# Worked-example output: Executive Report (real Studio run)

> Provenance: generated in AtomStorm Studio (Agentic mode, model `deepseek-v4-flash`
> via the AtomStorm gateway) on 2026-09-08, from the input package
> `sample-input-research-notes.md` and the prompt in `prompt.md`. Two generation
> passes (83.3s + 54.3s; 10,593 + 6,390 output tokens). Text recovered from the
> product UI and lightly reformatted (table line breaks); wording unedited —
> **including the model's own misreading in R5** (it claims note 7 was
> truncated; the source file is complete).
> **This output is a worked example, not a benchmark — see the workflow page's
> limitations.**

---

## 1. Bottom line (≤ 120 words)

Recommendation: approve a portability policy for AI-generated documents and decks, conditionally — effective only after two checks we run ourselves pass before Sept 30. Findings: Cloudflare's official docs pin redirect semantics (splat `*`, topmost rule, Functions bypass); official Anthropic and Cloudflare docs split AI crawlers into search/agent/training with independent switches; Bing's official blog defines Citation Share per grounding query; Gmail's official guidance is a published, tiered sender contract. Gaps: export fidelity from template-first tools is untested (inference); no public revision-fidelity benchmark was found (author's search); citation metrics are vendor-defined. We propose requiring redirect rules to travel with content and an evidence trail per artifact (inference from Google guidance, which rewards first-hand experience).

## 2. What we decided to ask (and what we did not)

We decided to ask:

- Is portability a three-layer problem — file format, URL/redirects, and reachability (auth/bot policy)? The frame is ours (author's synthesis); layers 1 and 3 are grounded in notes 1–2 and 5.
- Must the redirect rule set travel with content when a doc site moves (note 1, author's note)?
- Should an evidence trail (inputs, raw outputs, method) be required per artifact? This is the author's extrapolation from Google's guidance, which speaks to experience, not AI documents (note 3) — flagged, not a finding.
- Should citation reachability be an acceptance criterion for published artifacts (notes 5–6, author's notes)?
- Is Gmail's tiered, written-down contract the policy pattern to copy (note 7, author's note)?
- The notes' open questions: Q1 — cheapest honest test for "revision fidelity" of a generated deck/doc, runnable monthly? Q2 — does our current toolchain export resolve the URL layer (redirects) as part of the artifact, or does that stay manual?

We did not ask:

- Which specific tool/product to standardize on (no evidence in notes).
- Content quality or style standards (not covered by notes).
- Code-editing model selection (note 4 is an ecosystem observation, not a docs/decks decision).
- Tooling cost or ROI (no numbers in notes).
- Whether our current redirect rules are actually configured correctly (ownership is open — see "People to ask": platform eng).

## 3. Evidence

Every row below is traceable to an archive file; the notes record archive paths, not URLs, so sources are cited by path. Inferences from the notes are included only where flagged as such — they are assumptions, not findings.

