Language Design Evidence

Status. A bounded evidence check of two shipped language-decipherment games against the team's language-play proposals. Sources are developer statements in official material and interviews; secondary summaries did not fill gaps. Nothing here shows that our version will be fun, readable or long enough. Three kinds of claim are kept apart: an implemented mechanic the developer describes, a stated intention, and empirical evidence from players. Neither game's sources supplied the third beyond developer impressions.

Heaven's Vault (inkle, 2019)

Sources: inkle's official page; Game Developer interview, 17 April 2019 (Jon Ingold, Joseph Humfrey); MCV/DEVELOP interview, 16 April 2019 (Ingold, Humfrey); TheSixthAxis interview, 27 April 2018 (Ingold); PCWorld preview, 18 April 2018 (one Ingold quote, otherwise a journalist's account).

Implemented, as the developers describe it. Inscriptions are broken into words and the player selects, for each word, from a small set of English candidates the player character offers; words are built from smaller "atoms", so related symbols hint at connections already made (Game Developer; inkle page). The game never says whether a choice is right: "What we don't tell you is if you're right" (Ingold, MCV). A guess is fed back into the story and into later translations, and a wrong idea "can feed into the narrative" (Ingold, MCV). There is no fail state; the design "encourages correct interpretations or eliminates options" as a player drifts (Game Developer paraphrase). Inscriptions are generated to match each player's dictionary progress and location (Game Developer). Ingold expects players to be confident about roughly a third of the language by the end (TheSixthAxis). The player still decides which candidate to commit, whether to trust an earlier guess when a word recurs, and what a place therefore is; the companion's hinting role is not described in these sources.

Stated intention. To feel like translating rather than to model a language: the "Guitar Hero of linguistics" (Humfrey, Game Developer); ownership through uncertainty (Ingold, TheSixthAxis).

Empirical evidence. Only a developer impression that guesses are "usually pretty good" (MCV paraphrase). No published player data.

Transfer limit. The uncertainty works because a story absorbs any guess at no cost and without a clock. In Hold The Horde a reading is spent under oxygen and horde pressure and can fail physically, and a fixed station cannot generate inscriptions to fit progress; placed redundancy must do what inkle's generation does.

Chants of Sennaar (Rundisc, 2023)

Sources: Game Developer interview, 29 February 2024 (Julien Moya); developer store description; Rundisc's page. Two further interviews (80.lv, Pocket Tactics) contained no mechanic detail.

Implemented, as the developer describes it. A redundancy rule: "every word you encounter in the game must be seen at least twice in two different situations" (Moya, Game Developer). Players are not expected to understand a word at once; they "jot down a few hypotheses and keep exploring" in a notebook (Moya; store text: "note each symbol, keep track of your surroundings"). Two systems are named but not explained in the reachable sources: an "Obra Dinn-inspired validation system" and "automatic translation of complete sentences" (Moya, Game Developer). The widely reported picture-page batch confirmation could not be verified from a developer source within the cap and is treated as unverified. Context comes from people speaking and acting.

Stated intention. "The feeling of the traveler who explores countries where he or she doesn't speak the language" and understands "through cultural immersion and practice"; "confronting the player with seemingly impossible challenges while discreetly giving them all the tools and clues they need" (Moya, Game Developer).

Empirical evidence. None in the sources; the systems emerged from development trial and error, not reported player data.

Transfer limit. Sennaar's context is living speakers in safe spaces. Our languages have no speakers, and observation costs oxygen. A batch-validation system also grants a certainty our proposals deliberately withhold; adopting one is an explicit change of stance, not a detail.

What this says about the released positions

"Every inference testable", "costs little", "cannot be solved by dialogue". Heaven's Vault is a working design in which inferences are not testable at all; the story absorbs them. Testability is a preference that presumes a machine response for every sign; no source supports it as a requirement. "Costs little" is contradicted by the mechanic taking inkle "about a year and a half" (Ingold, MCV) and Rundisc "extensive trial-and-error": the interface and consistency are the cost, not the instances.

"Any two comparable instances plus one that breaks the reading will do." Sennaar's stated safeguard is a per-word rule of at least two situations, and Heaven's Vault compensates for misses by generation. Neither shows that two-plus-one guarantees recovery, which still depends on the player noticing the exception. Context-dependent signs also still yield reusable rules: inkle's atoms are precisely reusable context, so an indexical language does not exclude a lookup.

"A hint, match or confirmed gloss removes interpretation." Heaven's Vault hands the player candidate words for every inscription and still leaves choice and consequence; Sennaar's automatic sentence translation is assistance the developer describes as part of the puzzle, not its removal. Help does not by itself dissolve interpretation, and required help is a gate only where a policy makes it one.

One inherited example should be dropped: a ledger tally matching the player's death count assumes a count-based mechanism no direction has adopted.

Design implication and the open choice

Both shipped designs give the player something firmer than open inference: candidates to choose among, or validation once enough is consistent. Both survive missed clues by redundancy or generation. In a fixed station priced in oxygen, the transferable safeguard is a stated per-sign redundancy rule, and the paused journal can present a sign's own recorded instances side by side without confirming a meaning. The choice the sources cannot make for us: withhold confirmation entirely and let readings feed consequences, or confirm in batches once enough is consistent. The first fits discovery first and personal ownership; the second gives certainty and pacing. Neither source shows which holds up when reading is done under threat.