Designing an AI that never states a fact — a deterministic engine, with the LLM only writing prose

Fortune telling is the sloppiest possible use of generative AI. Write something plausible and it reads as accurate, so when the model invents a fact, nobody notices.

That makes it an ideal place to try the opposite. Whether a reading is true is unfalsifiable — but whether the evidence the app cited actually exists in that user's data is decidable by machine. If you can refuse to publish prose that fails that check, the same structure moves to contract summaries, insurance and medical explanations, and internal-policy Q&A.

You cannot prompt your way out of hallucination. Tell a model to "answer only from the facts" and it will believe it is doing exactly that while describing a feature that isn't there.

Two genuine difficulties:

1. Deciding how much to move outside the model. Computing a Four Pillars chart is a pile of calendar-conversion rules. Hand it to an LLM and it fails quietly at the boundaries between solar terms.
2. Verifying the prose after you've let the model write it. Once you accept free-form natural language, correctness is normally no longer traceable.

I split it into three layers. The VLM only extracts structured palm features. A deterministic engine computes the chart. The LLM is permitted only to turn those into Japanese prose.

The calculator produces the answer; the AI only reads it aloud. The AI never does the math — so numbers that don't exist can't appear.

For verification, every generated section must supply evidence_refs. Those references are restricted to a canonical vocabulary such as lines.heart.tip, and prose citing a reference that does not resolve fails validation and is regenerated.

What I gave up was expressive freedom. Removing the model's room to "say something nice" leaves prose that is visibly constrained by structure. I chose verifiability over being a more entertaining fortune teller.

Solar-term data covers 97 years (1940–2036) to the minute, and the year and month pillars switch on the exact time of the term — on the day of 立春 itself, anything before that instant still belongs to the previous year.

Models are assigned by role: Gemini 3.5 Flash for palm-feature extraction, Flash-Lite for daily readings, Claude Opus 5 for paid deep reports. Every call — success or failure — writes one row to a cost ledger with the unit price snapshotted at that moment, so a later price change cannot rewrite historical COGS.

The spend guard has three layers: per IP, per verified user, and an account-wide ceiling — on the assumption that an IP can be spoofed or tunnelled.

Palm images exist only for the duration of the request and are never persisted, on device or server.

86 commits between 2026-07-21 and 07-29; 114,822 lines of ts/tsx/sql/mjs. 369 test files, with 100% coverage enforced as a threshold on every workspace. The database has 22 tables and 36 RLS policies, plus 3 tables that deliberately carry no policy at all — meaning anon and authenticated can read zero rows — and are reachable only by service_role. RLS itself is verified by 13 pgTAP suites.

CI is 7 workflows. One is security-only, bundling CodeQL, Gitleaks, Secretlint, a dependency audit, a license allow-list, SBOM generation, Trivy and OSV-Scanner.

If you want generative AI in production somewhere a single wrong sentence becomes an incident, it has to look like this: fact computation outside the model, the model writing prose only, and a machine checking the evidence that prose cites. Contract summaries, insurance and medical explanations, internal-policy Q&A — they all fit the same shape.

The CI checks for regulated Japanese advertising and pharmaceutical-claim wording are implemented too, and portable as-is. The interesting part is that they are denial-aware: an article is allowed to quote a superstition in order to refute it.

I take on this kind of generative-AI work as modern / AI development. See also the client engagement running a production voice AI.