diff --git a/.github/scripts/add-to-project/classify-initiative.sh b/.github/scripts/add-to-project/classify-initiative.sh
new file mode 100755
index 000000000..30408e1e4
--- /dev/null
+++ b/.github/scripts/add-to-project/classify-initiative.sh
@@ -0,0 +1,309 @@
+#!/usr/bin/env bash
+# classify-initiative.sh — assign the Initiative single-select field (and its
+# rolled-up Theme) on Initiatives-project items that don't have one yet.
+#
+# Why this exists (petry-projects/.github#415): the add-to-project automation
+# only ever places items ON the board (addProjectV2ItemById); nothing set the
+# Initiative field, so nearly every item sat on the board unassociated with any
+# initiative. This is the continuous, deterministic classifier that fills that
+# gap — and its first run back-fills the existing blank items.
+#
+# Classification is RULE-DRIVEN, not AI: each item's title + labels + repo are
+# flattened into a lowercase "signature" and matched against ordered regex
+# rules (initiative-rules.tsv, first match wins). A matched item gets its
+# Initiative set, plus the Theme that Initiative rolls up to
+# (initiative-taxonomy.tsv). An item that matches NO rule is left blank and
+# reported for triage — the classifier never guesses a bucket.
+#
+# Safety: this only ever writes per-item field VALUES
+# (updateProjectV2ItemFieldValue via lib.sh's set_item_single_select_value).
+# It never touches the field SCHEMA (updateProjectV2Field), so it cannot trip
+# the single-select option-wipe footgun documented in
+# standards/initiatives-project.md. By default it only fills items whose
+# Initiative is currently empty, so a human's manual assignment is never
+# overwritten; RECLASSIFY=all re-evaluates every item.
+#
+# Required env:
+#   PROJECT_ID        ProjectV2 node ID of the Initiatives project
+#   GH_TOKEN          Token with org Projects: Read+write
+#
+# Optional env:
+#   PROJECT_URL       Logged in human-readable messages only
+#   DRY_RUN=1         Log intended writes, mutate nothing
+#   RECLASSIFY=all    Re-evaluate items that already have an Initiative
+#   RULES_FILE        Override path to the rules TSV
+#   TAXONOMY_FILE     Override path to the Initiative→Theme TSV
+#   PAGE_SIZE         Items fetched per GraphQL page (default 100)
+#   INITIATIVE_FIELD  Initiative field name (default "Initiative")
+#   THEME_FIELD       Theme field name (default "Theme")
+#
+# Functions (sourceable / unit-tested):
+#   normalize_signature <title> <labels_json> <repo>
+#   classify_by_rules   <signature>          → Initiative name or "" (no match)
+#   theme_for           <initiative>         → Theme name or ""
+#   decide_for_signature <signature>         → "<initiative>\t<theme>" or ""
+#   resolve_fields                           → populate field-id / option maps
+#   sweep_project                            → paginate + reconcile the board
+
+set -euo pipefail
+
+_ci_dir="$(cd -- "$(dirname -- "${BASH_SOURCE[0]}")" && pwd)"
+# shellcheck source-path=SCRIPTDIR
+# shellcheck source=lib.sh
+. "${_ci_dir}/lib.sh"
+
+# Field ids + name→optionId maps, resolved once per run by resolve_fields.
+declare -gA CI_INIT_OPT CI_THEME_OPT
+CI_INIT_FIELD_ID=""
+CI_THEME_FIELD_ID=""
+
+# normalize_signature <title> <labels_json> <repo>
+#   Flatten an item into one lowercase line "title | l1,l2 | owner/repo" that
+#   the regex rules match against. A non-array labels_json (null/object from
+#   odd payloads) degrades to no labels rather than aborting under set -e.
+#
+#   Gate labels are STRIPPED from the signature (SIGNATURE_IGNORE_LABELS,
+#   default = the required + excluded noise-gate labels). This matters:
+#   every board item carries the `dev-lead` required label, so leaving it in
+#   would make the `dev-lead agent` rule match every single item. Only labels
+#   that actually signal an initiative survive into the signature.
