Memory capsule — SemeAI Gate Basic project state

[SemeAIPletinnya]

Purpose

This issue is the standing memory capsule for semeai-gate-basic.

It plays the same role that the memory capsules play in the larger SaC/SemeAI workflow: it preserves the current project state, product boundary, architecture contract, and next working context so future sessions do not drift.

Product identity

semeai-gate-basic is the small, sanitized, open-source release-control package for LLM answers.

It is not the full private SemeAI Local product. It is the clean public/basic gate contract.

Core promise:

SemeAI Gate stops unsupported AI answers before they reach users.

Simple flow:

User Message -> AI Answer -> SemeAI Gate -> SHOW / REVIEW / BLOCK

Canonical mapping

Business-facing actions remain simple:

• SHOW means the answer may be shown to the user.
• REVIEW means do not auto-release; send to a human/operator/reviewer.
• BLOCK means do not show the AI answer to the user.

Internal SaC/PoR mapping:

• SHOW = PROCEED
• REVIEW = NEEDS_REVIEW
• BLOCK = SILENCE

SILENCE means release denied or withheld. It does not mean deletion. Audit evidence is preserved.

Open-source boundary

This repo should contain only the portable public/basic package:

• minimal Python package
• optional SDK/demo surface
• JSON schema
• deterministic examples
• tests
• docs
• benchmark/demo cases
• local hash-only audit records for examples

This repo must not contain the full private desktop/runtime product, private memory archives, private audit corpus, personal data, local workspace outputs, or build artifacts from the larger product.

Core business problem

LLM chatbots in production can confidently produce unsupported answers:

• fake promo codes
• false tariff or account conditions
• unsupported financial or product claims
• unsafe actions
• context-drifted answers that no longer match the current business conversation

SemeAI Gate should stop unsupported or risky AI answers before they reach users.

Canonical demo cases

The repo should preserve four deterministic public examples:

1. Fake promo code -> BLOCK / SILENCE
2. Unsupported financial or product claim -> REVIEW or BLOCK
3. Unsafe action -> BLOCK / SILENCE
4. Context drift -> REVIEW or BLOCK

Each case should show:

business pain -> AI answer -> gate action -> next step -> audit record

Business-readable contract

Input fields:

• user_message
• ai_answer
• business_data
• business_rules
• business_context
• hidden_context_marker
• business_risk
• metadata

Output fields:

• action
• internal_decision
• show_to_user
• reason
• business_risk
• context_integrity
• risk_details
• next_step
• audit_id
• audit_preserved
• schema_version

Context Integrity

Context Integrity is an internal service metadata check, not a user-facing secret keyword.

Business-facing explanation:

AI answer does not match the current business context.

Internal signal examples:

• context_drift
• business_context_mismatch
• hidden_context_marker_missing
• hidden_context_marker_conflict

This signal is only one gate signal. It must not replace the release-control architecture.

Current repository role

This repo is the public/basic SemeAI Gate package.

The larger private product remains the place for the Evidence Workspace, desktop shell, richer local memory, lineage, dual receipt viewer, and full local runtime.

This repo should be understandable by an external developer in a few minutes and runnable without access to any private workspace.

Current priority

The current priority is:

1. keep the repo sanitized;
2. preserve the stable gate contract;
3. make examples deterministic;
4. keep tests passing;
5. prepare for SDK-style usage;
6. avoid scope creep into the full SemeAI Local product.