Investigator report — 2026/06/28
Verdict
A coherent edition with a genuine cross-domain spine — the LAB lead and THE QUESTION both land well, and the PELOTON coverage of the Belgian nationals is solid. The edition is let down by three specific failures: a stale Movistar source that the fact-checker identified but the pipeline kept anyway; a broken promise inside THE LAB article ("THE LONG READ covers it today" when the section is empty); and THE QUESTION re-treading Tour heat territory that was already the question two days ago. The run itself was clean and fast, with no agent failures. These are editorial judgment issues, not infrastructure ones.
Frontpage
The rendered PNG reads clearly as a newspaper front page. The LAB headline dominates the upper left at 68px, appropriately reflecting its priority-80 lead status. The PELOTON image is the pen-and-ink editorial illustration specified in the prompt — Belgian countryside road, lone rider crossing a finish line strung between barriers, peloton dissolving in the distance — and it is well-matched to the story and well-placed. The image occupies a 644px-wide landscape panel, giving visual weight to the lower half of the page.
The LAB headline ("China's CPU-Only Supercomputer Beats El Capitan — Without a Single GPU") runs to six stacked lines in the upper left and the last word is clipped at the canvas edge: the full headline ends with "GPU" but on the rendered page it reads "Singl CPU" before cut-off. The issue is purely visual — the headline is complete in the HTML — but a reader glancing at the front page without clicking through will see a broken line.
THE WORLD column is headline-only per the rules (frontpage_display: "headline_only") and occupies 240px on the bottom rail, which is appropriate. FROM THE ARCHIVE and ALSO NOTED fill the remaining columns. THE FUNNIES is correctly absent from the front page (frontpage_display: "skip"). Section ordering on the page respects priority order within the layout structure. The daily-decision strip is present and correctly rendered.
The index.html at pd.thep3000.com renders all sections including THE FUNNIES and FROM THE ARCHIVE. THE LONG READ is present as a section header but contains only a blank body — the template correctly handles the empty sentinel file.
Priority ranking
| Section | Priority | Length | Image | Notes |
| THE LAB | 80 | 519 words | no | Lead, upper left |
| THE QUESTION | 78 | 428 words | no | Secondary column, row 1 |
| THE PELOTON | 72 | 1,041 words | yes | Row 2, with lead image |
| THE WORLD | 62 | 1,151 words | no | Headline-only on front page |
| FROM THE ARCHIVE | 40 | 220 words | no | Bottom rail |
| ALSO NOTED | 12 | 391 words | no | Bottom rail |
| THE FUNNIES | 8 | 65 words | no | Hidden from front page |
| THE LONG READ | — | 0 words | no | Held, section empty |
The rankings are broadly defensible. THE LAB at 80 earns its lead: LineShine topping the TOP500 without a single GPU is a genuine story with geopolitical legs. THE QUESTION at 78 is high for what is ultimately a cross-domain editorial; it reflects the pipeline's enthusiasm for the cross-domain bridge more than the story's objective news weight, but the piece is strong enough to live there.
THE PELOTON at 72 is slightly low for a national championships day five days before the Tour de France Grand Départ. Wellens winning in a genuine surprise is solid racing news. The priority could defensibly sit at 75–76. THE WORLD at 62 is appropriate.
No priority inflation or compression issues. THE ARCHIVE is correctly capped below 45 and never leads the paper.
Editorial reading
THE QUESTION and the recency violation. The QUESTION agent ran the recency check, found the conflict, and pressed ahead anyway. Its reasoning — that the cross-domain bridge (ASO tweaking shade vs. China scrapping GPU dependence) is structurally different from Jun 26's "Can the Tour Stay in July?" — is plausible but thin. Both questions are ultimately about whether an institution can adapt proportionally to a structural threat, with the Tour as the primary example. The agent correctly noted the overlap and then rationalized around it. The result is that the reader has now seen the Tour-heat angle framed as THE QUESTION two days in a row. The agent should have either committed to the China-only angle (export controls failing as a ceiling is entirely self-standing) or drawn the question from Wellens-Evenepoel, where institutional rules about mandatory participation created a situation nobody wanted. Either would have given the edition more range.
