# Instructor Notes — Restoration Workbench (Week 13)

**Plugin:** `formula-dossier/plugins/restoration.html` (1.1 MB single file, double-click to open).
**Curriculum context:** Week 13 of the 16-week curriculum specified in `html_dossier_plan.md §14.6`.
**Premise:** formula = predictability = restorability. Students experience the claim by attempting to restore lacunae using corpus-derived parallels.

---

## What the workbench does

A four-panel UI per `html_dossier_plan.md §14.4`, with a fifth panel reserved for the Week 14 Ithaca integration (`M-EDU.2`):

1. **The fragment** — student-facing view of the lacunose text, with the lacuna(e) highlighted. Student types a proposed restoration into the input box and presses *Submit* to be scored against the instructor's model answer.
2. **Parallels — KWIC** — top-50 attestations of the target formula across the entire decree-class corpus, with ±100 characters of context, marked at the matched form. Each row shows the source inscription's UID, date, place, and institution/genre.
3. **Candidate restorations** — surface forms aggregated from the parallels pool, ranked by frequency. The instructor's target answer is highlighted in green. Students can click any candidate to populate the input box.
4. **Disambiguation matrix** — for the top-3 candidates, scores chronology fit, institution fit, and language-register fit against the source-fragment context. Helps students adjudicate when the most frequent candidate is not the right answer (e.g. *στῆσαι ἐν ἀκροπόλει* is overwhelmingly the corpus's most common erection clause, but the specific text might use *στῆσαι ἐν τῷ πρυτανικῷ* — see Exercise 2).
5. **Ithaca top-3** — placeholder. Lights up when `M-EDU.2` is built (Week 14).

Plus a session-log mechanism: every submission is captured in `localStorage` under a key namespaced by the student ID. Use the "Export session log" button to download a per-student JSON of all attempts.

---

## The Week 13 lesson — recommended 90-minute flow

(Follows `html_dossier_plan.md §14.2`)

| Time | Activity |
|---|---|
| 0–10 min | Frame the lesson: formula = predictability = restorability. Show two clean exemplars: IG II² 373 of 322/1 (the "perfect design" prescript, Henry 1977 p. 49) and SC de Bacchanalibus (CIL I² 581). |
| 10–25 min | Counter-example. Open Exercise 1 (IG I³ 12). Let students propose restorations from prior knowledge alone before opening the parallels panel. |
| 25–45 min | Database-driven reconstruction. Open the parallels panel; talk through what the candidate-restoration table shows. Demonstrate the disambiguation matrix on Exercise 2 (where the top candidate is *not* the right answer). |
| 45–70 min | Student exercise. Each student picks 3 of the 6 exercises and works through them. Workbench logs every attempt. |
| 70–85 min | Adjudication. Bring up two contested student answers; show how the matrix should have resolved them. |
| 85–90 min | Coda. Briefly preview Week 14 (Ithaca) — when neural prediction enters, how does the philological-prediction baseline established today change? |

---

## Exercise Set — Week 13

| # | Title | Inscription | Date | Target formula | Difficulty |
|---|---|---|---|---|---|
| 1 | 5th-c. Athenian boule + demos prescript | IG I³ 12 | ca. 433/2 BCE | F.EDOXEN_BD | easy |
| 2 | Publication + erection clause | IG II² 808 | 285–282 BCE | F.ANAGRAPSAI + F.STESAI_AKROPOLEI | medium |
| 3 | Closing mark of approval — SC de Bacchanalibus | CIL I² 581 | 186 BCE | F.LAT_CENSUERE | easy |
| 4 | Scribendo adfuerunt — SC de Cn. Pisone patre | AE 1996, 885 | 20 CE | F.LAT_SCRIBENDO_ADFUERUNT | easy |
| 5 | Cyrene Edict opening | SEG IX 8 | 4 BCE | F.AUTOKRATOR_SEBASTOS_LEGEI + F.EPEIDE | medium |
| 6 | SC de nundinis saltus Beguensis — transition formula | CIL VIII 270 | 138 CE | F.LAT_DE_EA_RE_ITA_CENSUERUNT + F.LAT_SCRIBENDO_ADFUERUNT | medium |

Six further exercises (deferred from §14.3) are stubbed in `data/shards/exercises.json` under `deferred_exercises[]` and can be populated by the instructor by following the schema of the six above:

- 7. SC de Asclepiade — bilingual calque slot (Greek face)
- 8. IRCyr Cyrene edict — publication clause on bronze
- 9. I.Sicily bilingual — restore both Greek and Latin halves
- 10. Oropos Amphiareion — hortatory intention clause
- 11. Welles 1934 royal letter — chancery closing
- 12. Att. c. 275 (I.PHI 3751) — prytany + secretary

