Decision-Tree Gap List — agenda for the stage-2 design session

Date: 2026-07-05 · Status: mapping exercise; triage session CLOSED 2026-07-06 — all four stranger findings resolved (orphan-triage lane G25, typed-exclusion schema G2, review-queue taxonomy G21, devolution two-lens G18); five decision-tree nodes reclassified OPEN → DECIDED-UNBUILT (N1/G1, N3/G5, orphan, review, devolution — plus the four EXCLUDED nodes with the exclusion schema); allocation pile → confirmed stage-2 design agenda (see below); parked pile → confirmed pilot-gated with named revisit triggers (G10, G12, G15). Reclassified gaps: G1, G5, G2, G18, G21, G25. Companions: pipeline_architecture.mmd (systems view), candidate_decision_tree.mmd (logic view). Method: every decision node in the per-candidate tree was annotated rule · executor · status; this is the flat list of every node where the rule is missing/vague, the executor is unallocated, or the required data/implementation does not exist. Each entry says what is missing and what kind of decision closes it — it does not propose the answer.

Closing-decision kinds: RUBRIC (a rule the project lead writes/validates) · ARCH (an architecture / data-contract choice) · LIST (a reference list or vocabulary to build) · IMPL (code to write). Many gaps need more than one.

Honest framing: almost the entire per-candidate tree is OPEN or DESIGN. Layer-0 corpus and Layer-1 extraction are built and verified; Layers 2–5, the counting spine, and every feedback loop are ratified-design-or-open. The tree is not 30 small holes in a working machine — it is the map of the machine that stage-2 has to build. The value below is the shape of each hole.

Index

ID Node / layer What’s undecided Closes with Status
G1 N1 §4 exclusions rule (§4 list) SETTLED; residual = executor allocation (deterministic vs LLM) + where the drop happens ARCH+RUBRIC DESIGN
G2 exclusion outcome RATIFIED: fine-grained two-tier typed exclusion (exclusion_family + exclusion_subclass); writer unbuilt IMPL DESIGN ★
G3 N2 amendment detect §3 rule ratified, but executor unbuilt & is_amendment_insertion broken IMPL+ARCH DESIGN ✦
G4 N2 amendment residence policy for suppressing amending-SI candidates vs the target; backlog tail RUBRIC+IMPL OPEN
G5 N3 cross-ref resolve rule (§3 count-at-source) SETTLED; residual = executor allocation (L2 pattern vs L3) + cross-ref resolver IMPL+ARCH DESIGN
G6 N4 statutory bodies canonical statutory-bodies registry does not exist LIST+IMPL OPEN ✦
G7 N4 function test statutory-vs-contractual executor unallocated ARCH OPEN
G8 N4 context terms §5D term-flip needs per-Act definition resolution IMPL+RUBRIC OPEN
G9 N5 decomposition Layer-3 architecture question (unit → model) ARCH OPEN ✦
G10 N5 discriminator burden-vs-condition rule not empirically tested RUBRIC OPEN
G11 N5 → BERT production model can’t emit per-section multiplicity ARCH OPEN
G12 N6 category axes enum conflates trigger-timing & defence-structure RUBRIC DESIGN
G13 N7 polarity review “review” threshold not operationalised RUBRIC+IMPL DESIGN
G14 N8 obligated party no controlled party vocabulary ARCH+LIST DESIGN
G15 N9 capacity boundary boundary deferred; axis REDEFINED v2.6 (voluntary-market-activity + compulsion carve-out) RUBRIC+ARCH OPEN ✦
G16 N9 economic floor floor reporting needs capacity field + query IMPL OPEN
G17 N10 provenance Citation resolver unbuilt; must be fragment-grain IMPL OPEN
G18 spine dedup RATIFIED: two lenses — residence-based primary count + family_id enrichment relation (uniform/divergent); linker unbuilt, post-labelling IMPL DESIGN ✦
G19 spine join body↔schedule / amend↔target join does not exist ARCH+IMPL OPEN ✦
G20 L5 set-vs-set no burden-matching rule when model counts differ ARCH+RUBRIC OPEN ★
G21 L5 review queue RATIFIED as SCHEMA (review_reason + feeds fields), not tooling; fields unbuilt pending label store IMPL DESIGN ★
G22 L5 label store label store / version-control unbuilt (blocks all) IMPL DESIGN
G23 BERT inference no auto-accept-vs-review confidence policy ARCH OPEN ★
G24 L1 recall gate recall ceiling = word_list union (silent FN) LIST+IMPL OPEN
G25 L1 orphans RATIFIED: typed orphan-triage lane (clear non-operative / escalate → standard path, orphan=true); lane unbuilt IMPL DESIGN ★
G26 L1/L3 EU recitals no soft-law fast-path; retained-EU effect status RUBRIC OPEN ★
G27 L1 definitions def capture bounded (≤8) & single-instrument IMPL OPEN
G28 L0/L1 cache pipeline still fetches live; cache repoint unbuilt IMPL OPEN