THE LAB's broken cross-reference. The final paragraph of THE LAB article reads: "The full exchange is worth reading; THE LONG READ covers it today." THE LONG READ is empty. The LAB fact-checker ran before the longread writer resolved its held status, which is why the cross-reference survived. The sentence now misdirects the reader. This is a sequencing problem — the LAB FC ran in parallel with the longread writer — but the practical result is a broken sentence in the published article. The longread writer's "Done" summary explains the hold correctly; what's missing is a mechanism to propagate that hold signal to sibling sections that have already written cross-references.
THE PELOTON's stale Movistar paragraph. The Movistar/Quantum Pacific investment story dates from April 11, 2025 — 437 days before this edition, well beyond the 7-day recency cap. The fact-checker explicitly identified this, removed the "thread moves" framing, and corrected "last December" to "in December 2024." But the paragraph itself survived into the final article because the FC chose to fix the framing rather than excise the item. The result is a paragraph about a 14-month-old deal that sits alongside same-day race reports. A reader noticing the source date in the citations (Apr 17, 2025) will lose confidence in the edition's freshness. The FC call to correct rather than cut was wrong here; the recency cap exists precisely for this.
ON THE TRAIL's missing per-day mileage format. The newspaper.yaml spec for ON THE TRAIL picks is explicit: per-day mileage AND elevation gain are required in a "Day 1 in: X mi, +Y ft / Day 2 out: X mi, –Y ft" format. All three picks in this edition give total round-trip figures only. Iron Bear: "6.5 miles round trip, 2,100 ft gain." Sheep Lake: "~2 miles in, 300 ft gain" (one-way only, no Day 2 format). Thorp Lake: no mileage at all. The reader planning a backpacking trip cannot size their days from aggregate totals; this is the information gap the format is designed to close. The world writer followed the picks criteria correctly but skipped the format requirement entirely.
THE WORLD China bullet duplicates the LAB lead. The world section includes "China's fastest supercomputer — LineShine, installed at the National Supercomputing Center in Shenzhen, is now ranked the world's fastest, built without GPUs despite U.S. export controls" as one of its three world-news bullets. This is the LAB's lead story, already covered in 519 words. Running a world-news bullet on a story that is simultaneously the paper's tech lead dilutes the value of both placements. The bullet also runs 26 words against the 25-word hard cap — a minor violation, but the cap is described as "HARD CAPS" in the config. The world writer should have used one of the three bullet slots for something the LAB does not already cover, particularly given the availability of the UCI power passport story (riders' union furious, three sources, active thread as of today) or the Paul Seixas €13M transfer story (opened Jun 27 as its own thread), neither of which appeared anywhere in this edition.
Coverage gap: Paul Seixas and the UCI power passport. The search results and threads carried multiple sources on two genuinely interesting pre-Tour stories: Seixas being offered what would be the largest cycling contract in history (€13M, thread opened Jun 27), and the UCI power passport feasibility trial with the riders' union declaring itself "100% against" it. Both threads are listed in research.md. Neither appears in section-peloton.md's dropped array, nor in also-noted.md's dropped array. The Peloton writer was clearly focused on the Belgian nationals results (four separate story blocks) and ran out of room, but with the Tour Grand Départ six days away, the Seixas transfer story is the kind of item the reader wants in their head before Stage 1. It belonged in ALSO NOTED at a minimum.
Pipeline observations
Starting commit. The dispatch commit (95fdc46) runs from parent 51bd7df, which is the 2026-06-27 investigator commit — same-day start. No concern.
Agent inventory. 19 subagents ran: Scout, Researcher, writers for THE WORLD, THE PELOTON, THE LAB, THE LONG READ, FROM THE ARCHIVE, ALSO NOTED (sweep), THE QUESTION (reflector), THE FUNNIES (comic); fact-checkers for THE LAB, THE PELOTON, THE WORLD, FROM THE ARCHIVE, THE QUESTION, ALSO NOTED; Meta-Writer, Art Director, Thread Editor. All expected agents are present. No duplicates. No dedup subagent — the dedup function runs inline in the dispatcher as a build_coverage_index.py call, which is expected.
