# Independent editorial and interface review

The main agent reviews the generated material alongside the serial model passes. These checks do not count as additional Ralph iterations.

## Initial source and scoring review

- Inspected the older Leadership Principles poster, German latency/action scenarios, authentication failover, outage response, logging discovery, failed search, application protection, DR objectives, replication and backup choice pages, and paired learning-preference statements directly from the photographs.
- Confirmed that effectiveness ratings use five levels and must not be relabeled as agreement. Only endpoint labels are visible on the rating controls. Middle labels are teaching aids.
- Confirmed that some later images show multiple questions and cropped lower answer options. Complete reconstruction means preserving the visible corpus and documenting missing text, rather than inventing unseen options or treating numbered scenario gaps as supplied content.
- Sampled batch 05 against `image-1788636195773.jpg` and `image-1788636176984.jpg`. The original cloud/data-center best-practice and custom inspection-server actions are accurately represented. The explanations distinguish transferable security objectives from unchanged implementation and identify the single inline inspection server's reliability/operating risks.
- Reviewed the unspecified firewall item: the conditional rating correctly avoids assuming that the unnamed firewall is either only a port filter or a fully configured WAF. Inspection, rule quality, placement and TLS visibility matter.
- Changed test-mode headings and question-map labels to neutral scenario labels because some descriptive study titles disclosed the intended judgment before answering.
- Added an interactive RTO/RPO recovery-budget illustration with explicit sequential-step, successful-backup and no-log-replay assumptions. On 2026-09-05, opened [AWS recovery-objective guidance](https://docs.aws.amazon.com/wellarchitected/latest/framework/rel_planning_for_recovery_objective_defined_recovery.html) and [AWS business-continuity guidance](https://docs.aws.amazon.com/whitepapers/latest/disaster-recovery-workloads-on-aws/business-continuity-plan-bcp.html). The arithmetic and example timings are original teaching analysis; they do not certify actual recovery performance.

## Technical source checks during chapter generation

- On 2026-09-06 UTC, independently reopened [Amazon's current Leadership Principles](https://www.amazon.jobs/content/en/our-workplace/leadership-principles) and verified the 16 names. The final validator now checks the complete set with harmless apostrophe/case normalization. Principle-specific official explanations are included in the editorial handoff for richer leadership treatment.