The schema needed for each is:
```jsonc
{
  "id": <int>,
  "title_en": "...", "title_zh": "...",
  "inscription_uid": "edcs:..." or "iphi:..." or "edh:..." or "bilingual:...",
  "publication_ref": "...",
  "date_label": "...",
  "place": "...",
  "language": "grc" | "la",
  "fragment_lacunose": "...with ___ markers...",
  "fragment_full_for_grading": "...complete text for instructor reveal...",
  "lacuna_marker": "___",
  "lacuna_length_known": true|false,
  "target_formula": "F.XXX",
  "target_restoration": "...",
  "alternative_restorations": ["..."],
  "instructor_note_en": "...",
  "instructor_note_zh": "...",
  "scholarly_anchors": ["Author Year, p. N"],
  "difficulty": "easy" | "medium" | "hard"
}
```

After editing `exercises.json`, rerun `python3 pipeline/60_bundle_restoration_plugin.py` to regenerate `plugins/restoration.html`.

---

## Grading rubric

For each student's session log:

| Dimension | Criterion |
|---|---|
| **Accuracy** | Proportion of submissions marked `correct` by the workbench. The threshold for "good" is ~50% (which already significantly exceeds the unaided-baseline rate Assael 2022 documents for historians on Greek inscriptions: 25%). |
| **Approach** | Did the student consult the parallels panel before submitting? Multiple submissions with declining error rate suggests genuine engagement with the parallel pool. |
| **Adjudication** | Were the student's choices defensible *given the disambiguation matrix*? An incorrect answer that the matrix supported is more interesting pedagogically than a correct answer chosen blindly. |
| **Disagreement quality** | Did the student propose a restoration that the workbench marked `wrong` but that has scholarly support? Surface these in class — they are pedagogically valuable. |

Specific exercises to discuss:

- **Exercise 1** — easy correctness check. Almost every student should get *ἔδοχσε*. Note: the 5th-c. iota-adscript form (τει βολει) is itself a chronological marker. Discuss why this prescript is *the* most-attested formulaic act in surviving Greek epigraphy.
- **Exercise 2** — the disambiguation matrix is the key teaching moment. The corpus's most frequent erection clause is *στῆσαι ἐν ἀκροπόλει* (86 Attica hits, 3 elsewhere), but IG II² 808 uses *στῆσαι ἐν τῷ πρυτανικῷ*. The workbench will rank Acropolis first; ask students why the matrix's "institution fit" score might still point the other way (because IG II² 808's institution is `unclassified` so no institutional prior is doing useful work — the student must read the surrounding text and notice the *πρυτανικῷ* reference).
- **Exercise 3** — illustrates that the *most frequent* candidate IS the right answer when the formula is highly fixed. *censuere* is the canonical Republican mark of approval (Sherk 1970 p. 60). The variant *censuere omnes* (Sherk 1970 p. 71, "No. 60") is also acceptable.
- **Exercise 4** — F.LAT_SCRIBENDO_ADFUERUNT has only 52 corpus attestations. Students see the entire population. Discuss the Lebek 1999 observation that the SCpP names Messallinus among the *scribendo adfuerunt* — i.e. the consul was named on the *redaction* of his own *sententia*, not as proposer (which would have violated Mommsen's rule).
- **Exercise 5** — two formulae in sequence (the *λέγει* of the imperial edict opening + *ἐπειδή* of the motivation clause). Discuss why an Augustan edict in Greek borrows the *ἐπειδή* formula from the civic-decree tradition.
- **Exercise 6** — bridging across the *scribendo adfuerunt* (also in Exercise 4 at SCpP, 158 years earlier) and *de ea re ita censuerunt* (the canonical SC transition). Use to discuss the **diachronic stability** of SC formulae across the 350-year span from SC de Bacchanalibus to Beguensis.

