manga-recap-pipeline/ARCHITECTURE.md

# ARCHITECTURE

The target shape of the pipeline. Written 2026-08-13 from the user's design, rewritten the same day under
the no-Magi constraint. This is **not** what the code does. `NEXT.md` holds the live state and `AUDIT.md`
holds the current pipeline. Every section ends with what exists today and what would make it done. The
gap stays legible and testable without reading both.

The governing principle:

> Do not make the next panel understand the previous panel. Make it understand the current world state
> produced by all previous panels.

Vision produces observations. A persistent chapter graph owns identity and relationships. Everything below
follows from that split.

## Ruled out

Read this before proposing any of it again.

| ruled out | why | who decided |
| --- | --- | --- |
| Magi, Magiv2, Magiv3 as a detector or as pair heads | project constraint, not a measurement | user, 2026-08-13 |
| A trained `t2c` pair model, or any learned head | no labelled pages exist and no labeller is planned | follows from the above |
| Crop-to-crop cosine as a link signal | measured: two men reach 0.93, one man reaches 0.96, no threshold exists (`caveats/audit-open.md#cosine-not-identity`) | 2026-08-12 |

The consequence runs through the whole document. **Every structure below comes out of a gemma prompt
field, or out of plain Python over gemma's output.** Nothing below is trained. A learned score
in the original design becomes a hand-weighted sum. The weights are read off the labelled chapter, and
the DoD is the accuracy number rather than the mechanism.

## The measurement spine

Nothing below can be called done without this. **The character half is built, 2026-08-13**, and
reproduces the table below on the 19:44 run. The dialogue half is still empty
(`decisions/measurement-spine.md#truth-scope-is-38`).

The only ground truth in the project is the eyeball pass over the 19:44 run of 2026-08-12. It lives in
prose in `NEXT.md`. Write it to `eval/chapter-truth.json` against chapter
`7c944dd4-e972-42c7-ba60-9f6939548e80`, scoped to what was already checked by eye rather than to all 119
detections:

- the three characters walked crop by crop, each assignment marked as the real person or not
- 30 dialogue lines with their true speaker, typed as `visible | offscreen | narrator | unknown`

The baseline it records, from that run:

| character | assignments | correct | purity | note |
| --- | --- | --- | --- | --- |
| the lead | 16 | 14 | 0.88 | plus a photograph at `order 17` and a chibi at `order 20` |
| `character_2b1b12a1` | 13 | 13 | 1.00 | a main character the registry never named |
| `character_f0d4e901` | 9 | 7 | 0.78 | the other 2 are `2b1b12a1` |

Woman A is `2b1b12a1`. She has 15 occurrences split across 2 ids, so her fragmentation is 2.

`audit_registry.py` already walks panels, reads `identity_assignments` and counts per character. Extend
it to print purity and fragmentation against the truth file. Do not write an eval harness.

**Done when:** `audit_registry.py <chapter>` prints purity per labelled character and fragmentation per
labelled person, and reproduces the table above on the 19:44 run.

**Met for the character half.** All 38 labelled occurrences match an assignment at IoU 0.5, and all six
numbers print `= baseline`. The definitions had to be pinned down first. Purity is a share and
fragmentation is a count of ids, so neither needs a `character_id` from the truth file
(`decisions/measurement-spine.md#purity-is-id-free`). Run it with:

```bash
docker cp audit_registry.py manga-orchestrator:/app/
docker cp eval/chapter-truth.json manga-orchestrator:/app/eval/
docker exec manga-orchestrator python3 /app/audit_registry.py
```

## 1. The page is a region graph, not a list of panels

```
page
 ├─ regions
 │   ├─ panel
 │   ├─ inset_panel
 │   ├─ embedded_art
 │   ├─ text
 │   ├─ tail
 │   └─ character_occurrence
 │
 └─ edges
     ├─ contains(region, region)
     ├─ reads_before(text, text)
     ├─ tail_of(tail, text)
     ├─ points_to(tail, character)
     ├─ spoken_by(text, character)
     └─ same_identity(character, character)
```

A flat set of panels cannot express a television inside a room. That is the defect the current pipeline
shows most often.

With no detector to train, the graph comes out of the detection prompt. Gemma already returns a box per
character and per text. Two fields per detection buy most of the graph with no new model:

```
plane   = story | screen | photo | poster | drawing | flashback | dream
species = human | animal | object
```

`plane` is the containment edge in disguise. A detection whose `plane` is not `story` sits inside embedded
art, and that is the fact every stage below needs. `species` is a separate axis and exists because vision
boxes cats as people and dresses them.

Tail regions stay unbuilt. The `det`/`seg` heads exist and are unused
(`caveats/speaker-attribution.md#tail-is-not-geometry`), and section 3 says why they are not the first
thing to spend on.

**Today:** the crop stage emits a flat panel list with a bbox each, plus `context_fragments`, a
non-destructive caption-to-face link. Vision emits per-panel characters and dialogue. There is no
containment edge, no region type, no `plane` and no `species`.