+normalize_signature() {
+  if [ "$#" -ne 3 ]; then
+    printf '[normalize_signature] expected 3 args (title labels_json repo), got %d\n' "$#" >&2
+    return 64
+  fi
+  local title="$1" labels_json="$2" repo="$3" labels=""
+  # Drop routing/process labels that carry no initiative signal, so they can't
+  # drive a classification. Two mechanisms:
+  #   1. an explicit ignore list (SIGNATURE_IGNORE_LABELS) — the noise-gate
+  #      labels (dev-lead + the excluded set);
+  #   2. a family-prefix strip for `dev-lead*` and `initiative*`, which also
+  #      removes their colon-variants (`dev-lead:needs-human`,
+  #      `initiative:auto`, …). Those variants contain "dev-lead"/"initiative"
+  #      and would otherwise make the `dev-lead agent` / `Initiatives Project`
+  #      rule tokens match every pipeline-routed item.
+  local ignore="${SIGNATURE_IGNORE_LABELS-dev-lead,compliance-audit,health-check,fleet-tracker,daily-report}"
+  if printf '%s' "${labels_json}" | jq -e 'type == "array"' >/dev/null 2>&1; then
+    labels=$(printf '%s' "${labels_json}" | jq -r --arg ig "${ignore}" '
+      ($ig | ascii_downcase | split(",") | map(gsub("^\\s+|\\s+$";"")) | map(select(length > 0))) as $drop
+      | [ .[].name
+          | select((ascii_downcase) as $n
+              | (($drop | index($n)) | not)
+                and (($n | test("^(dev-lead|initiative)(:|$)")) | not)) ]
+      | join(",")')
+  fi
+  printf '%s | %s | %s' "${title}" "${labels}" "${repo}" | tr '[:upper:]' '[:lower:]'
+}
+
+# classify_by_rules <signature> → first-matching Initiative name, or "".
+classify_by_rules() {
+  if [ "$#" -ne 1 ]; then
+    printf '[classify_by_rules] expected 1 arg (signature), got %d\n' "$#" >&2
+    return 64
+  fi
+  local sig="$1" name rx
+  local rules="${RULES_FILE:-${_ci_dir}/initiative-rules.tsv}"
+  if [ ! -f "${rules}" ]; then
+    printf '[classify_by_rules] rules file not found: %s\n' "${rules}" >&2
+    return 65
+  fi
+  while IFS=$'\t' read -r name rx || [ -n "${name}" ]; do
+    name="${name%$'\r'}"
+    rx="${rx%$'\r'}"
+    case "${name}" in ''|'#'*) continue ;; esac
+    [ -n "${rx}" ] || continue
+    if printf '%s' "${sig}" | grep -Eiq -- "${rx}"; then
+      printf '%s' "${name}"
+      return 0
+    fi
+  done < "${rules}"
+  return 0
+}
+
+# theme_for <initiative> → the Theme it rolls up to, or "".
+theme_for() {
+  if [ "$#" -ne 1 ]; then
+    printf '[theme_for] expected 1 arg (initiative), got %d\n' "$#" >&2
+    return 64
+  fi
+  local want="$1" init theme
+  local tax="${TAXONOMY_FILE:-${_ci_dir}/initiative-taxonomy.tsv}"
+  if [ ! -f "${tax}" ]; then
+    printf '[theme_for] taxonomy file not found: %s\n' "${tax}" >&2
+    return 65
+  fi
+  while IFS=$'\t' read -r init theme || [ -n "${init}" ]; do
+    init="${init%$'\r'}"
+    theme="${theme%$'\r'}"
+    case "${init}" in ''|'#'*) continue ;; esac
+    if [ "${init}" = "${want}" ]; then
+      printf '%s' "${theme}"
+      return 0
+    fi
+  done < "${tax}"
+  return 0
+}
+
+# decide_for_signature <signature> → "<initiative>\t<theme>" (theme may be
+# empty), or "" when no rule matches.
+decide_for_signature() {
+  if [ "$#" -ne 1 ]; then
+    printf '[decide_for_signature] expected 1 arg (signature), got %d\n' "$#" >&2
+    return 64
+  fi
+  local sig="$1" init theme
+  init=$(classify_by_rules "${sig}") || return $?
+  [ -n "${init}" ] || return 0
+  theme=$(theme_for "${init}") || return $?
+  printf '%s\t%s' "${init}" "${theme}"
+}
+
+# resolve_fields — query the project once for the Initiative + Theme
+# single-select field ids and their live option name→id maps. Fails loudly
+# (75) if the project node is unreachable, (65) if the Initiative field is
+# absent. Theme is best-effort: its absence is not fatal.
+resolve_fields() {
+  _atp_require_env resolve_fields || return $?