✦ = named in the calibration note (expected). ★ = newly surfaced by this mapping (not on the anticipated list).

Gate & structural rules (Layers 1–2)

G1 — §4 non-operative exclusion: executor allocation (rule settled). [N1 · ARCH+RUBRIC · DECIDED-UNBUILT] The §4 non-operative exclusion list is settled — it survived three reviews untouched, so the rule is decided; N1 is DECIDED-UNBUILT, the same shape as N2/G3 (rule ratified, executor unbuilt). What remains is an executor-allocation question, now on the stage-2 allocation agenda: (a) which §4 classes are safe as deterministic L2 drops (shall be deemed, shall be defrayed) versus which need L3 judgement (list-of-contents vs list-of-requirements, definitional-embedded-in-prohibition); (b) which layer runs the drop — noting Layer 1 is deliberately high-recall and never drops (§4 signals are non-blocking hints), and the ratified Stage-2 output is a burden-set per section with no obvious slot for “this candidate is non-operative, drop it” (see G2). Closes with: the deterministic-vs-judgement allocation per §4 class (RUBRIC), and the exclusion-stage architecture (ARCH).

G2 — Typed EXCLUSION record: fine-grained two-tier schema RATIFIED. [exclusion outcome · IMPL · DECIDED-UNBUILT ★] A candidate surfaced by L1 that carries no burden gets a first-class typed exclusion record — for the extraction→count funnel, for audit, and as negative training data for Legal-BERT (rubric §7 wants worked negatives). “Empty burden-set” and “typed exclusion” are not the same thing. Ratified design (2026-07-06) — fine-grained, two-tier: every excluded candidate records exclusion_family + exclusion_subclass.

Consumers: the four EXCLUDED tree nodes and the orphan-triage lane (G25), which emits the same fields. Closes with: implementing the exclusion writer + store (IMPL; folds into the label store, G22).

G3 — Amendment detection: rule ratified, executor unbuilt, existing flag broken. [N2 · IMPL+ARCH · DECIDED-UNBUILT ✦] The §3 amendment-machinery rule is ratified (amendment = machinery; count at the consolidated target), so this node is DECIDED-UNBUILT at best — the rule is decided but nothing executes it. The one field that promises the capability, is_amendment_insertion, is actively broken: _in_amendment() fires only when a whole section nests inside <Addition>, so it flagged 0 of 1,084 candidates despite 16,366 <Addition> nodes (real amendments are inline fragment insertions). It is logged as a bug in project_decision_log.docx (2026-07-05, pattern #3 Named ≠ actual — sibling of the section_ref “JOIN KEY” entry). Detection must be at the <Addition>-fragment grain, not the section flag. Before Stage 2, the field is fixed or renamed so nothing builds on a false signal. Closes with: the fragment-grain detector (IMPL) and an executor decision — markup-rule L4 vs LLM L3 (ARCH); the rule is not the gap.

G4 — Amendment residence policy is incomplete. [N2/spine · RUBRIC+IMPL · OPEN] The rule says count the burden once, at the amended target in consolidated form. But amending SIs are themselves separate corpus items, so their amendment provisions will surface as candidates. There is no policy suppressing those so the count lands once at the target, and the legislation.gov.uk editorial backlog means a small tail of targets is not yet updated → zero-count. Closes with: a residence/suppression rule (RUBRIC) and its enforcement in the spine (IMPL). Overlaps G18.