**Done when:** every detection carries `plane` and `species`. On the labelled chapter, `order 17` and
`order 20` of the lead are not `story`, none of his 14 correct crops is demoted, and `p081` and `p108` are
`animal`. Measured by `audit_registry.py`, which already reads the vision blob per panel.

## 2. Identity exists independently of names

```
occurrence c42
    -> identity char_07
         name = null
         aliases = []
```

`char_07.name` may be filled later, or stay null forever and display as `unknown character #7`. The
occurrence is the observation, the identity is the cluster, the name is an optional label on the cluster.
Three levels, never collapsed into one.

**Today:** already true. `identity_assignments` is the occurrence, `characters` owns the identity, `name`
is nullable and downstream already falls back to an anonymous display.

**Done when:** already done. No work item. The clustering is section 4 and the naming is
`decisions/identity-naming.md`.

## 3. Speaker attribution is a scored edge, not a procedure

Do not write `find bubble -> find tail -> nearest character`. Score every plausible edge:

```
score(text, character) =
    w1 * gemma_answer
  + w2 * spatial_evidence
  + w3 * same_panel
  + w4 * same_plane
  + w5 * conversation_continuity
  + w6 * character_activity_prior
```

The original design put a learned `t2c` head in the first term and called it load-bearing. No labelled
pages exist, so that term does not. **Gemma's answer becomes one term of six rather than the whole
procedure.** The geometry terms overrule it when they agree against it. The weights are constants read off
the 30 labelled lines. Six numbers in a module, not a training run.

Geometry carries normalized relative position, distance, overlap, same-panel and containment depth. Tail
direction is absent until a tail region exists, and it is not the first thing to build. `conversation
continuity` is free, and turn-taking is the strongest prior for a tail-less bubble.

Then the cases fall out of one mechanism instead of four:

| case | what carries it |
| --- | --- |
| bubble with a tail | gemma plus spatial, usually decisive |
| bubble with no tail | conversation continuity plus spatial |
| speaker outside the panel | recent identities plus an offscreen candidate |
| narration | the narrator candidate |
| nothing resolves | unknown speaker |

**A dialogue line must not be required to resolve to a visible character.** That is a failure mode, not a
safeguard. The speaker type is a union:

```
speaker = visible(character_id) | offscreen(character_id?) | narrator | unknown
```

**Today:** `speaker_ref` is already a typed union of `character_id | name | unknown | narrator`
(`decisions/audit-phase1.md#speaker-ref-is-canonical`). `offscreen` is the missing arm. Attribution is a
prompt to gemma over a window of panels, with no geometry term at all.

**Done when:** `audit_speakers.py` reports accuracy over the 30 labelled lines, split by true type, and
the scored version beats the recorded gemma-window baseline. Two numbers must move the right way, and both
are reported. Correct assignments go up. **Forced** errors go down, where forced means a line given a
visible character while the truth is `offscreen`, `narrator` or `unknown`. Record the baseline before
touching the code.

## 4. Character recognition is occurrence, then identity, then name

```
character detection
       ↓
occurrence embeddings
       ↓
pairwise same_identity scores
       ↓
chapter-wide constrained clustering
       ↓
char_001, char_002, ...
       ↓
optional name claim
       ↓
name or unknown
```

Three rules the current code gets wrong.

**The embedding is not the character crop alone.** The crop embedding measures scene, not person, which is
why two men reach 0.93. Combine the face or head crop with the person crop instead of replacing one with
the other. `face_detect` already finds the face and pairs it for `has_face`, so the face box is free. This
is the queued experiment in `NEXT.md` item 1: re-embed the same 22 detections and recompute the matrix.

**Cluster chapter-wide, not page by page.** `tracklets.link_tracklets` groups within an 8-panel window.

**Two characters in the same panel may be one person.** Seven things break that rule. Mirrors,
photographs, flashbacks, insets, screens, imagined scenes, repeated action drawings. Make it a weak
cannot-link, and only between detections on the same `plane`.

That last rule has an ordering trap. Same-panel co-presence is currently a **hard** constraint and it is
load-bearing precisely because cosine cannot separate people. Weakening it before the embedding improves
will regress purity. The dependency is the embedding fix, not the `plane` field alone.

**Today:** the embedding is the person box only (`caveats/audit-open.md#cosine-not-identity`). Clustering
is greedy and local. `tracklets.cannot_link` treats same-panel co-presence as hard. Naming is
`db.add_name_claim`, corroboration over `name_claims`.

**Done when:** no labelled character holds more than one wrong assignment, and woman A's fragmentation is
1. Baseline is 2 wrong, 0 wrong, 2 wrong, and fragmentation 2. The bar is stated in errors rather than in
a purity ratio on purpose. The three characters hold 16, 13 and 9 assignments. At those counts any ratio
above 0.94 means zero tolerated errors, and the ratio hides that.

The face-plus-person embedding lands first and carries its own smaller check. On the 22 measured
detections, the highest different-person pair must fall below the lowest same-person pair.

