Handback, stage D4: the taxonomy from adaptations

Legs D4a-01 to D4a-03, 24 and 25 October 2026. The taxonomy is read off the adjudicated ledger (the re-check, all 705 rooms), adapted against the model frozen at D3 (46384686…). The model is unchanged: the P05 and A11 amendments of D4a-01 were reverted (QD4-001) and are held proposals. The detail and the scripts are in tree-D4.md, notes-D4.md, branches.jsonl, forms/, crosslinks.jsonl, chunk-share-D4.* and projections-D4.jsonl.

The claim

The rooms adapted the model in many small shared ways, not into a few forms. 35 branches hold 476 of the 705 rooms, and they overlap: 759 cross-links join 348 pairs of branches that do not nest. A room makes several element-level changes at once. The branches are facets of those changes, not exclusive types. The comparison built without the ledger agrees in its own way: its partition beats its null but barely separates the rooms (silhouette 0.024), does not survive a change of distance, and largely sorts the rooms by how much the coders coded. What the branches say is specific:

• Rooms add work the model does not state. They administer the April Fools play (F01). They carry out deals: they move work, retarget scripts and relay terms (F10). They keep the shared instruction current, because a stale template or bot target undoes their own new work (F15, F33, F08). They keep the chat open (F16). They make help wait on their own work and spare hands (F22, F28). And they carry on by renewing the work elsewhere, often before the close (F11 to F14).
• They limit P05's added clause. In design, at the border and in keeping (F07, F26, F27), informed members reject a planned change, or the record does not show the untold keepers the clause needs.
• They show two links the model lacks. After a move the room builds again (F09). After a celebration it defends again (F32).

The room that best supports it: fivenightsatfreddys. It is in eight branches across seven elements. It runs a takeover as a temporary regime (F01, F02), does not show the telling clause at its design (F07), builds again after a move (F09), carries out a deal (F10), carries on past the close (F23), depends on chat access (F29) and hosts by one-off permission (F31). One room, many facets, no single form.

The room that most threatens it: ainbowroad. It is the largest room in no branch (2,054 concepts). Its shared changes sit on keys held by fewer than 20 rooms, which the pass did not read. A room can hold rich organizing that the branches do not show.

The condition between them is whether a room's change lands on a node that many other rooms use. Where it does, the room appears in a branch, often in several. Where its changes are its own, or shared by only a few rooms at one node, it appears in none. The branches are a census of common changes, not of the rooms' organizing.

The return value

The branches, in the order cut, with the verdict and canvas percentiles

Verdicts are projections.py's map (my last ruling at D3, else the readers' agreed verdict), so the D3 versions are mixed. Each row: id, parent, level, members; composes, with residue, does not compose, cannot tell; the does-not-compose percentile; the median area and its percentile. The canvas percentiles follow room size (tree-D4.md).

The held proposals (QD4-001; never served, not re-checked)

• P05 (move 198; history lines 173 to 176; reverted by move 203, line 181). The statement would gain: "This holds where the keepers can learn of a planned change only by being told, and no informed member holds it." Status bounded. The bound names completeanarchy, polska, satania, halo, bangtan and thesilphroad (the shape read) and armok, maryland and oneshot (informed members), with their nine re-check lines. The branches F07, F26 and F27 hold 108 rooms where the clause holds only in part.
• A11 (move 199; history lines 177 to 180; reverted by move 203, line 182). The arrow from P02 to P04 would be withdrawn under rule 14: it never beat its null; template first is nearly as common as the transition; the failure clause stands alone in 117 rooms, 31 of them talk only.

Questions, departures and defects

• QD4-001 (answered): the amendments were reverted, and the model is the D3 version.
• QD4-002 (answered): the comparison engine was patched to v1.1; declared as a departure (departures.jsonl line 10).
• For the operator. validate.py --open D4 reads the startup packet's pattern name rooms-x-adaptations.<level>.<kind>.tsv literally. stage_D4.py resolves branch keys against elements only, so arrow keys, the largest change keys in this ledger, cannot key a branch. Its member-name check is a substring check: the room 'de' forbids 'decisions' and 'outside'.

Licences exercised (METHOD.md)

• The count set by kinds, not by a head's capacity: exercised. The branches number what the keys and their lines support (35), under rules written before the first cut; the cuts stop at a declared key size.
• An in-sample return in Eisenhardt's spirit: its ledger is the taxonomy's material; the taxonomy cannot corroborate it (section 7).
• Langley's strategy: not exercised at D4; the two links the model lacks wait for an order test.
• Gioia's explosion at the front end: not a D4 licence.

Slips

• At D4a-01 the session ended inside one command; its log line and snap were written, and the empty F04 dry run was repeated.
• F06's form lists the arrow A12 among the elements its members carry unchanged; the forms are written once.
• Two spec names (F16, F30) were changed before writing, after the member-name check refused them.
• At D4a-03 three Edits to this handback's script ran in one turn, against the rule; each landed on a separate string.