This page is a preview. Do not publish it yet.

Blocker 1 Nothing on this page is buyable.

There is no checkout, no license file, no download, and navy-evals is not a git repository. The page offers a free validator download and two paid tiers. None of the three can be delivered today. The offer section is a mockup of an offer.

Needs: a delivery decision (repo? Gumroad? private?) before any traffic.

Blocker 2 Every price is invented.

$249 for The Forge, $1,490 for the Command License, free for The Gate, and the 30-day guarantee: all four are my calls, not Tim's. I picked them from the shape of the market (rapidepr.com and navyevalwriter.com sell AI eval generation; phrase libraries are free) and from the structure of the product, not from any pricing research or willingness-to-pay signal.

Needs: Tim to set or reject the numbers.

Blocker 3 Tim's own Navy affiliation may make selling this a conflict of interest.

Selling an evaluation-writing tool to Sailors, while affiliated with the Navy, raises a Joint Ethics Regulation question about using a position for private gain, and a second question about soliciting subordinates or peers. I cannot resolve this and did not try. It is the single item most likely to make the whole page moot.

Needs: Tim, and possibly a command ethics counselor, before anything ships.

Blocker 4 The live gate publishes unverified line and character limits as authoritative.

The gate embedded on the page is real and it enforces the numbers in _config/constraints.json. Those numbers carry the repository's own open item navfit98a-v33-line-verification at medium severity: they trace to the NAVFIT98A v30 user guide from 2012, and current software is v33. The page discloses this in section 11, but the gate itself does not, and a chief could trim a block to fit a limit that is wrong.

Needs: either verify against a v33 constraint table, or put the caveat inside the gate output, not just in the prose above it.

Blocker 5 Zero customers, and the page's credibility depends on that staying visible.

No real Sailor has ever been run through the pipeline. Both fixtures are synthetic. The page says so plainly in section 11, and that admission is doing real persuasive work. If anyone later adds testimonials, social proof, or a user count that is not literally true, the section 11 admission stops reading as candor and starts reading as a trick.

Needs: a standing rule that section 11 gets updated before any proof claim is added anywhere else.

Blocker 6 Publishing the gate publishes the encoded rule set.

site/gate-core.js is 71 KB and inlines the whole of constraints.json and quality-matrix.json. Anyone can read them from the browser. That is fine if the rules are the giveaway and the pipeline is the product, which is how the page positions it. It is not fine if Tim considers the encoded instruction to be the moat.

Needs: a deliberate answer, because the free-gate strategy depends on it.

Blocker 7 The name contains "Navy."

"Navy Eval Forge" is Tim's own name for the system and I did not invent it. On a public commercial page it may read as implying affiliation, which the footer disclaimer explicitly denies. The disclaimer is probably necessary and probably not sufficient.

Needs: a look at whether the name survives contact with a lawyer.

Blocker 8 The Forge tier requires Claude Code, and the page mentions it once.

A buyer at $249 who does not already run Claude Code cannot use what they bought. The page states the dependency in the offer card and nowhere else. That is thin for a hard requirement, and it also makes the price partly a function of a third party's pricing.

Needs: either a prominent requirements line, or a decision to package it differently.

Blocker 9 A public tool that invites eval text needs a real privacy statement.

The gate genuinely runs client side: no fetch, no analytics, no logging, and the claim is checkable in devtools. But the page asserts that in prose only. A tool soliciting CUI-adjacent text should say what it does and does not do in a durable, linkable place, and should probably carry a Content-Security-Policy that makes exfiltration impossible rather than merely absent.

Needs: a privacy note and a CSP header before this is public.

Review 1 I corrected one of my own claims mid-build.

The P.S. originally read "for E7 and above, silence on retention is a recommendation against retention." Reading validate.mjs line 301 showed the rule is scoped to Significant Problems reports for E7 through E9 and for all FITREPs. The P.S. now says that. Worth a second look because it is the most quoted-looking sentence on the page.

Review 2 The house standard is uncalibrated and the page says so.

quality-matrix.json floors are a first calibration. No real report has been scored against them. The live gate shows STD-WARN findings from those floors to strangers, which means strangers are being shown an opinion that has never been tested. Section 10 frames the two channels honestly, so this is defensible, but it is worth knowing.

Review 3 The example draft in the gate is mine.

"Load an example that fails" inserts a draft I wrote to trip two specific checks: the missing E5 RSCA line and the Ch. 13-5 medical clause. It is not from the repository fixtures and it is not a real Sailor. It reads as plausible eval prose, which is the point, and that is also the risk.

Review 4 The audience read is inferred, not researched.

The kitchen-table scene, the three-week window, the eight-to-twenty report load, and the "three people gave you three answers" line are all my construction from the domain material. They are plausible and they are not sourced. If Tim knows the real numbers, they should replace mine.

Review 5 Voice: this page does not sound like Tim's Slack voice, on purpose.

~/.claude/tim-writing-style-guide.md is a Slack and team-comms guide: warm greetings, exclamation marks. That register would be wrong in front of an E8 at 2100. The page mirrors the navy-evals repo's own prose instead. The mechanical rules from the guide were kept: no em dashes in the page copy, colons over dashes, active voice. The gate's own findings do contain em dashes, because those strings are quoted verbatim out of the product's config and rewriting them would make the page's output stop matching the CLI's.