+  local json
+  # shellcheck disable=SC2016  # $projectId/$initName/$themeName are GraphQL variables
+  json=$(gh api graphql \
+    -F projectId="${PROJECT_ID}" \
+    -F initName="${INITIATIVE_FIELD:-Initiative}" \
+    -F themeName="${THEME_FIELD:-Theme}" \
+    -f query='query($projectId:ID!,$initName:String!,$themeName:String!){
+      node(id:$projectId){
+        ... on ProjectV2 {
+          initiative: field(name:$initName){ ... on ProjectV2SingleSelectField { id options{ id name } } }
+          theme:      field(name:$themeName){ ... on ProjectV2SingleSelectField { id options{ id name } } }
+        }
+      }
+    }')
+
+  if [ "$(printf '%s' "${json}" | jq -r '.data.node?')" = "null" ]; then
+    printf '[resolve_fields] GraphQL returned data.node:null for PROJECT_ID=%s — token may lack access, or the project was deleted.\n' "${PROJECT_ID}" >&2
+    return 75
+  fi
+  CI_INIT_FIELD_ID=$(printf '%s' "${json}" | jq -r '.data.node.initiative.id? // ""')
+  CI_THEME_FIELD_ID=$(printf '%s' "${json}" | jq -r '.data.node.theme.id? // ""')
+  if [ -z "${CI_INIT_FIELD_ID}" ]; then
+    printf '[resolve_fields] Initiative single-select field %q not found on the project.\n' "${INITIATIVE_FIELD:-Initiative}" >&2
+    return 65
+  fi
+
+  local id name
+  while IFS=$'\t' read -r id name; do
+    [ -n "${id}" ] && CI_INIT_OPT["${name}"]="${id}"
+  done < <(printf '%s' "${json}" | jq -r '.data.node.initiative.options?[]? | "\(.id)\t\(.name)"')
+  while IFS=$'\t' read -r id name; do
+    [ -n "${id}" ] && CI_THEME_OPT["${name}"]="${id}"
+  done < <(printf '%s' "${json}" | jq -r '.data.node.theme.options?[]? | "\(.id)\t\(.name)"')
+}
+
+# _ci_report <total> <already> <matched> <unmatched> <skipped>
+_ci_report() {
+  local total="$1" already="$2" matched="$3" unmatched="$4" skipped="$5"
+  local mode="apply"; [ "${DRY_RUN:-}" = "1" ] && mode="dry-run"
+  printf '\n=== classify-initiative summary (%s) ===\n' "${mode}"
+  printf '  board items scanned : %d\n' "${total}"
+  printf '  already associated  : %d (skipped; RECLASSIFY=all to re-evaluate)\n' "${already}"
+  printf '  newly matched       : %d\n' "${matched}"
+  printf '  unmatched (blank)   : %d\n' "${unmatched}"
+  printf '  option-missing skip : %d\n' "${skipped}"
+  if [ -n "${GITHUB_STEP_SUMMARY:-}" ]; then
+    {
+      printf '### classify-initiative (%s)\n\n' "${mode}"
+      printf '| scanned | already | matched | unmatched | skipped |\n'
+      printf '|--:|--:|--:|--:|--:|\n'
+      printf '| %d | %d | %d | %d | %d |\n' "${total}" "${already}" "${matched}" "${unmatched}" "${skipped}"
+    } >> "${GITHUB_STEP_SUMMARY}"
+  fi
+}
+
+# sweep_project — resolve fields, page through every board item, and set the
+# Initiative (+ Theme) on each qualifying item. Idempotent; DRY_RUN-aware.
+sweep_project() {
+  _atp_require_env sweep_project || return $?
+  resolve_fields || return $?
+
+  local page_size="${PAGE_SIZE:-100}" items
+  # --paginate concatenates one JSON document per page; jq -s slurps them and
+  # flattens all item nodes into a single array.
+  # shellcheck disable=SC2016  # GraphQL variables, not shell
+  items=$(gh api graphql --paginate \
+    -F projectId="${PROJECT_ID}" \
+    -F pageSize="${page_size}" \
+    -F initName="${INITIATIVE_FIELD:-Initiative}" \
+    -f query='query($projectId:ID!,$pageSize:Int!,$endCursor:String,$initName:String!){
+      node(id:$projectId){
+        ... on ProjectV2 {
+          items(first:$pageSize, after:$endCursor){
+            pageInfo{ hasNextPage endCursor }
+            nodes{
+              id
+              initiative: fieldValueByName(name:$initName){ ... on ProjectV2ItemFieldSingleSelectValue { name } }
+              content{
+                __typename
+                ... on Issue        { title labels(first:20){ nodes{ name } } repository{ nameWithOwner } }
+                ... on PullRequest  { title labels(first:20){ nodes{ name } } repository{ nameWithOwner } }
+                ... on DraftIssue   { title }
+              }
+            }
+          }
+        }
+      }
+    }' | jq -s '[.[].data.node.items.nodes?[]?]')
