How The Taste of Words was built

This record reconstructs the decisions registered during edition 002. Version `1.5.2` governed both the narrative method and the format of this public explanation. The complete artefacts remain private: what follows provides public traceability, not an external audit or an objective measurement of literary quality.

Scientific starting point

The source was version 1 of Lindsay Blackwell’s preprint *Content Moderation Futures*. The fidelity check confirmed identity, version, method, declared results, the author’s interpretation, and limitations. It neither reproduced the study nor independently validated its scientific conclusions.

The qualitative study brought 33 professionals together in six participatory workshops. It recorded tensions between contextual, transparent, accountable moderation and pressures created by scale, automation, underinvestment, and growth. The sample was mostly Western, white, and highly educated, with limited representation of more precarious workers. The story retained that boundary and did not present the findings as universal rates.

Alternatives considered

Six possibilities were developed. The selected one turned sensory associations supplied by moderators into traces reused by a global AI. The finalist made the hidden labour signatures behind every automated decision visible as shadows.

The other four explored a city without irony, working copies of dead moderators, six governance principles distributed between districts, and appeals paid for with autobiographical memory. They lost on relative fit: risk of repetitive procedure, displacement towards the consciousness of copies, excessive conceptual load, or weaker focus on moderation labour.

Prose probes

Two opening scenes of 350–450 words were written. Both passed B1 and B2: someone wanted something concrete, an opposing force acted, and the scene ended in a materially different state.

The selected probe moved from a routine decision to discovering retention and receiving a coercive offer. The finalist produced a strong image but left its protagonist without an equally concrete next action. The candidate was therefore not switched.

Adversarial-role evaluation

The first complete version received `APROBADO_CON_CAMBIOS` (“approved with changes”). It had no critical failures, but the climax omitted the exact moment when Atlas Global became active, and the workers’ unanimous final response simplified the conflict.

The model-assisted evaluation adopted an adversarial role: it looked for localisable contradictions and weaknesses. Its scores were editorial judgements, not objective measurements or independent human review.

Material changes

There was one full rewrite and no redesign. The final version:

• shows Jonás bringing the launch forward as a counteraction;
• replaces a clause added that same day with ambiguous consent covering “derived features”;
• replaces unanimity with a partial coalition and visible disagreement;
• reduces percentages and retains a single decisive threshold;
• requires Jonás to sign the suspension and assume responsibility for delays in provisional mode.

The opening taste of burnt milk, the offer of an individual exit, the appeals channel, the real threat, the bodily burn, and the final tasteless word were preserved.

Causal summary and removed elements

An AI copies sensory associations to interpret ambiguous language. Alba discovers that these traces will become infrastructure and tries to revoke them. Jonás brings the launch forward to contain real threats. Alba opens a collective withdrawal and burns her interface: she stops the global model but loses the ability that sustained her profession.

The rewrite removed the overly explicit contractual fraud, the unanimous burning of all thirty-two Traces, and a sequence of percentages that made the climax harder to follow.

Registered controls

• CF1: the novum appears in the first sentence, is identified as an artificial-intelligence interface, and causes several observable changes.
• CC1: the isolated evaluation reconstructed `novum -> objective -> decision -> cost` in ordinary language.
• T1: Jonás protects immediate safety and reduced exposure; Alba protects control over context. Both exercise power and pay a price.
• B2: both probes changed state within the scene; the selected one left a more concrete next action.
• J1: discovery, offer, isolation, revocation, early launch, forced cooperation, and the burn produce qualitative changes.
• L1: the isolated evaluation identified concrete expectations at 25%, 50%, and 75%, and an ending that pays off the planted Trace. L1 does not measure real-reader behaviour or desire.
• Bilingual QA: the adaptation preserves facts, figures, opposition, cost, and outcome.

Limit of this record

This explanation discloses the source, alternatives, probes, verdict, changes, and registered results. It does not publish complete drafts or internal instructions. Nor does it prove that the story will be enjoyed, that the evaluations are objective, or that the paper is scientifically valid beyond its own scope.

[Return to the story](/en/stories/the-taste-of-words) · [General methodology](/en/methodology) · [Versión en castellano](/relatos/el-sabor-de-las-palabras/proceso)