Sol

Sol is a research-first, local prototype for building a user-consented, evidence-backed, and user-correctable personality and style model for personalized generation.

It does not claim to detect a person's true personality, read minds, or provide clinical diagnosis. Assessment results are treated as self-report evidence, not identity facts.

Current product thesis: authorized evidence becomes inspectable provisional profile atoms; users correct and scope those atoms; only eligible atoms may influence a specific generation task.

The first product wedge is the Personal Creative Style Kit: combine writing samples, preferences, examples, and direct calibration into an editable style profile, then generate text and visual direction the user can evaluate.

The working implementation is currently a narrower assessment-first foundation. It proves this local loop:

consent
  -> OCEAN self-report assessment
  -> validated responses and scores
  -> provisional profile atoms
  -> evidence inspection and user correction
  -> lifecycle/scope review
  -> generation-safe context export
  -> model-free writing-guide prompt dry run
  -> structured feedback notes

The next product slice is a faster participant-link validation MVP: share a no-auth link, assign an opaque participant ID, show a bounded fictional scenario, collect the participant's organic written response, run an assessment, generate an assessment-informed predicted response that excludes the organic response, and ask whether the prediction sounds like the participant. The full Style Kit guidance/source workbench follows only after this first validation slice is interpretable.

The planned hosting path for that slice is Vercel plus hosted Postgres. Vercel can host the public participant UI and stateless API routes, but the current local JSONDB-backed server should not be deployed unchanged.

The local assessment MVP supports:

• 11 permissive/open OCEAN and OCEAN-adjacent instruments;
• browser administration, autosave, resume, scoring, and result display;
• persisted consent plus instrument/scoring fingerprints for new sessions;
• response ID/range validation and protected completed-session evidence;
• provisional, non-diagnostic profile atoms;
• confirm, reject, review-only, claim-edit, and user-note controls;
• immutable original claims and timestamped review history;
• expandable item, keying, score, source, reliability, and uncertainty evidence;
• whole-session, raw-response-only, or derived-atom-only deletion;
• scoped profile context export with ineligible atoms excluded by default;
• a model-free writing/communication guide prompt dry run;
• structured generation feedback linked to inspectable mapping notes;
• serialized and atomic local JSONDB mutation for the single-process MVP.

The Creative Style Kit foundation now also includes five versioned record schemas, a cross-linked example bundle, an offline validator, an ADR covering ownership/consent/retention/deletion, and a contract-validated atomic local repository. The repository defaults to ignored tmp/ storage and exposes no UI or network service yet. A guidance-domain service now enforces explicit user review, immutable original wording, prompt-safe edit history, evidence eligibility, and disable behavior.

Persisted pilot runs now support credential-free dry-run:// and deterministic mock:// providers. Each run stores generic and personalized variants, exact prompt-safe guidance snapshots, task/context/request hashes, provider versions, and bounded output validation results.

Evaluation events now enforce a persisted two-step blinded lifecycle: the opaque choice is recorded before identity reveal, and later feedback is bound to the exact run and guidance actually used. Events are exportable and independently deletable without changing assessment evidence or guidance.

The current automated suite has 61 tests, plus rendered desktop/mobile QA for the Administer and Workbench views.

The repository does not yet provide:

• writing-sample, liked/disliked-example, or moodboard ingestion;
• browser-based contextual generation-guidance authoring;
• browser guidance, run-history, blinded comparison, and evaluation surfaces;
• participant-link scenario validation UI;
• model-backed artifact generation;
• “feels like me,” usefulness, or generic-baseline evaluation reporting;
• a validated original Sol personality assessment;
• production authentication, authorization, multi-user storage, or deployment;
• multimodal generation adapters.

Model-backed generation is intentionally unapproved until the entry conditions in generation-contract-review.md are satisfied.

The repository contains one project-authored design-review candidate:

• Sol-OCEAN-Quick-v0
• 30 original candidate items
• five OCEAN domains
• 15 candidate subconstructs
• balanced positive/negative keying
• explicit sensitivity and expected-failure metadata

The candidate is stored at assessments/ocean/experimental/sol_ocean_quick_v0.json. It is excluded from the administrable manifest, and the product scoring/session boundaries reject it.

Passing repository validation means the candidate is ready for independent review—not that it is reliable, valid, normed, invariant, or product-ready. Expert review, cognitive interviews, a pilot ADR, and empirical validation remain required.