## 5. The narrative plane is what stops art-in-art

The page is a hierarchical scene graph:

```
page
└── panel A                     plane=story
    ├── character c1
    ├── text t1
    └── television/poster       plane=screen
        ├── character c2
        └── text t2
```

Perfect classification is not the point. The output that matters is one predicate:

```
same_narrative_plane(a, b)
```

It has exactly two consumers, and they are the reason the field is worth adding at all:

- section 3, as the `same_plane` term. A real character beside a poster does not get the poster person's
  line.
- section 4, as the guard that makes the weak cannot-link safe.

**Today:** nothing models this, and it is the whole of the remaining identity error on the lead. On the
19:44 run his 16 assignments were 14 correct plus a photograph of another man and a chibi drawing. Both
are art inside a panel.

**Done when:** section 1's DoD, plus both consumers wired, plus section 4's purity DoD holds with the
cannot-link demoted to weak. If purity regresses when the constraint is weakened, the embedding is not
ready and the demotion reverts.

## 6. Narrative understanding is carried state, not a per-panel description

Today each stage reads its predecessor's blob for one panel or one beat. `recent`, a rolling list of the
last few dialogue lines, is the only carried state. That is the root of the invented narration.

The version worth building is one record per scene, inherited forward:

```
scene_31:
  location: school_rooftop
  time: evening
  participants: {char_03: present, char_07: present, char_11: offscreen}
  narrative_mode: present
  last_speaker: char_07
  addressee: char_02
```

A panel inherits this unless visual evidence overrides it. That alone kills a class of errors. A character
absent for one panel has not left. A panel with no background has not changed location. A tail-less line
keeps the offscreen participant as a candidate. A close-up still belongs to the scene. `last_speaker` and
`participants` are what section 3's continuity term reads.

A panel produces a **delta** against that record, not another standalone prose interpretation:

```
panel 142:
- character_07 enters room_03
- character_02 is already present
- character_07 says "..."
```

Classify the transition, not just the panel:

```
CONTINUE_SCENE | NEW_SCENE | LOCATION_CHANGE | TIME_SKIP | FLASHBACK_START
FLASHBACK_END | DREAM/IMAGINATION | POV_CHANGE | EMBEDDED_SCENE
```

`EMBEDDED_SCENE` is `plane != story` at scene granularity, and is what stops a television's contents
mutating the room around it.

**Scene state is written by a resolver, never by the vision model.** The path is
`observation -> resolver -> state transition`. The resolver may reject an impossible update. It is plain
Python over gemma's per-panel delta, and it is where the constraint lives.

**Today:** none of it exists. Chapter boundaries are a reset. Narration asserts things no panel shows
(`NEXT.md` item 6).

**Done when:** a scene record carries location, participants and `narrative_mode` across panels. A
character absent from one panel stays a participant. On the next full run the four invented-fact
timestamps do not recur. Those are 0:43, 2:03, 2:05 and 2:15, and they are the regression list. The
correctness verifier passed 116/116 over them because it checks quotes and names, never invented claims.
So the check is a re-watch of those four points, not a stage counter.

### Not building yet

Each of these was in the original design. Each is deferred with a trigger, not dropped.

| deferred | trigger to revisit |
| --- | --- |
| Facts, hypotheses and unknowns as separate records with confidences | when scene state exists and narration still asserts unshown claims |
| The seven-check consistency checker | when a scene record exists for it to check against |
| Unresolved references that survive and back-propagate | when a second chapter of the same manga runs |
| Chapter checkpoints and the two-memory split | when a second chapter of the same manga runs |

One reason covers all four. They sit on an identity layer still wrong on 2 of the lead's 16 crops. State
machinery over wrong identity produces confidently wrong state.

## 7. Build order

Detection, vision and character embeddings already exist. The order below is chosen so each step is
falsifiable by the step's own DoD before the next one starts.

```
0. eval/chapter-truth.json + purity and fragmentation in audit_registry.py   DONE, characters only
1. plane + species per detection                      -> section 1 DoD
2. face-plus-person embedding                         -> section 4 embedding check
3. chapter-wide clustering, weak cannot-link on plane  -> section 4 purity DoD
4. scene record carried forward                       -> section 6 DoD
5. scored speaker edge, offscreen arm                 -> section 3 DoD
6. tail regions from the unused det/seg heads         -> only if 5 misses its DoD
```

Steps 1 and 2 are independent and can land together. Step 3 depends on 2, which is the ordering trap in
section 4. Step 5 depends on 4, because the continuity term reads the scene record. Step 6 is
conditional on purpose: build a tail detector only after the cheap terms have been measured and found
insufficient.

The VLM keeps judging ambiguous edges. What changes is that it stops rediscovering every character and
dialogue relationship from raw pixels on every panel. The chapter graph carries the answer forward.

## Sources

Three ideas come from published comics-transcription work. Detection and association as graph generation,
the text-to-character pair head, and the character bank of exemplar images plus names. The
formulation is kept. The models are ruled out, see **Ruled out** above. No source in this document is a
runtime dependency.