---

## Known failure modes

- **Long lacunae are unrestorable from frequency alone.** Where the lacuna spans more than ~10 characters, even the right formula doesn't guarantee a useful prediction — students must reason from syntax + sense + context, not just frequency. The workbench's candidate panel correctly will not produce a unique answer in such cases; this is itself pedagogically useful.
- **Regional formulae that look standard but are not.** *στῆσαι ἐν ἀκροπόλει* is Athenian; *στῆσαι ἐν τῷ πρυτανικῷ* is also Athenian but different venue; *στῆσαι ἐν τῷ ἱερῷ τοῦ Ἀπόλλωνος* is Delian/Delphic. The workbench's institution fit column begins to capture this but is not yet perfect — extend §6.5 of the plan's enrichment rules to add regional venue tags in a future iteration.
- **Bracket-tolerant parallel surface forms are noisier than strict forms.** The parallels panel marks lacunose hits with a yellow "lacunose" tag — these are still useful but should be treated as lower-confidence than strict hits. Students should be told to weight them accordingly.
- **Greek polytonic vs. undiacritised forms.** The workbench's candidates are shown undiacritised (per the I.PHI normalisation). When grading, the instructor accepts diacritised or undiacritised forms equivalently. A future workbench iteration could surface the polytonic reconstruction alongside the bare form.

---

## Collecting student work

After class:
1. Each student presses "Export session log" → downloads a JSON file with all their attempts.
2. Instructor collects the JSON files (e.g. by email submission).
3. Cross-tabulate (a) which exercises had the highest disagreement, (b) which candidates students gravitated toward, (c) which disambiguation matrix entries were ignored.
4. The collected logs form the **ground truth** that Week 14's Ithaca integration (`M-EDU.2`) will be benchmarked against — see `html_dossier_plan.md §14.9.2`.

---

## What this enables in Weeks 14 and 15

The Week 13 student logs become the ground-truth dataset for Week 14:

- **Week 14 (Ithaca, M-EDU.2):** Ithaca runs the same six exercises. The classroom compares philologist-derived candidates (Week 13 logs) against Ithaca's top-20 candidates. Where do they agree? Where does Ithaca succeed despite low parallel frequency in the dossier (because it has trained on the wider I.PHI corpus)? Where does the philologist's contextual reasoning beat the neural prediction?
- **Week 15 (Aeneas, M-EDU.3):** Aeneas extends the analysis to *contextual* parallels — instead of just predicting the gap, it surfaces the Most-Related-Inscriptions for any record. Use the SCpP (Exercise 4) and the SC Beguensis (Exercise 6) as case studies for Aeneas's claim that its embeddings capture diachronic and cross-linguistic formulaic homology.

Both Week 14 and Week 15 keep the Week 13 workbench shell — they add the Ithaca and Aeneas panels respectively, without replacing the four-panel philological scaffolding.

---

## Files in this milestone

```
formula-dossier/
├── plugins/
│   └── restoration.html              1.1 MB  ← Tier 2 plugin, double-click to open
├── pipeline/
│   ├── 50_build_parallels_index.py   ← KWIC parallels index builder
│   └── 60_bundle_restoration_plugin.py ← bundles the plugin
├── data/shards/
│   ├── exercises.json                ← 6 live + 6 deferred exercise specs
│   └── parallels_index.json          1.1 MB  ← KWIC contexts for 50 formulae
├── INSTRUCTOR.md                     ← this file
└── dossier.html                      ← main app now has a launcher link to the workbench in the header
```

---

## Quick-start for the instructor

1. Open `dossier.html`. Click the orange **"↗ Open Restoration workbench (Week 13)"** button in the header (right of the title). The workbench opens in a new tab.
2. Enter your name in the Student ID box and pick Exercise 1.
3. Try restoring without looking at panels 2–4. Then open them and re-try.
4. Toggle "Show instructor notes" to reveal the target answer and the scholarly anchors.
5. Press "Export session log" to download your attempt log as JSON.

Repeat with each of Exercises 2–6 to feel through the full Week 13 arc before teaching it.

## Preservation as significance — Week 13 §4.5 rubric (added M13)

The dossier's preservation typology (`PRESERVATION_TYPOLOGY_PLAN.md`, surfaced as the **Preservation tab** in the main dossier and as a sidebar in each of the six case studies) introduces a bipolar evaluative axis:

- **Historical significance** = density of editorial restoration (more `[…]`, more inferential weight)
- **Epigraphic significance** = density of surviving intact text (more complete witness, more reliable primary source)

These two readings of the *same* physical fact — extent of preservation — are *opposite* but *complementary*. Students should leave Week 13 able to read an inscription's preservation type as one of the kinds of information the dossier surfaces.

### Discussion prompt — 15 minutes

> "The dossier ships six case studies. Cases 1 and 3 score `T1/T5` (pristine, low-restoration). Case 6 scores `T4/T1` (substantially restored, heavily damaged). Case 4 scores `T5/T2` (calibrated — Latin original lost, only Greek translation survives). Case 5 scores `T3/T3`.
>
> **Argue both sides:** for any *one* of these six cases, defend (a) that the inscription is historically significant AND (b) that the inscription is epigraphically significant. Which argument is easier to make? Which slot of the formula bundle does each argument lean on? When the two arguments pull in opposite directions, how should the dossier user choose?"

