# Sophia · One question, and what the answer may claim · receipt · 12 September 2026

Chapter four of _Shared systems_, delivered in the Eve edition on 12 September
2026. Forty inherited slides become a forty-eight-slide chapter under the
`sophia` id. The brief is
[authoring/sophia-chapter.md](../../authoring/sophia-chapter.md). Machine-readable
state is in [state.json](state.json).

## What was delivered

Forty-eight slides, all authored here. All forty inherited slides are rewritten
in place and nothing merges away; eight are new — a cover, six section dividers
and a close.

The forty arrived spread across four manuscripts.
`content/71-v1-sophia-depth-a.json` and `content/72-v1-sophia-depth-b.json` held
nothing but Sophia slides, so both empty and retire.
`content/06-evidence-memory-runtime.json` loses five slides and
`content/23-shared-deepdives.json` three; both keep everything else.

## Evidence

No capture. Sophia's only direct surface is a workbench and the chapter's
subject is an envelope, so every claim is executed or read at the pin.

- **The extractive composer, executed here.** It tokenises by lowercasing,
  splitting on anything outside letters and digits, and dropping thirty-three
  stop words, then orders citations by Jaccard overlap with the query. Verified
  deterministic: reversing the input returns the same ranking, and ties keep
  input order. Each citation yields exactly one `retrieved` claim whose text is
  the summary verbatim, or the title when the summary is empty. Identical
  summaries produce one rendered line and two claims, each still tracing to its
  own source.
- **The grounding state is a count.** Executed here, three or more citations
  gives `grounded`; one or two gives `partial`; zero abstains, splitting into
  `ungrounded` for a real question and `abstained` for a blank one. Three
  citations about bicycle maintenance, sourdough and tax law, against a question
  about solstices, produced `grounded` — every lexical overlap score was zero.
  The word reports how much evidence was attached, not whether it bears on the
  question.
- **The response sanitiser, executed here.** A claim citing an identifier that
  does not exist is dropped; a claim citing nothing is dropped; a claim citing
  one real and one invented identifier survives with the real one; any label
  other than an exact `retrieved` is coerced to `synthesized`, `model-only`
  included. A code fence is stripped and non-JSON throws, which returns the
  coordinator to the composer. When every claim is dropped the result abstains
  **and the model's answer string is still returned** — a well-formed envelope
  carrying confident prose, no claims, and a confidence band computed from the
  citation count before synthesis ran.
- **Fail-closed configuration, executed here.** An empty environment, a key
  without a model, and unrelated variables each resolve to no synthesizer at
  all, and the route uses the deterministic composer instead.
- **The source scorer, executed here.** Four weights — 0.4 for the domain tier,
  0.2 each for recency, citation volume and peer review — summing to exactly
  one. Six tiers from 0.95 down to 0.3, with an unknown domain at 0.5,
  deliberately above a known-poor one. A hundred named domains across four sets.
  Citation volume saturates at exactly 1,000: ten thousand and a million score
  identically. Two different pages on the same host score identically, because
  the tier is a property of the host.
- **No reputation clears the floor, measured here.** The floor is 0.5, and the
  most authoritative tier alone reaches 0.38. Every tier fails unaided —
  academic 0.36, reputable 0.30, standard 0.24, unknown 0.20, low 0.12. One
  maximal second signal takes the authoritative page to 0.58; an unknown domain
  needs all three. Corroboration is arithmetically required rather than advised.
- **The BM25 index, executed here.** Canonical parameters — k1 at 1.2, b at 0.75
  — with field weights of 2.4 for a title, 2 for keywords, 1.6 for a summary and
  1 for body or any unnamed field, 28 stop words, a two-character token floor
  and accent folding. None of it is used by the answer route, which orders by
  Jaccard overlap and has no term frequency, document length or field weighting
  in it at all.
- **Read at the pin.** The coordinator calls one corpus adapter with a limit of
  six and catches any retrieval exception into an empty result set. Two envelope
  vocabularies differ by a single character — `retracted-source` in the answer
  contract, `retracted_source` in the adapter — and a third vocabulary,
  `grounding status`, describes a different thing entirely and is the only one
  that can express conflict. The fact-check loop's five thresholds (0.45, 0.55,
  0.18, 0.68, 0.42) and three latency budgets (6s, 90s, 15min) and the
  contradiction loop's five (0.18, 0.18, 0.62, 0.42, 0.25) were all read at the
  pin; the answer path imports neither loop.

### What this chapter's evidence tops out at

Executing a pinned module proves what that code does with the inputs it was
given. Reading a route proves what an author wrote. Neither establishes that a
model was called, that a provider responded, or that any member saw anything.
Two files in the shared AI library have moved since the pin —
`libs/shared/ai/src/types.ts` and `providers/openai.ts` — and the probe records
that drift rather than hiding it; the run still executes the pin. Provider SDKs,
a logger and a Redis client were stubbed at their boundaries so the module graph
would load; each stub throws rather than returning a fabricated result, and the
repository's own configuration resolver and response sanitiser stayed real,
because they are what was under test.