G5 — Count-at-source: executor allocation + cross-ref resolver (rule settled). [N3 · IMPL+ARCH · DECIDED-UNBUILT] The §3 count-at-source family — cross-refs, compliance/contravention hooks, enabling powers, penalty-as-consequence, offence primary-vs-secondary — is settled (several members ratified this month), so N3 is DECIDED-UNBUILT, the same shape as N2/G3. What remains is an executor-allocation question, now on the stage-2 allocation agenda (L2 pattern vs L3 model), plus its dependency: count-at-source and the offence primary-vs-secondary distinction both require resolving “a requirement under section X” / “contravenes section 4” to the referenced provision — possibly in another instrument — and no such resolver exists. Without it, offence cross-references and compliance hooks can’t be reliably separated from primary burdens. Closes with: the reference-resolution component (IMPL), and the deterministic-vs-LLM allocation (ARCH).

Actor classification (Layer 4)

G6 — Statutory-bodies registry does not exist. [N4 · LIST+IMPL · OPEN ✦] The §5C function test needs a canonical list of statutory bodies (FSCS, FOS, regulators, named officeholders) with their public/private status. word_list.PUBLIC_BODY_SUBJECTS / PRIVATE_ACTOR_SUBJECTS are keyword cues built for the retired analyser and are not imported by extract_candidates.py — they are unwired, and a keyword list is not a registry. Closes with: building the registry (LIST) and wiring a lookup (IMPL).

G7 — Function-source test executor unallocated. [N4 · ARCH · OPEN] “Statutory source = public / contractual source = private, funding-irrelevant” is a judgement with a hybrid→review escape. Whether it runs as an L4 lookup, an L3 LLM call, or both (lookup then LLM for misses) is undecided. Closes with: an architecture decision (ARCH).

G8 — Context-dependent terms unresolved. [N4 · IMPL+RUBRIC · OPEN] §5D terms (“scheme manager” = public in FSMA, private in occupational pensions) flip by Act. Resolution needs the Act’s own definitions; the extractor attaches ≤8 same-document definitions and no cross-instrument ones. Closes with: definition-resolution code (IMPL) and a fallback rule for the unresolved case (RUBRIC → route to review).

Decomposition & the unit of count (Layer 3 / counting spine)

G9 — Decomposition executor = the Layer-3 architecture question. [N5 · ARCH · OPEN ✦] How a section becomes N burden units — per-candidate classifier, counting head, BIO/span tagging, or decompose-then-classify — is the single decision that fixes both the labelling I/O contract and the production-model shape. Formally deferred to stage-2 (rubric §7 item 1, notes item 6). Selection criterion: the chosen architecture must emit usable confidence outputs — production review depends on confidence-thresholding replacing disagreement as the model-side uncertainty signal (G21, G23), and the four candidate architectures differ in how naturally they provide it. Closes with: the stage-2 architecture choice (ARCH).

G10 — Burden-vs-condition discriminator not empirically tested. [N5 · RUBRIC · OPEN] The unit rule’s core call — obligation-leaf (count each) vs condition/factor-leaf (count once) — is defined but not yet run against real chapeau-plus-list provisions from the corpus (§7 item 3). Until tested, the two models will diverge on multiplicity. Revisit trigger (parked, pilot-gated): Jethro’s hand-labelling pass + the dry run — the §7 field-test. Closes with: a validation pass and any resulting rule tightening (RUBRIC).

G11 — Legal-BERT cannot emit per-section multiplicity. [N5→BERT · ARCH · OPEN] A per-candidate production classifier structurally cannot reproduce “3 burdens in this section” or cross-candidate merges without a counting/tagging scheme. The labelled unit (burden) and the model’s emission grain (candidate) are mismatched. Companion to G9. Closes with: the same architecture choice, viewed from the production side (ARCH).

G12 — Category enum conflates two axes. [N6 · RUBRIC · DESIGN] Cats 1/2/5 classify by trigger-timing; Cats 3/4 (IB/IBA) by defence-structure. Resolved for now by the defence-dominance precedence rule (notes item 7) as a stated simplification — logged as pilot-contingent (may need an orthogonal defence flag if defence-revealed burdens show materially different trigger profiles). Revisit trigger (parked, pilot-gated): a pilot showing defence-revealed burdens with materially different trigger profiles. Closes with: a post-pilot rubric decision (RUBRIC).

Per-burden attributes (Layers 3–4)

G13 — Polarity “review” threshold not operationalised. [N7 · RUBRIC+IMPL · DESIGN] review is reserved for genuine post-rule uncertainty, “not a dumping ground” — but nothing measures or enforces that discipline, so the bucket can bloat once a batch runs (§7 item 9). Closes with: a threshold/audit rule (RUBRIC) and its monitoring (IMPL).