| # | Claim | Source (archive) | Confidence (why) |
|---|---|---|---|
| 1 | Splat `*` greedily matches all characters | `sources/cloudflare-pages-redirects.md` (official doc, updated 2026-08-25) | Single source — official doc; not reproduced by us |
| 2 | When multiple redirect rules share a source path, the topmost applies | Same | Single source — official doc |
| 3 | Pages Functions bypass `_redirects` on the routes they serve | Same | Single source — official doc |
| 4 | Redirect rules must travel with content; portability includes the URL layer | Author note on excerpt 1 | Inference — author's framing, not the source's |
| 5 | Template gallery advertises "100+ templates" across categories incl. Marketing and Reporting, with "Browse templates" / "Start for free" actions | `sources/gamma-templates.md` (product page, 2026-09-04) | Vendor claim — product page, not independent |
| 6 | Template-first tools make in-tool editing the path of least resistance, so export may be second-class | Author note on excerpt 2 | Inference — author's; explicitly "worth testing" |
| 7 | Google guidance rewards first-hand experience and expertise as signals of helpful content | `sources/google-helpful-content.md` (official guidance) | Single source — official guidance |
| 8 | An evidence trail (inputs, raw outputs, method) makes an AI-assisted document defensible | Author note on excerpt 3 | Inference — author's extrapolation; page speaks about experience, not AI docs |
| 9 | Per-model code-edit accuracy benchmarks are published and maintained | `sources/aider-leaderboards.md` (product leaderboard, 2026-09-04) | Vendor claim — product leaderboard |
| 10 | No equivalent public benchmark for "revise a 30-page document without wrecking the styles" was found | Author note on excerpt 4 | Inference — absence from author's search, not a proven absence |
| 11 | Anthropic documents distinct search/citation vs training crawlers | `sources/anthropic-crawlers.md` (official docs) | Single source — vendor official docs |
| 12 | Cloudflare's AI traffic controls separate Search / Agent / Training with independent switches | `sources/cloudflare-ai-traffic-options-2026.md` (official docs) | Single source — vendor official docs |
| 13 | An artifact behind a login or bot block loses the citation channel; portability includes reachability | Author note on excerpt 5 | Inference — author's, built on claims 11–12 |
| 14 | Bing AI Performance reports citations across Copilot/Bing AI answers; Citation Share is per grounding query, alongside Intents and Topics | `sources/bing-ai-performance-2026.md` (official blogs, Feb + Jun 2026) | Single source — vendor official; two posts corroborate each other |
| 15 | Citation Share can serve as an acceptance criterion for the public version of this report | Author note on excerpt 6 | Inference — author's |
| 16 | All Gmail senders need SPF or DKIM, valid forward/reverse DNS, TLS, spam rate <0.3% | `sources/gmail-sender-guidelines.md` (official help, 2026-09-04) | Single source — official help |
| 17 | Senders of ≥5,000 messages/day need SPF and DKIM, DMARC, alignment, one-click unsubscribe (visible link for marketing/subscribed mail) | Same | Single source — official help |
| 18 | Gmail's tiering is a model of a portability policy done right | Author note on excerpt 7 | Inference — author's |
| 19 | Elicit documents sentence-level citations and extraction accuracy reporting | `sources/elicit.md` (product pages, 2026-09-04) | Vendor claim — product pages |
| 20 | Source-linked claims age well if sentence-level citation becomes table stakes | Author note on excerpt 8 | Inference — one vendor is a signal, not a trend |

## 4. What we verified ourselves (and how)

What we checked:

- Read and archived each source with its date: Cloudflare Pages redirects (updated 2026-08-25), Gamma templates (2026-09-04), Aider leaderboard (2026-09-04), Gmail sender guidelines (2026-09-04), Elicit (2026-09-04), Bing AI Performance (Feb + Jun 2026 blog posts), plus the Anthropic and Cloudflare crawler docs.
- Distinguished facts from author's notes in every excerpt ("Facts:" vs "My note:"); the synthesis section is explicitly the author's, not the sources'.
- Cross-checked the Bing claim across two blog posts (Feb and Jun 2026) — both report Citation Share per grounding query.
- Checked that the Gmail contract is genuinely tiered (all senders vs ≥5,000/day) and recorded the exact thresholds.

What we did NOT verify:

- Cloudflare redirect behavior — no hands-on reproduction of splat/precedence/Functions bypass.
- Gamma export fidelity — no export/re-import test run (Q1 open).
- Aider benchmark methodology or reproducibility — vendor claim only.
- Elicit's extraction accuracy claims — vendor claim only.
- Whether AI-crawler policies (Anthropic/Cloudflare) actually apply to our content.
- Whether our toolchain emits redirects automatically (Q2 open) and who owns redirect rules on release day ("People to ask": platform eng).
- Whether our mail volume crosses the 5,000/day tier or our spam rate is under 0.3%.
- The existence of document-revision benchmarks beyond the author's single search (absence from a search ≠ proven absence).
- What breaks first when a deck is re-opened after export ("People to ask": design).

## 5. Risks and failure conditions

| # | Risk / failure condition | Grounding | Confidence (why) |
|---|---|---|---|
| R1 | **File-format portability is unverified.** If export from template-first tools is lossy, decks/docs won't survive a tool switch even with perfect redirects — the policy would rest on the URL layer alone. | Note 2: "100+ templates" incl. Reporting (product page) + author's note that export is the "second-class citizen" | Vendor claim for the page facts; **inference** for the export-is-second-class framing (flagged in notes) |
| R2 | **Single-source dependency on vendor redirect semantics.** Splat `*`, topmost-rule-wins, and Pages Functions bypass are documented only by Cloudflare; not reproduced by us. If semantics change or our Functions routes differ, inbound links break. | Note 1, official Cloudflare doc (2026-08-25) | Single source (official doc; unverified in our stack) |
| R3 | **Citation metrics are vendor-defined.** Adopting Citation Share as an acceptance criterion ties policy to a metric Bing defines per grounding query; it may not transfer across platforms. | Note 6, official Bing blogs (Feb + Jun 2026) | Single source / vendor claim |
| R4 | **Crawler/citation controls are independent switches.** Search, agent, and training traffic are separated per vendor; misconfiguring one switch exposes artifacts to the wrong crawler. Documented but unverified in our environment. | Note 5, official Anthropic + Cloudflare docs | Single source (vendor docs) |
| R5 | **Deliverability requirements incomplete.** Note 7 truncates at "one-click unsubscribe (o…"; full ≥5k/day tier (DMARC etc.) unread. Policy may assume compliance we haven't confirmed. | Note 7, official Gmail sender guidelines (truncated) | Single source (incomplete) |
| R6 | **Contradiction tension (not factual).** No factual contradiction among notes, but note 1 frames portability as URL-layer while note 2's inference says file export is second-class. Until export fidelity is tested, treat the URL/redirect set as the durable artifact. Note 4 (Aider, code editing) is out of scope for docs/decks — using it here would be scope creep. | Notes 1 vs 2 (author notes); note 4 | Inference |

