Created pitches/a.md: a 230-word pitch on Ethereum history preservation, supported by five checked primary sources.

Verified word count, sentence limits, listed banned words, citations, and distinction from prior leads. Only the requested file was created.

I've written the pitch to pitches/c.md and committed it (197 words, only that file).

The argument: North Korea's TraderTraitor hacking group has stolen from three exchanges in 20 months without breaking any cryptography. Each time, it got the exchange's own approval process to sign the withdrawals:

• Bybit, 21 February 2025: about $1.46bn, through a Safe{Wallet} developer's compromised machine.
• Bitget, 24 September 2026: $387.5m, through altered transaction data.
• Binance, 29 October 2026: about $6.4bn, when valid signed authorizations (EIP-7702, which lets an ordinary wallet run code) handed its hot wallet to the attacker's code.

What actually saved money was the stablecoin issuers' ability to freeze funds: $2.8bn frozen, about 44% of the Binance total. The bitcoin and ether stay out of reach.

Selection criteria:

• Consequence: it is the largest crypto theft on record.
• Freshness: the Binance theft was two days ago.
• Sourceability: six sources, three of them primary.
• Not covered to death: no rehearsal edition led with an exchange theft (prior leads were the Hoodi delay twice, stateless workers, SEACAST and signature-free authorization). Coverage so far counts dollars rather than how the approvals failed.

Sources: I fetched each URL and saw it load with the content cited.

• Primary: FBI notice I-022625-PSA, Safe's own statement on X, and the EIP-7702 specification.
• Other: Chainalysis's on-chain tracing (30 October 2026) and Elliptic's Bybit analysis.
• Secondary, disclosed in the pitch: Wikipedia. It is the only source I have for the Bitget loss figures.

Uncertainty, stated in the pitch:

• The Binance attribution is an assessment by analytics firms. The FBI has named TraderTraitor only for Bybit.
• Binance's explanation (a compromised employee laptop) is preliminary.
• The open question is why the signing controls approved the withdrawals.

Checks: zero em dashes, no banned words, no relative dates, and no sentence over 22 words. I could not open CoinDesk (bot check) or Cryptotimes and Cryptobriefing (both refused access), so none of them is cited. Some figures, like the frozen total, may change as the investigation goes on. The wc word count includes the URLs.

Created pitches/b.md: 208 words, six checked primary sources, and no repeated prior lead.

Style, citation accessibility and file-scope checks passed. Only the requested file changed.

I wrote topic.txt, and it picks Pitch B as the winner. That's the only file I changed. I left it uncommitted because you didn't ask for a commit. There was nothing to build or run tests on for this task. My checks were opening the cited sources, reading dates from GitHub's API, and checking the file against the style guide: no em dashes, none of the banned words it lists, and the 43% figure computed rather than typed by hand.

Winner: Pitch B, "Ethereum's past needs a maintenance budget." It scores best on sourceability and on being under-covered.

• Its news hook holds up. On 28 September 2026 a change to EIP-4444 cut how long nodes keep old history from 33,024 to 14,299 epochs. That's 43% of the old window, and GitHub's records confirm the date. The proposal is still a draft.
• Its argument has a real basis. The proposal itself says paying for and monitoring the preservation of old records is out of scope.
• It's fair to the other side. It includes the strongest counterargument in the words of Matt Garnett, who announced the change for the Ethereum Foundation.
• It's new ground. None of the five rehearsal editions led with this subject.

Ranking of the losers:

• 2nd, Pitch A: Same argument and sound sources, but a weaker hook. It hangs on one client release from 20 August 2026, which came before the change B cites. It also has no named counterargument to test the thesis against.
• 3rd, Pitch C (the Binance theft): The freshest and most consequential pitch, but the weakest on sourceability and the most heavily covered story of the week. Several of its claims didn't hold up when I checked:
  • Chainalysis's post never mentions EIP-7702, so that claim rests on Wikipedia.
  • How the Bitget theft was carried out is sourced only to Wikipedia.
  • The FBI has named the TraderTraitor group only for the Bybit theft, not Bitget or Binance.
  • The time attached to Chainalysis's "$2.8bn frozen" figure is still an unfilled placeholder in their post.

The file also has a short section of verification notes for the record, marked not for publication. The vote picks one pitch and doesn't combine them.

Created feature.md: 2,668 words, six sections, full citations, SOURCES and METHOD.

Structure and style checks pass. The method discloses that review was adversarial self-review; no separate reviewing agent was available.

Verdict: publishable under its own rules, yes, with two medium defects an editor should fix before it ships. No claim in feature.md contradicts its source, no banned word appears, no sentence exceeds 35 words, no em dash appears, every paragraph is cited, and the body runs about 2,670 words.

What I checked. I fetched the style guide and all 23 cited sources, compared each paragraph's claim to the source text, and ran mechanical checks for sentence length, paragraph size, em dashes, banned words, relative dates, first person, and uncited paragraphs. The Garnett figures, the Geth 1.16.0 and archive-guide sizes, the Eth Docker epoch count and date, the amendment's commit date and values, the 43% ratio, the EIP statuses, the Era1 cap, the history-endpoint dates, the Portal, FAQ, Glados, weak-subjectivity, sync-mode and checkpoint-update wording all match. The ethPandaOps file counts match the provider's checksum lists exactly.

Findings, ranked. All six are in .imd-findings.json with reproductions.

• Medium, topic.txt line 20. The vote record says nothing from Pitch A was merged into B. The feature opens its news section on Pitch A's Eth Docker peg, which the vote itself called the weaker peg, and borrows A's Download Era source and its follow-a-transaction device. The record and the piece disagree.
• Medium, feature.md lines 29 to 121. The style guide says uncertainty is said plainly once. The feature carries 13 disclaimer sentences across 11 paragraphs, several restating the same limit, plus 18 "X, not Y" contrasts. The voice flattens into one rhythm.
• Low, topic.txt line 6. The vote asserts access has become "a public duty that nobody has funded". EIP-4444 says only that risks exist if incentives are lacking. The writer rightly withdrew that claim at lines 119 and 121, so the headline rests on a premise the body does not support.
• Low, feature.md line 63. The cited ethPandaOps page serves zeros in static HTML; the figures appear only after browser rendering. Counts are verifiable from the uncited checksum lists. The gigabyte sizes could not be verified from any fetched text.
• Low, feature.md lines 9 and 113. "The Merge" and "trie" are never taught in one plain sentence, as the feature register requires.
• Info, feature.md line 51. The 8,192-block cap appears without a paired context number.

Early in the session a batch of source downloads landed in the repository root because a scratch directory was unwritable. I moved them out immediately. The working tree contains only the original files and the findings file.