Validate the candidate:

python3 tools/validate_sol_ocean_candidate.py

As of 2026-07-12:

See docs/current-state.md for the maintained handoff and docs/audits/2026-07-12-initial-plan-progress-audit.md for the full initial-plan audit. The complete current inventory is docs/inventory/2026-07-12-capability-inventory.md.

python3 tools/assessment_web_mvp.py --port 8765

Open http://127.0.0.1:8765.

Run automated QA:

./scripts/run_assessment_web_mvp_qa.sh

Capture isolated desktop/mobile visual QA:

./scripts/run_assessment_web_mvp_visual_qa.sh

Visual artifacts are written below ignored tmp/assessment-web-mvp-visual/ and do not modify tracked assessment storage.

After at least one atom has confirmed/edited feedback, active_atom state, and contextual or global scope:

python3 tools/generation_pilot.py \
  --pilot-id writing_guide_001 \
  --generated-at 2026-07-12T00:00:00Z

This renders the prompt and eligible context only. It does not call an external model.

The active product roadmap is plans/14-personal-creative-style-kit-roadmap.md.

Current order:

1. Build the participant-link validation MVP in plans/16-participant-link-validation-mvp.md.
2. Add the Vercel/Postgres hosting path in plans/17-validation-mvp-hosting.md.
3. Prove the predicted response uses assessment-derived candidate context while excluding the participant's organic response.
4. Store alignment/misalignment feedback as evaluation evidence, not as an automatic rewrite of assessment claims or confidence.
5. Run a small consenting pilot and decide whether the output sounds like users.
6. Expand into fuller Style Kit guidance, source intake, visual inputs, or platform extraction only after the validation gate is reviewed.

Independent expert/cognitive review of the experimental Sol OCEAN candidate is a parallel validation track, not a dependency of the product wedge.

Detailed ledgers:

• plans/11-mvp-hardening-and-profile-loop.md
• plans/12-post-audit-roadmap.md
• plans/13-sol-ocean-experimental-assessment.md
• plans/14-personal-creative-style-kit-roadmap.md
• plans/15-style-kit-validated-execution.md
• plans/16-participant-link-validation-mvp.md
• plans/17-validation-mvp-hosting.md

Blocked by default:

• clinical diagnosis or personality-disorder labeling;
• protected-class inference;
• hidden profiling;
• covert political persuasion;
• employment, housing, lending, insurance, legal, or education eligibility use.

High-sensitivity claims—such as trauma interpretation, emotional instability, honesty/exploitiveness, attachment, intelligence, or competence—must not be casually inferred or used as ordinary generation controls.

The repository includes reviewed source cards, metadata-only paper imports, Wikipedia background summaries, model contracts, assessment catalogs, and a local lexical RAG index.

Build the index:

python3 tools/rag.py index

Search or print context:

python3 tools/rag.py search "construct validity personality profile"
python3 tools/rag.py context "why broad traits should not control generation" --top-k 6

Metadata imports are background references, not reviewed knowledge. Research changes reach product behavior only through reviewed cards, model/contract updates, and ADRs where appropriate.

• app/assessment-mvp/ — local browser UI.
• assessments/ocean/instruments/ — administrable permissive/open instruments.
• assessments/ocean/experimental/ — inactive project-authored candidates.
• kb/cards/ — reviewed research source cards.
• kb/model/ — knowledge, signal, and profile-atom contracts.
• kb/assessments/ — assessment catalogs, mappings, and construct blueprints.
• docs/architecture/ — implementation and research architecture.
• docs/adr/ — architecture decisions.
• docs/audits/ — repository plan/progress audits.
• plans/ — committed product and research execution ledgers.
• jsondb/ — local queues and MVP persistence.
• tools/ — RAG, import, assessment, validation, and pilot CLIs.
• .codex/skills/ — repo-local reusable agent workflows.

• docs/project-memory.md
• docs/current-state.md
• kb/model/profile_atom_schema_v0.md
• docs/architecture/assessments/web-mvp.md
• docs/architecture/assessments/sol-ocean-experimental.md
• docs/adr/2026-07-12-style-kit-record-contracts.md
• docs/adr/2026-07-13-participant-link-validation-mvp.md
• docs/adr/2026-07-13-vercel-validation-mvp-hosting.md
• docs/architecture/rag/research-promotion-workflow.md