- On 2026-09-05 UTC, independently opened the current [regional NAT gateway documentation](https://docs.aws.amazon.com/vpc/latest/userguide/nat-gateways-regional.html). Confirmed that the chapter must distinguish regional and zonal modes and retain the regional mode's private-NAT limitation. Added a handoff so later cost, recovery and register passes do not contradict this current behavior.
- Independently opened [CloudFront origin failover](https://docs.aws.amazon.com/AmazonCloudFront/latest/DeveloperGuide/high_availability_origin_failover.html) and [cache expiration](https://docs.aws.amazon.com/AmazonCloudFront/latest/DeveloperGuide/Expiration.html). Confirmed the method-specific failover boundary and the effect of positive minimum TTL on private/no-store responses. These are relevant to the quiz's availability and session-protection reasoning.
- Reviewed the five pass-09 chapter drafts directly. Checked the 10,000-request cache example (2,800 origin requests, 74 ms illustrative mean), the 72-hour DNS preparation timeline, the firewall's positive and negative cases, 10/20-segment mesh arithmetic (45/190 links), the bandwidth-delay product example (8 MB), and the POST/302/GET sequence. Assumptions and the distinction between illustrative calculations and measured guarantees are explicit. The chapters connect to exact reconstructed question IDs and distinguish protocol concepts from product-specific mechanisms.
- Reviewed pass-10's five chapter drafts. Checked the target-loss timeline and surviving-capacity arithmetic, JWT revocation-window example, 172.8-million daily session-read calculation, Amdahl and database-connection limits, synchronous image-result requirement, independent-release boundaries, and the 390-to-120-second startup budget. Warm-pool depletion and stale state are explicit. The text distinguishes fail-open routing from successful service, DynamoDB TTL cleanup from enforced session expiry, and instance affinity from durable authentication state.
- On 2026-09-06 UTC, independently opened current [DynamoDB global-table design guidance](https://docs.aws.amazon.com/amazondynamodb/latest/developerguide/bp-global-table-design.html) and [Aurora Global Database recovery guidance](https://docs.aws.amazon.com/AmazonRDS/latest/AuroraUserGuide/aurora-global-database-disaster-recovery.html). Confirmed the MREC/MRSC distinction and feature constraints, and planned switchover versus unplanned failover. Added explicit cross-pass guidance to prevent database, recovery and register contradictions.
- Reviewed all four pass-11 chapter structures and their worked cases. Checked 43.2 minutes of monthly downtime budget, the 45/62-minute dependency critical paths, weekly-backup and completion-delay bounds, 128/248 ms assumed replication paths, 7.81 sequential commits/s and 64 in-flight commits, survivor-capacity calculations, and the 113.7/139.9-minute restore budgets. The calculator's immediate-backup assumption is explicitly narrower than the chapter's delayed-completion example. The cropped tier question remains provisional, and restoring old state is distinguished from transaction rollback, compensation and engine downgrade.
- Reviewed pass-12's four chapter structures and diagnostic examples. Checked the 69.12 GB/day log volume, 5.76 GB outage backlog and four-hour net drainage, nearest-rank percentile counterexamples, 20-times budget burn and 2.78% monthly consumption, 26,700 failed search attempts, and the 20-to-70-second report decomposition. The guide distinguishes event history from data-event coverage, shared AWS role identity from an authenticated human, swap occupancy from current pressure, and host spare capacity from container limits. It explicitly notes that EXPLAIN ANALYZE executes work and that diagnostic commands were not run against a live AWS workload.
- Reviewed pass-13's four security chapter structures and worked cases. Checked the narrowly scoped S3 object-read policy, 98.3% illustrative excess-permission matrix, old-session/reissuance timeline, 18,000 possible stale-secret connection attempts, approximately 1.91 TiB of envelope-encrypted payload, and the 75% discovery-precision example. Found a bare unresolved citation marker in the draft and strengthened the existing validator to catch both bare and labeled markers before pass acceptance. Independently opened [current KMS rotation guidance](https://docs.aws.amazon.com/kms/latest/developerguide/rotate-keys.html) on 2026-09-06 to verify eligible automatic rotation and the distinction from data-key rotation or data re-encryption.
- The pass-13 recheck passed with no errors or warnings after the citation repair. The pass also added the GetCallerIdentity limitation so identity inspection is not presented as proof that permissions have been revoked. Security chapter body text totals 8,345 words at that verified revision.
- Extended the security review to value parameterization versus identifier allowlists, tenant authorization, context-specific browser escaping, HttpOnly versus CSRF, WAF rule ordering and inspection limits, and IAM direct-session/resource-policy exceptions. Checked the rule-evaluation corpus calculations: 90% recall, approximately 69.2% precision, 0.4% false positives and an illustrative 8,000 affected legitimate daily requests. The examples distinguish count-mode observation from enforcement and preserve documented SCP scope limitations.
- On 2026-09-06 UTC, independently opened AWS's [ElastiCache durability overview](https://docs.aws.amazon.com/AmazonElastiCache/latest/dg/durability.html), [write options](https://docs.aws.amazon.com/AmazonElastiCache/latest/dg/Durability.Options.html), and [limitations](https://docs.aws.amazon.com/AmazonElastiCache/latest/dg/Durability.Limitations.html). Confirmed the current Valkey durability modes and deployment restrictions. Flagged the earlier session chapter for explicit integration repair and cross-linking, including the distinction between Multi-AZ durability and cross-Region recovery. That repair remains a final integration requirement until verified.