## Checks

| Check         | Result                                                              |
| ------------- | ------------------------------------------------------------------- |
| Build         | 1,674 slides, 713 sources; freshness check clean                     |
| Python suite  | 490 tests pass                                                       |
| Node suite    | 508 tests pass                                                       |
| Browser sweep | 48 slides × 1440, 1280 with audio chrome, 390 and print — 0 findings |
| PDF           | 48 pages, exact authored visual text; the chapter has no table       |
| Narration     | 48 tracks, 69.0 min; full library `--check` verifies 1,674 tracks    |
| Bindings      | 49 reviewed source units, regenerated and re-fingerprinted           |

The sweep passed on its first run. The previous chapter's overflow findings had
already taught the two causes — Eve card grids hold three or four nodes, and
diagram grids are zero-based — and those were applied while authoring rather
than discovered afterwards.

## Binding work

Thirty-two of this chapter's slides are cited in
`assignments/v1-sophia-depth.json`, which is generated. The regenerated file
holds forty-nine units, the same count as before the rewrite.

Getting there needed a fix to the assembler. It restores each slide's original
heading citations and then folds duplicate citations so each path appears once —
but the fold preferred whichever form came first, so a bare path string could
displace the heading citation that had just been restored onto it. Two units
came unbound that way before the fold was changed to prefer the heading form and
to replace a bare path in place. The same bug would have been silent: the slide
reads identically either way.

## Tests updated

Six, each with the reason written beside it.

Two had a named specimen that this rewrite moved to another layout — and both
had already been repointed once for an earlier chapter, so this time they were
made specimen-agnostic instead:

- `test_ownership.py` took `sophia-live-flow` (and `architecture-read-handoffs`
  before it). It now takes whichever `ownership-handoffs` slide the library
  ships and asserts against that slide's own owner and exchange counts. Two
  assertions that encoded the old specimen's shape became structural: exchange
  placement is counted rather than requiring both directions, and the orphan
  case strands the least-connected owner so the edge list stays non-empty.
- `test_composition.py` took `sophia-scoring-model` (and Meet Lilith's four
  evidence questions before it). It now takes a four-node `evidence-comparison`
  slide from the library and reads node ids and the caption from the specimen.

The other four: the Sophia assignment test (slide count 40 → 48, and the
sequence now bounded by the cover and close); the rollback guard (`sophia` moved
from `APPROVED_DEEPENED_GUIDES` to `REWRITTEN_CHAPTER_GUIDES`); the
tracked-chapter count, 1,151 → 1,199; and `pinned-sophia-answer.test.mjs`,
repointed at the new manuscript after the teaching it guards was restored to the
rewritten slides — the count caveat back onto the composer slide where the rule
lives, and the fall-back wording onto both slides that describe it.

## Treatment deviations

Nine of the forty inherited slides were tables and the plan allowed a table
where readers need exact values side by side. None of them needed one: every one
became a card grid, a compare panel or a stat panel, and the chapter ships with
no table. Beyond that, three slides depart from the layout the plan named:

- **`sophia-scoring-model`: card grid → `stat-panel`.** The teaching turned out
  to be four executed measurements — the weights, the tier table, the domain
  count and the saturation point — which is what a `measured` stat panel is for.
- **`sophia-live-flow`: `ownership-handoffs` → `sequence-lanes`.** Four lanes
  exchanging in a fixed order is a sequence; the ownership layout does not carry
  the ordering.
- **`sophia-bm25-engine`: card grid → `stat-panel`.** Same reason as the scoring
  model: k1, the title weight, the stop-word count and the zero uses on the
  answer route are all measurements taken here.

## What this chapter does not establish

No model was called, no deployed service was reached, and no member data was
used. Nothing here shows what any interface displays, which the chapter states
on its own slides rather than only in this receipt.

## Correction (12 September 2026)

`state.json` claimed that `check-center-coverage.py --check` reported no stale bindings, and its note
said the regenerated assignment held fifty-eight units. Both were wrong: the file holds forty-nine, the
figure the checks table gives, and the sentence was carried over from the previous chapter's receipt.
The generated assignment was clean, but rewriting the answer slides left all four reviewed bindings in
`assignments/sophia-answer-boundaries.json` stale. They were re-reviewed later the same day with the
Iris chapter. One moved: the grounding-threshold unit now cites `sophia-thresholds`, the slide that
actually carries the thresholds after the rewrite, and that slide gained the architecture-page citation
it teaches from.
