Destination web provision, Indonesia & Montenegro

Model identity is taken from the gateway's configuration, not from the models' own statements — and that is not pedantry here, because both hosted coders misidentified themselves when asked. One answered “I'm Claude, made by Anthropic”; the other “Built by OpenAI, GPT-4”. Since coder independence is the entire basis of this check, a study that establishes it by asking the coders who they are has established nothing.

Read the V1 block as a diagnosis, not a disappointment. If the construct “can a visitor transact online” were genuinely ambiguous, the three LLM coders would disagree with each other too. They do not — they reach κ = 0.897 to 1.000, and a 27B open-weights model running on a laptop reproduces a frontier model's judgement exactly. Every one of them agrees with the regex instrument at chance. The construct is highly codeable; the instrument was not a valid measure of it.

Where the instrument went wrong — three error classes, all visible in the disagreement set. Missed transactional language: “Elektronske ulaznice” (Montenegrin e-tickets) on nparkovi.me, “Beli TLPJL” on Raja Ampat's site, the eRinjani booking system. Keyword without a route: a “Reserve” menu item, the word “reservation” inside a policy, a “checkout” string in a script — six domains scored 1 on that basis and 0 by all three LLM coders. Absence read as evidence: where a page delivered no text, the instrument recorded “no booking”; it had no “cannot determine” category at all, and 28 of 142 panel cells belong in one.

Live cross-section, 33 domains, 4 September 2026. The three valid coders find a larger gap than the broken instrument did, in the same direction, at p ≈ 0.06 on a single year. So after the withdrawal the hypothesis was unmeasured, not refuted — which is what made re-running the full panel worth doing.

Panel, 142 archived domain-years, all re-coded by Qwen3.8-27B locally; 142/142 parsed, zero failures. The regex instrument roughly doubled both rates and inflated the gap. Note that the direction of the correction reverses between the two datasets — larger gap on the live pages, smaller across the panel. An invalid instrument does not err consistently, which is precisely why its output cannot be trusted even when it happens to point the expected way.

Neither country is adopting. Indonesia sits higher in every year under the valid instrument, but online booking capability on official destination sites is not spreading over 2022–2026 in either system. That flatness runs against the assumption that destination e-commerce steadily diffuses, and is worth reporting in its own right.

What may not be claimed. Three models agreeing establishes consistency, not accuracy — they may share a bias. That the LLM coding is correct and the regex wrong needs a human-coded subsample, which has not been run; adjudication_worksheet.json is prepared for it (~26 domains). What holds without any ground truth is the pair of measured facts above: the instruments disagree at chance, and the conclusion flips depending on which one is used.

One of three carries a claim. H2's row shows the re-measured figures, not the withdrawn ones; no multiple-comparison correction applies to it because it no longer enters the family as a positive result. H1 survives Benjamini–Hochberg, the pre-registered rule, and fails Bonferroni — which is stated here rather than omitted.

Reliability precedes significance. The pre-registered rule, written before any coder ran: if κ for the booking variable falls below 0.60, the finding does not stand whatever its p-value. It came in at 0.043. That rule is why a p = 0.026 result was withdrawn rather than defended, and it is the part of this project most worth copying.

H1 remains fragile. A single reclassification — one presence date read differently — pushes it to p = 0.085. Expanding the frame toward ~30 domains per country is the remedy, and on that sample H1 and H3 are genuinely pre-specified by 04_analysis_prespecification.md.

The archive-coverage confound was tested, not assumed. Montenegro is archived more densely than Indonesia (median 46 vs 32 snapshots), so an absent capture could have meant an unarchived site rather than no site. Domain registration dates settle it from outside the archive: all four Indonesian domains the archive dates to 2023–24 register in the same window, and Raja Ampat’s kkprajaampat.id — live, serving HTTP 200, registered 2026-05-16, not yet crawled — is a fifth case the archive alone would have missed.

Montenegro's single decline is kotor.travel — the only domain in either country to lose presence. It is still live; it migrated to a client-rendered single-page app and became invisible to archival crawlers. Destination digital heritage is being lost silently.

HTTPS is shown for completeness but Montenegro's apparent decline is an artefact: the measure is the URL scheme the crawler happened to archive, and a site reachable over both schemes can be archived under either. Do not report it as a finding.

Six captures returned exactly 73 characters under the title “One moment, please…”. Without this screen they would have entered the content series as near-zero, and a site that started blocking crawlers would read as a site that lost its content. 18 of 142 captures are interstitials or empty shells, and the share is rising in both countries.