+
+  local total=0 already=0 matched=0 unmatched=0 skipped=0
+  local node
+  while IFS= read -r node; do
+    [ -n "${node}" ] || continue
+    total=$((total + 1))
+
+    local item_id cur title labels repo sig decided init theme optid
+    item_id=$(printf '%s' "${node}" | jq -r '.id?')
+    cur=$(printf '%s' "${node}" | jq -r '.initiative.name? // ""')
+    title=$(printf '%s' "${node}" | jq -r '.content.title? // ""')
+    labels=$(printf '%s' "${node}" | jq -c '.content.labels.nodes? // []')
+    repo=$(printf '%s' "${node}" | jq -r '.content.repository.nameWithOwner? // ""')
+
+    if [ -n "${cur}" ] && [ "${RECLASSIFY:-}" != "all" ]; then
+      already=$((already + 1))
+      continue
+    fi
+
+    sig=$(normalize_signature "${title}" "${labels}" "${repo}")
+    decided=$(decide_for_signature "${sig}")
+    init="${decided%%$'\t'*}"
+    theme="${decided#*$'\t'}"
+
+    if [ -z "${init}" ]; then
+      unmatched=$((unmatched + 1))
+      printf 'UNMATCHED  %-20s «%s»\n' "${repo:-draft}" "${title}"
+      continue
+    fi
+
+    optid="${CI_INIT_OPT[${init}]:-}"
+    if [ -z "${optid}" ]; then
+      printf '::warning::rule matched Initiative %q which is not a live project option; skipping «%s»\n' "${init}" "${title}" >&2
+      skipped=$((skipped + 1))
+      continue
+    fi
+
+    printf 'MATCH  %-22s <- %-20s «%s»\n' "${init}" "${repo:-draft}" "${title}"
+    set_item_single_select_value "${item_id}" "${CI_INIT_FIELD_ID}" "${optid}"
+    matched=$((matched + 1))
+
+    # Theme is best-effort: co-assign only when the field and matching option
+    # both exist live. A missing Theme field/option is silently tolerated.
+    if [ -n "${theme}" ] && [ -n "${CI_THEME_FIELD_ID}" ]; then
+      local topt="${CI_THEME_OPT[${theme}]:-}"
+      [ -n "${topt}" ] && set_item_single_select_value "${item_id}" "${CI_THEME_FIELD_ID}" "${topt}"
+    fi
+  done < <(printf '%s' "${items}" | jq -c '.[]')
+
+  _ci_report "${total}" "${already}" "${matched}" "${unmatched}" "${skipped}"
+}
+
+if [ "${BASH_SOURCE[0]}" = "${0}" ]; then
+  sweep_project
+fi
diff --git a/.github/scripts/add-to-project/initiative-rules.tsv b/.github/scripts/add-to-project/initiative-rules.tsv
new file mode 100644
index 000000000..ccc3e2b7a
--- /dev/null
+++ b/.github/scripts/add-to-project/initiative-rules.tsv
@@ -0,0 +1,38 @@
+# Initiative <TAB> ERE regex  — deterministic classification rules for
+# classify-initiative.sh, matched (case-insensitive) against a per-item
+# signature = "title | label,label | owner/repo", all lowercased.
+#
+# ORDER IS PRIORITY (first match wins). Two tiers:
+#   1. CROSS-CUTTING infra/agentic initiatives come FIRST — they win
+#      regardless of which repo the work lives in. This is deliberate: a
+#      SonarCloud/Scorecard sweep or an auto-rebase fix inside a product repo
+#      is Org Standards / Auto-rebase work, not product work (see
+#      standards/initiatives-project.md → Org Standards covers scorecard/
+#      sonarcloud across all repos).
+#   2. PRODUCT initiatives come LAST — a repo-scoped fallback. Anything in a
+#      product repo that no infra rule claimed, plus product-feature idea
+#      drafts (which have no repo), get the product bucket via repo name or
+#      product-feature keywords.
+# Unmatched items are left BLANK and reported for triage — never guessed.
