the current post snapshot and proj...

The post review panel asked users to judge “valid” or “mistaken” flags without explaining what was uncertain. It now asks “Does this sentence need fixing?” with visible Needs fi...

The evaluation page offered review links for every project, including projects with no problematic posts, and mixed replaced originals with unresolved work. It now shows counted...

Held and automatically archived posts currently expose automated flags but provide no way to record whether those flags were valid. This adds an admin-only Problematic filter an...

Admins currently have no persistent indication that the Jev evidence-selection comparison needs label review or has produced results. Add an admin-only status page, sidebar atte...

Jev made no accepted decisions in the observed production cycles, so default traffic was paying for Jev and still falling back to Gemini. Make all three stages default to off,...

Fixes #1587.

A production tutorial was marked truth-verified despite an obsolete free-plan allowance and a CSV import promise that extraction omitted. Current product promises now retain tab...

Citation verification, product-capability checks, and evidence selection currently send every semantic decision to Gemini. This adds TypeSafe Jev as a selective first pass: conf...

Search Console settings showed “Synced” after a refresh was skipped because property access was blocked. Only completed, non-skipped refreshes now show that success message; the...

Hosted blog publishes ignored the adapter's seo_title and meta_description, and those pages lacked the site's GA4 tag and a clear next step. Honor the search metadata and in...

The signup dashboard omitted anonymous audits entirely and scanned roughly 347,000 unrelated shared-table records to find 25 accounts. Existing users could appear missing behind...

Visitors who request a sign-in link after a free audit currently lose the path to monitoring, while the owner cannot distinguish requests from completed sign-ins. Search Console...

run_phase_advance re-raises so SQS redelivers, and the comment above that raise accepted the cost explicitly:

Two changes to the API's burst posture, both from live evidence gathered while re-checking the 2026-09-19 scanner incident.

Follow-up to #1555. The admin inbox (GET /api/v1/admin/leads) was the last paginated endpoint still handing out a raw DynamoDB sort key as its cursor.

vectraseo-prod-api-gateway-5xx fired 2026-09-19 23:33 UTC. The