G14 — Obligated_party has no controlled vocabulary. [N8 · ARCH+LIST · DESIGN] Free-text party (“employer” / “operator” / “any person”) is the field on which actor-capacity, dedup, and aggregation all depend. Without a controlled taxonomy those downstream steps are unstable. Closes with: a schema decision (ARCH) and a party taxonomy (LIST).

G15 — Actor-capacity boundary deferred to phase 2/3. [N9 · RUBRIC+ARCH · OPEN ✦] Phase 1 uses generous ambiguous; the economic figure is a floor, not a share. The capacity axis was REDEFINED in rubric v2.6 (2026-07-15): capacity turns on whether the duty attaches to voluntary market activity on the burdened actor’s side (exchange or production-for-exchange, either side, formal or informal), with a side-specific compulsion carve-out (scheme-compelled purchases stay personal for the compelled party; the provider side is economic). This supersedes the activity-profile formulation. The economic floor now explicitly includes consumer-side transactors, with a separate business-side cut (economic ∧ business-side obligated_party). Still new and pilot-gated. Revisit trigger (parked, pilot-gated): pilot data, including the both/either base-rate question and the consumer-inclusion effect on the floor. Closes with: the deferred-phase refinement plan (ARCH) and validation of the redefined axis on real cases (RUBRIC).

G16 — Economic-floor reporting needs schema + query support. [N9 · IMPL · OPEN] The reporting rule (floor = economic-only; both/either and ambiguous as separate lines) requires the capacity field to exist and a reporting query to segment it. Neither is built. Closes with: implementation once the field lands (IMPL).

G17 — Provenance resolver unbuilt (data is there). [N10 · IMPL · OPEN] introduced_by / introduced_year are ~100% machine-recoverable (Addition→Commentary→Citation, measured 1,403/1,403 burden-bearing insertions), but no code extracts them, and they must attach at the fragment/burden grain — 22.4% of amendment-touched sections were patched by >1 instrument (max 16), so a section-level field is wrong. Fields are nullable, so labelling won’t stall; the growth-over-time analysis depends on this being built. Closes with: the fragment-grain resolver (IMPL).

Counting spine

G18 — Devolution / parallel-instrument duplicates: two-lens design RATIFIED. [spine · IMPL · DECIDED-UNBUILT ✦] The “one rule, multiple textual homes” pattern means the same regulatory control can appear in several instruments; but intentional devolution duplicates (parallel E&W / Scotland / Wales / NI provisions) must not be silently collapsed. Ratified design (2026-07-06) — two lenses:

Mechanism (five stages, structure-first): (a) candidate instrument-family pairs from Layer-0 metadata (year, counterpart types uksi/ssi/wsi/nisr, title similarity, extent); (b) structural alignment of burdens within paired instruments (position + section skeleton); (c) text similarity as confirmer only — never creates a link without the structural scaffolding; (d) uncertain band → review queue (family_link added to review_reason, a production-era lane, G21); (e) attribute comparison → uniform/divergent tagging.

Acceptance bar — good-enough, asymmetric, measured: precision favoured over recall (a missed link biases the distinct-rules count conservatively = cheap; a false merge overclaims sameness = the expensive direction, kept fussy). Errors are bounded and non-propagating — this is an enrichment layer, and the primary count is untouched. Achieved link-precision is validated by a stratified hand-audited sample and stated in the methodology — the claim is characterised error, not minimal error.

Scope for publication: devolution counterparts only; cross-sector template parallels are future analysis, not publication-gating.

Within-instrument extent-split path (added 2026-07-15, from round-3 extraction). Candidate-family generation needs TWO paths, not one: (i) the cross-instrument instrument-pair path in stage (a) above (parallel uksi/ssi/wsi/nisr counterparts), and (ii) a within-instrument extent-split path — round-3 extraction found the same section_ref resolving to separate England/Wales and Scotland bodies inside ONE Act (EPA 1990 s.33/s.34, HSWA 1974, Wildlife 1981), because DOM-identity keying (correctly) keeps extent variants as distinct section_index candidates. The residence-based primary count treats them as separate burdens (correct); but without path (ii) the distinct-rules enrichment lens silently misses within-Act extent twins, under-linking the family relation. Path (ii) pairs same-section_ref candidates within one instrument that differ only by an extent qualifier.