**Failure condition:** approving without the section-4 checks would make the policy load-bearing on vendor claims and inferences.

## 6. Options

| Option | Description | Reversibility (if we stop/switch) | Portability (what happens to artifacts) |
|---|---|---|---|
| A. Approve now, no gates | Policy effective immediately | High — a text decision; revoking changes no tooling; no artifacts migrated | Artifacts stay in current tools; portability rests on untested exports (R1) and manual redirect handling |
| B. **Conditional approval (recommended)** | Approve in principle; takes effect after two checks pass (export-fidelity test; redirect + crawler-switch verification) | High — if checks fail, policy never takes effect; nothing to undo | Redirect rule sets travel with content (note 1); file exports verified before reliance; artifacts remain in-tool until migration is triggered |
| C. Reject / no policy | Keep ad-hoc practice | N/A — no change | No portability guarantee; inbound-link breakage stays a per-incident manual problem |
| D. Approve + mandate evidence trail | Policy plus per-artifact trail (inputs, raw outputs, method) | Medium — evidence files deletable, but retro-applying to existing artifacts is costly | Artifacts portable as in B; defensibility improved (extrapolation from note 3 — **inference**, flagged) |

## 7. Recommendation and next step

**Recommendation: Option B — conditional approval.** The load-bearing facts we hold are documented by vendors (notes 1, 5, 6, 7) and are single-source; the load-bearing assumptions (export fidelity, evidence-trail defensibility, deliverability) are inferred or unverified (notes 2, 3; note 7 truncated). Approving without checks would put vendor claims at the center of policy; rejecting would discard documented URL-layer mechanics we can use.

Two checks close the gap:

1. **Export-fidelity check:** generate a Reporting-category deck from the template gallery (note 2) and verify lossless export (structure, links, assets).
2. **Hosting check:** confirm splat/topmost/Functions-bypass redirect behavior on our stack (note 1) and configure search/agent/training crawler switches correctly (note 5).

**Next action:** Platform Tooling Lead (engineering) runs checks 1 and 2 and submits the final approve/reject recommendation to engineering leadership.
**Date:** Sept 30, 2026.

## Appendix

**Evidence traceability (claims in sections 5 and 7)** — supplements the section 3 table:

| Claim | Source | Confidence (why) |
|---|---|---|
| Redirect semantics: splat, topmost rule, Functions bypass (R2, check 2) | Note 1, official Cloudflare doc, 2026-08-25 | Single source (official; not reproduced) |
| Template gallery "100+ templates" incl. Reporting (check 1) | Note 2, Gamma product page snapshot 2026-09-04 | Vendor claim (product page) |
| Export is second-class in template-first tools (R1, Option D) | Note 2, author's note | Inference (flagged) |
| First-hand experience is a helpful-content signal (Option D) | Note 3, official Google guidance | Single source; extrapolation to AI documents = inference |
| Crawlers split search/agent/training, independent switches (R4, check 2) | Note 5, official Anthropic + Cloudflare docs | Single source (vendor docs) |
| Citation Share is per grounding query (R3) | Note 6, official Bing blogs, Feb + Jun 2026 | Single source / vendor claim |
| Gmail tiered sender requirements exist; note truncated (R5) | Note 7, official Gmail sender guidelines | Single source (incomplete) |

**Open questions (unsourced or unanswerable from notes):**

- **Q1:** Cheapest honest test for "revision fidelity" of a generated deck/doc, runnable monthly? (raised in section 2)
- **Q2:** Does our current toolchain export resolve the URL layer (redirects) as part of the artifact, or does it stay manual? (raised in section 2)
- **Q3:** Full Gmail requirements beyond the truncated excerpt (DMARC specifics) — to be confirmed against the complete guidance before the policy cites it.