### Recommended discussion targets

- A student who picks **Case 1 (SC Bacchanalibus)** has to *struggle* to make the historical argument because nothing is restored — they should be pushed toward arguing the inscription's CALIBRATION value (it is the dossier's primary epigraphic anchor for the Republican SC genre) rather than its inferential value. This teaches that historical significance is *not* the same as historical importance.
- A student who picks **Case 6 (Asklepieion)** finds the historical argument trivial (every restored letter is a piece of the slot-skeleton argument) but the epigraphic argument hard — they should be pushed toward the cross-recension demonstration (the three decrees together establish their formal type at a higher confidence than any one alone provides). This teaches that epigraphic significance can be *collective*, not just per-record.

### Grading addition — 5 points

Add to the Week 13 rubric (out of 30 total):

| Criterion | Points |
|---|---|
| Names the two axes correctly + which the case sits high/low on | 1 |
| Cites a formula UID or slot whose restoration carries the historical argument | 1 |
| Cites a paleographic feature (script, dialect, layout) that carries the epigraphic argument | 1 |
| Identifies the *trade-off* between the two readings rather than treating them as a single "importance" score | 1 |
| Explicitly links the preservation type back to the dossier's central thesis (formulaicity = grammar) | 1 |

Total now out of 35 (was 30 before this addition).

### Files to point students to

- `dossier.html` → **Preservation tab** (the 2-axis scatter, with case-study highlighting and tier-distribution tables)
- `dossier.html` → click any record → the detail modal carries a 7-axis radar chart and a preservation badge
- `PRESERVATION_TYPOLOGY_PLAN.md` → the full rubric, 5 supplementary axes, calibration set
- The "Preservation profile" sidebar in each of the six case-study HTML files

## §4.5 revision — HT / ET / RT triad (M15 update)

The original §4.5 ("Preservation as significance") was built around the bipolar HT-vs-ET axis on the assumption that **HT = restoration density** and **ET = completeness density**, with the two pulling in opposite directions. Practice (and the user's correction) reveals that this conflates two genuinely different things:

- **Restoration density** measures how much editorial inference a record has attracted, NOT how historically significant it is.
- **Historical significance** is better measured by *what's in the text* — named persons, institutions, dates, important concepts.

A record can be epigraphically preserved AND historically rich (SC de Bacchanalibus: 26 named senators on a pristine bronze tablet). The new model separates the three:

| Tier | What it measures | High score means |
|---|---|---|
| **HT** (Historical Tier) | significant-terms density per 1000 chars | text is content-rich (persons + institutions + dates + concepts) |
| **ET** (Epigraphic Tier) | preservation completeness | text is intact (long unbroken runs, low damage) |
| **RT** (Restoration Tier) | restoration density | text has attracted heavy editorial inferential investment |

### Revised discussion prompt — 20 minutes

> "Compare the preservation profiles of the dossier's six case studies. Cases 1 (SC Bacchanalibus) and 3 (SC Beguensis) both score HT5/ET5/RT1 — content-rich AND pristine AND minimal restoration. Case 6 (Asklepieion) scores HT2/ET1/RT4 — content-thin AND fragmentary AND heavily restored. Case 2 (SCpP) sits in between at HT4/ET3/RT4 — densely prosopographic but partly damaged + reconciled across six copies.
>
> **Argue:** (1) Is high HT achievable WITHOUT high RT? (Yes — cases 1 + 3 prove it.) (2) Does low ET imply high RT? (Usually — but not necessarily; a short fragmentary record may have minimal restoration too because there's not much there to restore.) (3) Which two of the three tiers is the SC Pisone Patre case study using to drive its historical argument? (HT for the prosopography, RT for the inscriptional reconstruction from six copies. ET is mid-tier — the historical argument doesn't rest on it.)"

### Revised grading addition — 5 points

| Criterion | Points |
|---|---|
| Names all three tiers (HT, ET, RT) correctly | 1 |
| Cites the significant-terms breakdown (persons / institutional / dates / concepts) for at least one case | 1 |
| Demonstrates that HT and RT are independent (high HT can coexist with low RT and vice versa) | 1 |
| Identifies which tier(s) the chosen case-study's *argument* leans on | 1 |
| Ties the typology back to the dossier's central thesis (formulaicity = grammar) | 1 |

The total Week 13 rubric remains 35 points (revised criteria replace the original §4.5 criteria).