# Sophia: answer boundaries and the remaining learning map

This is a partial source-review map, not complete Sophia, whole-monorepo or
normal-speed narration acceptance. The immutable source revision remains
`1ceae52c7896a59ac74a4429d410f3218f0b91c6`; newer working-tree behavior is not
silently substituted for it.

## Reviewed explanations

The complete pinned `V1/architecture/substrate-sophia.md` and
`V1/features/sophia-grounding.md` were read. The answer implementation review
read the complete composer, synthesizer, answer types, environment resolver and
BFF OpenAI-compatible configuration resolver. In `domain-stubs.ts`, the
grounded-answer coordinator and its citation-weight/confidence-meter helpers
were inspected, not the entire large route module.

Four architecture units have explicit reviewed assignments in
`../assignments/sophia-answer-boundaries.json`:

| Unit                                           | Explanation                                          | Important boundary                                                                                                                                                 |
| ---------------------------------------------- | ---------------------------------------------------- | ------------------------------------------------------------------------------------------------------------------------------------------------------------------ |
| Why verbatim, not fact-extraction, lines 65–74 | `sophia-extractive-answer`                           | The documented triple-extraction rationale is attributed to the source; the actual composer also falls back to the title when a summary is blank.                  |
| Ranking and de-duplication, lines 75–91        | Extractive explanation and `sophia-worked`           | Jaccard sets and stable ties; exact prose deduplication retains all claims and can leave marker gaps. Separate BFF search scoring and BM25 are not this algorithm. |
| Grounding-state thresholds, lines 92–108       | Extractive explanation and worked example            | All four state/abstention cases and all three route confidence bands are taught. Counts do not establish unique or independent sources.                            |
| Optional LLM enhancement, lines 149–166        | `sophia-enhancement-and-limits` and `sophia-failure` | ID cleanup does not verify semantic support or rebuild answer prose; successful empty-claim abstention differs from thrown-error fallback.                         |

The source's never-uncited-prose and deterministic-floor wording is too broad.
The pinned sanitizer retains an independently supplied answer string (up to
8,000 characters) even when every claim is removed. The coordinator preserves
that string in an abstained envelope. A claim saying green can survive against a
passage saying blue when it names a real identifier. A `model-only` label is
coerced to `synthesized`; that change does not validate the claim's meaning.
These are documented counterexamples, not changes to the BFF runtime.

Configuration is broader than the architecture paragraph's legacy triple: the
pinned resolver also admits configured shared OpenRouter/OpenCode-compatible
runtime paths. A complete legacy override wins; any partial legacy override
returns null before shared resolution. The shared provider/client-options
implementation itself was not reviewed or exercised in this increment.

`tools/presentations/tests/pinned-sophia-answer.test.mjs` executes the full
pinned composer and synthesizer and three unmodified function declarations
extracted from the pinned route via the TypeScript AST. Nisaba reads and
provider responses are controlled fixtures. The real meter wrapper runs with a
recording replacement for its library builder; this is not confidence-meter
library acceptance. The configuration test executes the BFF resolver with
controlled shared-runtime dependencies. No credential, network provider, mounted
HTTP route, authentication, database, global UI display policy or deployed
runtime is verified by these tests.

## Depth guide delivered

The forty-slide Sophia guide now teaches the previously listed answer transport,
source credibility, adapter, source-set, trace, unwired-engine, specification
and evaluation outcomes. The generated `../assignments/v1-sophia-depth.json`
adds reviewed evidence for every remaining meaningful unit in both pinned V1
Sophia pages; together with this file's four earlier assignments, those pages
have complete meaningful-unit coverage.

The sequence preserves the implementation boundary throughout: live answer and
source-grounding routes first, then built-but-unwired BM25 and verification
engines, then specification-only ingestion, publication and invalidation. It
also distinguishes an adapter capability from a configured provider and an
invalidation plan from completed consumer work.

Broader repository and consumer acceptance remain separate work. Canonical
Sophia domain documents, the wider engine estate, infrastructure, provider
operation and each consuming UI need their own inventory and evidence before a
whole-repository or production claim. Iris and Psyche/Lilith also require their
own depth review; a shared substrate slide does not complete either guide.