- Reviewed pass-14's four chapter structures, the new durability section and exact consistency/TTL qualifications. Checked the 32-second repeated-plan work example, 7,680 MiB possible sort demand, 12,000 DynamoDB write units and 24-way sharding arithmetic, 18,000 strong-read units, cache warm/cold demand, stale-refill and lease-expiry races, IOPS/throughput conversions, and the 94.1-minute restore budget. The text distinguishes logical key buckets from physical partition reservations, strong reads from transactions, S3 object readiness from asynchronous derivative completion, and EFS Archive from S3 archive-retrieval workflows.
- Reviewed pass-15's five chapter structures and worked examples. Checked the fictional €2,660/€2,090 operating-cost comparison, 1.44-million-event backlog and thirty-minute catch-up, one-zone-loss worker counts, and 120 TB transfer/catch-up arithmetic. The chapter ties all three lock-in ratings to workload-specific exit evidence. Added an integration wording check to distinguish capacity 20% above demand from keeping 20% of installed capacity spare; the denominator must be explicit. The messaging pseudocode is labeled illustrative and separates local atomic effects from externally coordinated payment effects.
- Reviewed pass-16's four chapter structures and the cost, performance and documentation examples. Checked €1,640/€925 monthly TCO, 6.3-month versus seventeen-month payback sensitivity, 12.5% hourly commitment utilization, 201-minute expected checkpointed runtime, and the 4.5-hour fallback boundary. Checked CPU/database capacity limits, Little's law, 15 MB bandwidth-delay product, IOPS/throughput and connection-pool arithmetic. Documentation coverage distinguishes source updates from successful publication, schema validation from business correctness, and compatibility repair from rewriting a guide. All 108 technical questions now have explicit links to substantive chapters; the three reflection items remain separate.
- Checked the 314 source URLs present after pass 16 on 2026-09-06 UTC. 312 returned HTTP 200. The two MySQL pages returned HTTP 403 to the automated client but both opened successfully through the browsing tool with the expected official content: [slow-query logging](https://dev.mysql.com/doc/refman/8.4/en/slow-query-log.html) and [implicit commits](https://dev.mysql.com/doc/refman/8.4/en/implicit-commit.html). `source-link-review.json` preserves the actual statuses and the supplemental checks. Link availability is distinct from claim support; the content review above supplies the latter checks.
- Reviewed the eight pass-17 leadership guides, each with 727–835 explanatory body words and eight distinct prompts. Inspected all fictional examples, STAR workshops, misuse/tension discussions and prompt topics. The guides preserve unsuccessful and partial outcomes, shared credit, evidence limits and truthful experience scope. The recovery example improves to 70 minutes while explicitly missing its one-hour target; the connectivity example distinguishes a three-segment pilot from an uncompleted rollout. Hiring/coaching links are labeled transfer exercises because the photos contain no direct hiring question. The four-choice learning-preference pair and both experience reports remain unscored and are explained separately.
- Reviewed the eight pass-18 leadership guides, totaling 6,447 explanatory words plus 64 distinct prompts at this revision. Inspected their STAR workshops, fictional cases, principle tensions and quiz connections. Checked the 40-to-30-unit cost comparison (25% reduction, with implementation costs explicitly excluded from that simple arithmetic). Disagreement does not require concealing new material risks; delivery extends through publication, adoption and operation. Employer and broader-responsibility connections are explicitly indirect. Reduced paging is checked against suppressed reporting, pseudonymous data is not called anonymous, and lower compute cost is not presented as proof of lower environmental impact. The complete set now contains 128 prompts, with full-project validation still pending.
- Checked the twenty newly added leadership source URLs, including the official principle transcripts: all returned HTTP 200. The availability report now covers 334 distinct source URLs, retaining the separate successful browser checks for the two MySQL automated-client denials.
- During register review, found that the SNS technology recommendation needed a topic-type qualification for replay. Independently verified [SNS archiving and replay capabilities](https://docs.aws.amazon.com/sns/latest/dg/message-archiving-and-analytics.html) and [FIFO archive configuration](https://docs.aws.amazon.com/sns/latest/dg/message-archiving-and-replay-topic-owner.html) on 2026-09-06 UTC. Added an integration requirement to account for configured FIFO archive/replay while preserving its distinction from a streaming consumption model. The final repair still needs verification.
- Reviewed pass 19's published register: 313 terms and 42 technology decisions across twelve topic labels. Every one of the 35 technical chapters has a term link, and the prominent acronym handoff has no omissions. Sampled definitions across all subject areas and inspected all 42 technology comparisons. Checked 8 MB BDP, 156.25 MiB/s IOPS conversion, 100 Mbps versus 12.5 MB/s, current durability modes, least-connections versus least-outstanding-requests, and Macie's discovery/enforcement boundary. Flagged precise wording improvements for cache TTL, RPO and percentile aggregation; these and SNS replay remain integration checks until repaired.
- The pass-19 completeness audit expanded the register to 326 terms. Reviewed all thirteen additions, including Macie, Shield, Direct Connect, Image Builder, OpenSearch, Compute Optimizer and Little's law. Definitions retain the operational limits rather than implying that a named service automatically meets a workload objective. The pass validator succeeded with no errors or warnings.