The methodology reports both lenses with this rationale, which pre-empts the devolution-artefact critique (that the statute-book count inflates by counting parallel instruments). Closes with: building the family-linker (IMPL), post-labelling / pre-publication. Related: G19 (body↔schedule / amend↔target join), G20 (set-vs-set matching), G4 (amendment residence).

G19 — Body↔schedule / amend↔target join does not exist. [spine · ARCH+IMPL · OPEN ✦] A burden’s operative clause and its schedule detail — or an amending Act’s insertion and its target section — live in different sections or items. section_ref is non-unique (48 collisions on the 7-Act set) and DOM-identity keys are per-item only. There is no cross-section/cross-item burden-grouping key. Closes with: a grouping-key design (ARCH) and its implementation (IMPL). Related to G4/G17.

Adjudication & review (Layer 5)

G20 — Set-vs-set agreement has no burden-matching rule. [L5 · ARCH+RUBRIC · OPEN ★] Agreement is defined for the equal-count/equal-label case. When Claude and Gemini emit different burden counts for one section, there is no rule to align the two burden-sets, score partial agreement, or decide what routes to review. This is the common disagreement mode, not an edge case. Closes with: a matching/scoring algorithm (ARCH) and an adjudication rule (RUBRIC).

G21 — Review-queue taxonomy: ratified as SCHEMA (not infrastructure). [L5 · IMPL · DECIDED-UNBUILT ★] At least five distinct review reasons converge on one queue with no typed reason recorded. Ratified design (2026-07-06) — two label-record fields, present from the first pilot adjudication, not a queue tool:

Closes with: implementing the two fields (IMPL; folds into the label store, G22). Set-vs-set burden-matching mechanics stay OPEN — that is the Layer-3 architecture question (G20/G9, allocation agenda), not the taxonomy.

G22 — Label store / version-control unbuilt. [L5 · IMPL · DESIGN] Ratified as “established before any real label is emitted” — the persistence, versioning, and label-provenance layer. It blocks every downstream step (adjudication, training data, BERT). Closes with: building it (IMPL).

G23 — Production-inference confidence/routing undefined. [BERT · ARCH · OPEN ★] At corpus run-time there is no policy for when a Legal-BERT output is auto-accepted vs routed to review, nor what triggers the active-learning retrain loop. The feedback loops are drawn but ungoverned. Closes with: a thresholding/routing decision (ARCH).

Intake edges & recall (Layers 0–1)

G24 — Recall ceiling = the word_list union. [L1 · LIST+IMPL · OPEN] A genuine burden phrased with none of the cue vocabulary is never surfaced and is invisible downstream (silent false negative). This is an upstream gate with no “unknown burden” branch; the only control is empirical recall audit against ground truth. Closes with: ongoing recall additions (LIST) and a standing recall-audit harness (IMPL). Partly acknowledged in the implementation plan.

G25 — Orphan / partial-context candidates: typed triage lane RATIFIED. [L1 · IMPL · DECIDED-UNBUILT ★] material_type='orphan' records are per-sentence with n_leaves=0; context_quality='partial' records lack ref/heading, so they do not fit the section-with-leaves tree. Ratified handling (2026-07-06): orphans do not enter the main pipeline directly — they route to a typed orphan-triage lane.

  1. An LLM first-pass clears clearly-non-operative fragments, recording the typed §4 exclusion (exclusion_family/exclusion_subclass, G2). The verdict is asymmetric — clear only what is clearly non-operative, escalate everything else — and the triage pass does not label burdens.
  2. Escalated fragments are enriched with retrievable surrounding raw text and re-enter the standard classification path (decomposition → category → polarity → obligated_party → actor_capacity → provenance) under the normal dual-model + review machinery — same record shape, no parallel pipeline; the retrieved context is part of the candidate the models see, not just the human view.
  3. Every orphan-derived record carries orphan=true to the final burden record, so the slice stays distinguishable (auditable for elevated disagreement / reliability) without being differently shaped.

