# Research notes: AI output portability (sample input) This is the input package for the worked example in [Research to Executive Report](/en/workflows/research-to-executive-report/). It is a real research-notes file: excerpts from primary sources, margin notes, and half-formed conclusions — the shape of material most people actually have when they sit down to write a report. Provenance: each excerpt below is drawn from a source file archived in the AtomStorm sample library (`docs/research/workbench-samples/sources/`, fetched 2026-09-04); the archive path is noted per excerpt. The **Facts** lines stay within what the archived source states. The **My note** lines are the researcher's own interpretation — exactly the kind of unverified reasoning a report workflow has to separate from evidence. Working title: "Will our AI-generated deliverables still be editable in 12 months?" Audience (draft): engineering leadership, 10-minute read. Goal (draft): decide whether we need a portability policy for AI-generated documents and decks. --- ## Excerpts ### 1. Hosting-level redirects: pinned semantics - Archive: `sources/cloudflare-pages-redirects.md` (official doc, updated 2026-08-25) - Facts: a splat `*` greedily matches all characters; when multiple rules share a source path, the topmost one applies; Pages Functions bypass `_redirects` on the routes they serve. - My note: moving a doc site is cheap; moving it *without* breaking inbound links requires the redirect rule set to travel with the content. Portability is not just file formats — it's the URL layer too. ### 2. Template ecosystems organize output by category - Archive: `sources/gamma-templates.md` (product page, 2026-09-04 snapshot) - Facts: the template gallery advertises "100+ templates" across categories including Company, Consulting, Marketing, **Reporting**, Sales, Strategy, with "Browse templates" / "Start for free" as primary actions. - My note (inference, not from the page): template-first products make in-tool editing the path of least resistance, which suggests export is the second-class citizen — worth testing before we rely on it. Note "Reporting" as a template category lands squarely on our use case. ### 3. Search guidance rewards first-hand experience - Archive: `sources/google-helpful-content.md` (official guidance) - Facts: the guidance emphasizes demonstrating first-hand experience and expertise as signals of helpful, people-first content. - My note: whatever format we generate, the *evidence trail* (inputs, raw outputs, method) is what makes an AI-assisted document defensible — that part is my extrapolation; the page speaks about experience, not about AI documents specifically. ### 4. Code-edit benchmarks are public - Archive: `sources/aider-leaderboards.md` (product leaderboard, 2026-09-04) - Facts: per-model benchmarks for code editing accuracy are published and maintained. - My note (unverified): I could not find an equivalent public leaderboard for "revise a 30-page document without wrecking the styles" — absence from my search, not a proven absence. ### 5. AI crawlers split into roles with separate policies - Archive: `sources/anthropic-crawlers.md` · `sources/cloudflare-ai-traffic-options-2026.md` - Facts: Anthropic documents distinct crawlers (search/citation vs training); Cloudflare's 2026 AI traffic controls separate Search / Agent / Training with independent switches. - My note: an exported artifact that lands behind a login or a bot block loses the citation channel entirely. Portability includes *reachability*. ### 6. Bing reports AI citations per grounding query - Archive: `sources/bing-ai-performance-2026.md` (official blog posts, Feb + Jun 2026) - Facts: AI Performance reporting shows how publisher content is cited across Copilot and Bing AI answers; **Citation Share** is measured per *grounding query* (share of a query's citations), alongside Intents and Topics views. - My note: gives us a measurable definition of "our artifact got cited" — narrower than page-level, but still an acceptance criterion for the public version of this report. ### 7. Email deliverability has a published, tiered contract - Archive: `sources/gmail-sender-guidelines.md` (official help, 2026-09-04) - Facts: for all senders — SPF or DKIM, valid forward/reverse DNS, TLS, spam-rate under 0.3%; for senders of ≥5,000 messages/day — SPF **and** DKIM, DMARC, alignment, and one-click unsubscribe with a visible link for marketing/subscribed mail. - My note: example of a portability *policy* done right: the artifact (email) survives transport because the contract is written down and tiered by volume. The tiering itself is the lesson for us. ### 8. Research assistants ship citation-anchored extraction - Archive: `sources/elicit.md` (product pages, 2026-09-04) - Facts: the product documents sentence-level citations and extraction accuracy reporting. - My note (inference from one vendor): if this becomes table stakes, a portability policy that requires source-linked claims ages well. One vendor is a signal, not a trend. --- ## Synthesis so far (three observations — mine, not the sources') 1. Portability has three layers — file format, URL/redirect layer, and reachability (auth/bot policy). Debates that only argue about file formats are missing two-thirds of the problem. (Layers 1 and 3 are grounded in excerpts 1–2 and 5; the "three layers" frame is my own.) 2. The ecosystem measures code-edit portability (excerpt 4) and search citation (excerpt 6), but I found no public measurement of document-edit fidelity at revision time. Any claim we make about "still editable" needs a test we run ourselves. 3. Published contracts (email deliverability, robots semantics) show the pattern that works: write the transport contract down, enforce it mechanically, re-verify when versions change. ## Open questions - Q1: What is the cheapest honest test for "revision fidelity" of a generated deck/doc — and can it run monthly? - Q2: Does our current toolchain export resolve the URL layer (redirects) as part of the artifact, or does that stay manual? ## People to ask - Platform eng: who owns redirect rules on release day? - Design: what breaks first when a deck is re-opened after export?