THE LONG READ held correctly. The longread writer's 11-message transcript shows a well-reasoned hold: primary candidate (Carmack/Quake/Petersen, PCGamer Jun 28) was identical ground to the Jun 25 THE LONG READ; backup candidate (Anthropic/Fable export-control piece) was fetched-blocked and the Jun 26 edition had already run a full Anthropic piece. The agent's Done message ("section held per brief's explicit guidance") is correct and the empty sentinel file is appropriate. THE LONG READ is optional and the hold call is defensible.
FC: THE PELOTON — stale source not removed (278s run). The fact-checker correctly identified the Movistar source as 437 days old, corrected the framing, and flagged it as stale — but kept the paragraph. The section's recency_cap_days is 7. The FC's resolution log notes "stale sources" and lists the Movistar URL, then proceeds to write the corrected file without removing the paragraph. This is a fact-checker judgment error: a source outside the recency cap should trigger removal of the dependent content, not just a framing correction.
FC: THE LAB — cross-reference to empty section not flagged. THE LAB's article ends "THE LONG READ covers it today." The LAB fact-checker ran in parallel with (or before) the longread writer's hold decision and cleared all 18 claims without flagging this forward reference. A sequential check after all writers resolve would catch this, but the pipeline runs parallel by design. This is an architectural limitation more than a checker failure, but the practical impact on the reader is the same: a broken sentence in the published article.
No fetch failures. All 31 fetch results returned ok: true. No pipeline-alerts file. Clean on this dimension.
Also-noted "source unverifiable" drops. Several items were dropped from ALSO NOTED with "source unverifiable" as the reason, including Phoronix (Linux 7.2 pipe performance), an OxCaml essay from theconsensus.dev, and a Wired story on the Euclid telescope's Milky Way image. Phoronix is an established tech outlet and should not fail a verifiability check. The sweep writer appears to have failed to fetch these sources and treated the fetch failure as a verifiability failure rather than retrying or flagging the fetch problem. The TIMELESS OVERRIDE rule in the config explicitly guards against this for undated items; a similar exception should apply to known outlets where the fetch was simply blocked rather than the source being genuinely unreliable.
Trace highlights
Researcher at $1.39 vs. THE LONG READ at $0.04. The researcher spent $1.39 (703s) building a comprehensive brief that correctly identified the longread's primary candidate, flagged the Jun 25 overlap risk, and provided a backup. The longread writer spent $0.04 (19s) on 11 messages to confirm the hold. The ratio is expected — the researcher does the expensive work so writers don't have to — but the long-read hold means that $1.39 of research cost generated zero article output. On a day where two strong cycling threads (Seixas, power passport) were also in the brief but unused, the researcher's output was substantially over-built for what shipped.
FC: THE PELOTON at 278s is the longest fact-checker run. This reflects genuine work: the checker identified a false Evenepoel TT win claim (contradicted by multiple independent sources), the Movistar recency issue, and a Caja Rural founding-year framing problem. Three corrections in 38 checks is a healthy catch rate. The runtime suggests the checker was doing real web searches rather than just reading source files — appropriate for a section with 9 sources.
Orchestrator at $2.85 — the most expensive single agent. The orchestrator's cost exceeds the next most expensive agent (Researcher, $1.39) by roughly 2x. With 389 session messages, the orchestrator is carrying significant context across the full pipeline. This is expected for a run with 19 parallel subagents and full task-notification handling, but it is the right cost center to watch as the pipeline grows.
THE WORLD at 138s and 0 input tokens. Near-zero direct input tokens (6 tokens) combined with heavy cache reads (85K read + 68K created) shows the world writer read everything from cache. Fast and clean.