## Browser verification

The isolated Playwright test in `scripts/smoke_test.py` checks standalone-file use, state persistence before submission, per-level explanations, bookmarks, question navigation, test/review mode, search, export/import/reset, invalid-import handling, mobile and desktop overflow, print coverage, and JavaScript errors. It runs against a separate browser context and does not reset the learner's browser.

Initial browser verification passed with the first four reconstructed batches. Final verification must run again with the completed dataset and `--full`; see `browser-validation.json` for the latest actual result.

After pass 09, repeated the browser suite successfully with five real chapters. Independently inspected desktop and 390px chapter screenshots and confirmed that chapter anchors scroll to their content. Exact-title global search was improved so reference material is not buried behind incidental question-body matches; the focused lookup and mobile-fit checks passed. Final full-dataset verification remains required.

Keyboard action review reproduced focus dropping to `BODY` after question navigation, bookmarking and explanation reveal. Fixed action-specific focus placement without stealing focus from search/filter inputs. The browser suite passed again with checks for retained bookmark focus, question-heading focus after navigation and exact-title global lookup. Replaced the full question-panel live announcement with the concise question-count announcement to avoid duplicating long explanations.

After pass 16, the expanded browser suite passed with all 111 questions and 35 technical chapters. It additionally verified a defensible alternative answer and the unscored cropped tier-recovery item. Desktop/mobile fit, print coverage and the existing state-management checks passed with the larger document. Leadership and register interaction checks have been added and will run once those sections contain their completed data; this pass remains partial-dataset verification.

After pass 17, the browser suite passed with eight Leadership Principle guides. It verified the complete card/article mapping, exact-name search, expanded evidence guidance and a real leadership-to-quiz transition, alongside the existing mobile, print and state checks. The register and final 16-principle dataset still require the final full run.

Visually inspected the complete 16-principle overview on desktop and the final principle's long title/body at 390px. Found the two-digit sidebar badge wrapping its digits at desktop width; prevented badge shrinking/wrapping and slightly tightened desktop navigation spacing. Rebuilt and verified that the numeric badges stay on one line. Final full browser verification will cover this CSS change with the completed register.

After pass 19, the browser suite passed with 326 terms and 42 technology choices. Visual register review prompted two improvements: exact acronym matches now precede incidental text matches, and mobile entries display definition, practical relevance and chapter links in one column without horizontal scrolling. The expanded suite passed again after these changes, including alphabetical print order and restoration of the active lookup afterward. Verification rejects a document changed during the browser run. Final `--full` verification against the accepted twentieth-pass artifact remains pending.

## Final verification — 2026-09-06 UTC

Final editorial review: complete.

- Confirmed the complete inventory of 93 supplied images, comprising 92 question photographs and the historical poster, and 111 unique reconstructed questions. Integration reconciles batch-local gaps and lists nine numbered actions absent from the complete supplied corpus. No missing question or unreadable option is invented.
- The final quiz contains 76 ratings, 32 choice items and three reflection items. There are 106 reasoned study keys; the function-policy and tier-recovery choice items are unscored because their full choice sets cannot be read. The three experience/work-style items remain truthful reflection exercises. Each visible option retains its reasoning, and source selections are not treated as keys.
- Reviewed the actual integration repairs to ElastiCache durability and DynamoDB TTL/MRSC scope, the SNS standard/FIFO replay distinction, and the explicit capacity-headroom denominator. Independently verified the SNS EndingPoint pause/resume behavior against [AWS's subscriber documentation](https://docs.aws.amazon.com/sns/latest/dg/message-archiving-and-replay-subscriber.html). Checked the equal-sized percentile example: two 100-request groups at 10/100 ms have a combined nearest-rank p95 of 100 ms, not the average 55 ms. Cache TTL and RPO definitions now state their intended boundaries precisely.
- Added a final 492-word SSRF section with the request mechanism, distinction from excessive privilege, destination/redirect/metadata boundaries, an original controlled-test example and diagnostic tradeoffs. Added the IMDS/IMDSv2 register entry and direct question/chapter links. This independent editorial expansion follows the twenty accepted model passes and is not counted as another pass; its change record is `editorial-final-changes.json`.
- Final content contains 35 technical chapters and 65,342 section-body words, 16 current Leadership Principles with 12,552 explanatory body words and 128 distinct prompts, 329 register terms, and 43 technology decisions. Every technical question has explicit substantive background coverage, and every chapter has register links. The full validator confirms the required chapter/principle depths, names, source records and internal links.
- Final source availability review covers all 373 source URLs: 371 returned HTTP 200, and the two automated-client MySQL denials were separately retrieved through the browsing tool with the expected official content. No unresolved missing-link finding remains. Source availability is not a claim that every service was exercised; infrastructure examples and the new SSRF fixture are teaching material, not live AWS test results.
- `python3 scripts/validate.py --full --execution` passed with zero errors and zero warnings. It verifies twenty genuinely completed, material GPT-6 Astra/high iterations with per-attempt evidence and excludes the failed attempt. The runner is finished, with no active model pass.
- `python3 scripts/smoke_test.py --full` passed against the same stable final HTML. It covers both unscored crops, defensible alternatives, all interface sections at 390/768/1440px, state persistence/import/export/reset, leadership prompts, exact register search, mobile register reading, chapter navigation, complete print coverage and restoration, and no JavaScript runtime errors. Final overview, quiz, leadership and register screenshots were visually inspected during review.

Both full verification reports identify HTML SHA-256 `e287771ad8b48bb261e6d5a77464e2184a3dbdecd3beb0665963430dbcff5b5c`. Final acceptance and PRD story completion are recorded only after these results.