The lane is the second consumer of the exclusion-label schema (G2; N1’s §4 exclusions is the first). Closes with: building the lane (IMPL). Related measurement: corpus-wide orphan-rate check queued pre-labelling (implementation plan) — an anomalous rate signals a Layer-1 sectioning failure, not a queue to process. Update (2026-07-16): the round-3 old-drafting orphan spike is now fixed at source by the tier-3 anchor (G29); this lane handles only the post-fix residue, and the pre-labelling certification now monitors size distributions alongside orphan rate (a post-fix taxonomy gap shows as a mega-section, not an orphan spike).

G26 — EU recital / retained-EU material has no fast-path. [L1/L3 · RUBRIC · OPEN ★] eu_recital sections are surfaced (with a recital_suspect hint) but recitals are near-always non-operative soft-law; without a routing rule they either add noise or draw inconsistent LLM calls. Retained/assimilated-EU direct-effect and applicability status is also unaddressed in the tree. Closes with: a material-type routing rule (RUBRIC).

G27 — Definition capture is bounded and single-instrument. [L1 · IMPL · OPEN] The extractor attaches ≤8 defined terms from the same document. A classification hinging on a 9th term, or on a term defined in another instrument, can misfire — most acutely for the §5D context-term flips (G8). Closes with: raising/relaxing the cap and cross-instrument definition lookup (IMPL).

G28 — Cache repoint pending. [L0/L1 · IMPL · OPEN] extract_candidates.py still fetches /data.xml live; the ratified local-cache read (revised-current primary / best-collection fallback, three-tier variant priority, variants read from files not the stale index) is unbuilt. Affects reproducibility, scale, and — critically — which text edition is classified. Closes with: the repoint (IMPL).

G29 — Layer-1 old-drafting anchor: tier-3 extension. [L1 · IMPL · DECIDED-BUILT ✦] Round-3 extraction over fresh instruments found an anomalous orphan rate on pre-1900 UK schedules and EU annexes (Merchant Shipping 1894 62%), because tiers 1–2 recognise a section only under a P1…P7 P-level or a numbered Division/Para; Victorian schedule prose/forms and EU annexes live under bare <P>/<Pblock> or OrderedList/ListItem with no <Number> child. Built (2026-07-16, ratified): a third anchor tier (fires only on the former (None,None)) routes such content into three new material_types (uk_schedule_unnumbered, eu_annex, uk_body_tail; taxonomy 5→8), with a per-entry / whole-form / paragraph-block grain rule and list-marker section_refs. Tiers 1–2 frozen → the four documented counters (727/862/106/611) provably untouched (verified old-vs-new on identical XML). Round-3 pool: 36 orphans → 0, re-homed 20-sentence→19 uk_schedule_unnumbered + 16-sentence→10 eu_annex, sizes sane. Full record: _layer1_orphan_investigation.md §C (design) + §F (verification). The orphan-triage lane (G25) and the anomalous-rate trigger remain in force for the post-fix residue; and note (per G25) that post-fix a taxonomy gap now surfaces as a mega-section or type-misassignment, not an orphan spike — monitor size distributions per leg_type × material_type, not orphan rate alone.

G30 — Candidate context-provisioning spec (what Act-level context is served with each candidate). [L0/L1 · ARCH+IMPL · OPEN] Raised by the v2.6 sign-off read (2026-07-17): resolving some rubric calls crosses section boundaries. The passive-voice duty-bearer one-liner (§5) resolves the obligated party “from the context provided with the candidate”; context-dependent terms (§5D) and defined-term lookups need the Act’s interpretation section; count-at-source needs cited/neighbouring provisions. This makes the provisioning contract load-bearing: what travels with each candidate — its own section + leaves[]; the Act’s definitions / interpretation section; cited and neighbouring provisions; the enclosing Part/Schedule heading — determines how many calls resolve vs route to review (context_term). Today the extractor attaches only the section’s own tree plus ≤8 same-document defined terms (G27) — no interpretation-section or neighbouring-provision service. Closes with: a context-provisioning spec (what to serve, at what radius, and how it is delivered to the labeller/model) — ARCH + IMPL, on the stage-2 allocation agenda. Related: G27 (definition cap), §5D, and the passive-voice one-liner. The rubric’s “never guess — route to review” instruction is the safe default until this lands.

Stage-2 allocation agenda — CONFIRMED (2026-07-06)

The allocation pile transfers to the stage-2 design session as-is — nothing in it needs a decision before that session. The confirmed agenda is the set of UNALLOCATED executor questions (the rules above them are settled; these are ARCH/IMPL allocations, not open rule questions):

Reading of the map