Trace summary
| Agent | Dur | Input | Output | Cache Read | Cache 5m | Cache 1h | Cost |
| Scout | 323s | 12284 | 329 | 460431 | 77837 | 0 | $ 0.47 |
| Researcher | 703s | 33 | 1210 | 2924845 | 132120 | 0 | $ 1.39 |
| THE WORLD | 138s | 6 | 4 | 85049 | 68454 | 0 | $ 0.28 |
| THE PELOTON | 116s | 10 | 129 | 172315 | 36684 | 0 | $ 0.19 |
| THE LAB | 80s | 9 | 82 | 108435 | 21421 | 0 | $ 0.11 |
| THE LONG READ | 19s | 6 | 8 | 37924 | 6389 | 0 | $ 0.04 |
| FROM THE ARCHIVE | 36s | 7 | 11 | 74712 | 15071 | 0 | $ 0.08 |
| FC: FROM THE ARCHIVE | 76s | 9 | 14 | 123723 | 26604 | 0 | $ 0.14 |
| Meta-Writer | 42s | 6 | 4 | 45293 | 19386 | 0 | $ 0.09 |
| FC: THE LAB | 102s | 8 | 52 | 112817 | 23446 | 0 | $ 0.12 |
| FC: THE PELOTON | 278s | 1617 | 181 | 531336 | 49741 | 0 | $ 0.35 |
| FC: THE WORLD | 213s | 992 | 12 | 389443 | 56783 | 0 | $ 0.33 |
| THE QUESTION | 87s | 11 | 83 | 211962 | 28436 | 0 | $ 0.17 |
| Illustrator | 108s | 209 | 5488 | 0 | 0 | 0 | $ 0.22 |
| FC: THE QUESTION | 79s | 9 | 16 | 141218 | 22920 | 0 | $ 0.13 |
| ALSO NOTED | 185s | 17 | 307 | 484086 | 53434 | 0 | $ 0.35 |
| Draw today's TWO parody comic strips for | 103s | 12 | 96 | 211832 | 29343 | 0 | $ 0.18 |
| Funnies (OpenAI) | 116s | 379 | 5488 | 0 | 0 | 0 | $ 0.22 |
| FC: ALSO NOTED | 163s | 9 | 84 | 164921 | 32212 | 0 | $ 0.17 |
| Art Director | 119s | 5 | 3 | 42526 | 36330 | 0 | $ 0.15 |
| Update story threads for today's edition | 219s | 5 | 5 | 41498 | 44112 | 0 | $ 0.18 |
| Orchestrator | | 152 | 25341 | 6124397 | 0 | 105215 | $ 2.85 |
| TOTAL | | 15795 | 38947 | 12488763 | 780723 | 105215 | $ 8.21 |
Suggestions for next edition
Wire THE QUESTION recency check harder. The agent found the Jun 26 Tour/heat collision, decided it was different enough, and proceeded. The structural test should be: if you are drawing the primary domain from THE PELOTON and the Jun 26 question also drew its primary domain from THE PELOTON on the same underlying tension, the check has failed regardless of framing. Tighten the check to compare primary domains and trigger events, not just structural phrasing.
Remove content whose source is outside the recency cap, not just flag it. The fact-checker's current behavior on stale sources is to correct framing and annotate. When a source is more than 3x the recency cap (7 days × 3 = 21 days), the dependent paragraph should be removed unless the writer explicitly marked it evergreen. The Movistar paragraph is the third instance of a stale source surviving into the published article because the FC corrected rather than cut.
Add Paul Seixas and the UCI power passport to the PELOTON's regular brief. Both stories are active threads, both have multiple sourced articles, and neither has been covered in this paper. The Grand Départ is in six days. These stories need to appear in the edition before Stage 1 makes them retroactive.
Add an ON THE TRAIL per-day format check to the world fact-checker. The format spec is unambiguous (Day 1 in: X mi, +Y ft / Day 2 out: X mi, –Y ft) and has been missed in multiple editions. The fact-checker for THE WORLD should verify that each pick has a per-day breakdown, not just a round-trip total. One line in the FC prompt would close this gap.