+Auto-rebase	auto.?rebase|update.?branch|rebase (bot|loop|collision)|conflict.?comment|dismiss.?stale|last.?push.?approval|stale.?base
+pr-review agent	pr.?review|(code|bot|single|advisory) review|review.?(one.?pr|decision|enumeration|loop|comments?|bot|gate|findings|annotation|enhancement|dismissal|tier|cycles?|retry)|single.?review|max_review|rubber duck|cascade|advisory.?bot|coderabbit|(gemini|copilot|sonarcloud|codex) review|fix.?request|extract_verdict|verdict|mention.?(ack|trigger)|changes_requested|@petry-review|@donpetry-bot|review-one-pr|list-prs|copilot suggest|github models|churn breaker|self.?trigger|downstream.?impact|pr_diff|mcp.*(review|tiers?|knob|config|server)|lsp (index|pilot|tool)|reviewer|awaiting review|#860
+Business Analyst	feature.?ideation|idea.?enhanc|idea.?triage|idea-enhancer|business analyst|enhance .*ideas?|single-idea enhancement|idea..?initiative pipeline
+Model Selection	model.?(selection|fallback|tier|switch)|fallback model|opus.?fallback|retire .*(opus|gpt|claude-|model)|update to opus|opus 4|cross.?provider|provider.?switch|additional providers|providers,? .*model|engine.?(availability|model|upgrade)|multi-model chain|rate.?limit resilience
+Agent Shield	agent.?shield|prompt.?injection|input sanitiz|sanitiz|agent.?(security|hardening|governance)|jailbreak|secret.?redact|mcp.*(governance|allowlist|registry)|credential isolation|owasp|posture scoring|egress (firewall|policy|filter|rule)|firewall policy|trust escalation|agent autonomy
+Self-healing	self.?heal|auto.?remediat|auto.?fix|self.?repair|stall detect|pipeline stall|recovery escalation|kill switch|circuit breaker|runaway (protection|agent|pr)
+dev-lead agent	dev.?lead|deep.?impl|bmad|engine.?(timeout|error)|story churn|phantom.?spec|thin.?caller|inlined (claude|workflow)|claude\.yml|claude-fix|claude-?issue|claude code workflow|claude workflow|issue automation|prompting framework|devops agent persona|duplicate .*(pr|agent)|dedup|umbrella issue|in-progress label
+GH-AW	gh-?aw|agentic.?workflow|(^|[^[:alnum:]])aw([^[:alnum:]]|$)|ci.?failure|failure.?analyst|ci-?analyst|failure-to-agent|lock\.?file|lock\.yml|eval scorer|held-out|rubric|proposer|skill.?eval|skillopt|skill.?candidate|plan_json|artifact (type|contract)|self-improving|eval-case|evals/|agents\.md (spec|align)
+Copilot Instructions	copilot.*instruction|copilot coding agent|copilot governance|copilot custom|\.github/copilot|copilot-instructions
+Cost Observability	cost (attribution|observ|meter|arbitrage|report|api)|token.?(cost|savings|report|metric|observ)|budget (enforcement|circuit breaker|guardrail|tracking|aware)|credit budget|observability dashboard|opentelemetry|otel|effective token|\bet\b .*anomaly|usage & cost|billing-aware|batch api cost
+Fleet Monitor	fleet.?(monitor|tracker|view|health)|health.?check|hourly runs?|cron entr|scheduled .*(runs?|actions)
+Daily Reports	daily.?report|nightly.?report|status.?digest|org.?status
+Compliance Blitz	compliance.?blitz|compliance.?day|compliance backlog
+Org Standards	sonarcloud|sonar[ ._-]|s7[0-9]{3}|s86[0-9]{2}|s59[0-9]{2}|s3776|s5906|scorecard|ruleset|codeowners|branch.?(protection|ruleset)|push.?protection|ci.?standard|baseline|standards.?(sync|deploy|reconcile)|nosonar|secret.?scan|gitleaks|dependency.?(harden|pinning)|pin.*(sha|ref|reusable|version|action)|actionlint|yamllint|lint check|dependabot|org.?standard|org-wide .*(standard|instruction|reusable)|code.?quality|zizmor|oidc|access control|standards drift|drift.?(guard|detection)|integrity verification|execution protection|security linting|pull.?request.?limits|pr.?limits|admission gate|agents\.md|claude\.md|codeql|reusable workflow|apply-repo-settings|repo settings|profile readme|coding standard|ci pipeline|markdown lint|concurrency group|check-suite|slsa|attestation|bootstrap.?new.?repo|repo.?template|non-file policy|require .*discussions|has_discussions|structured logging|cqrs|stacked pr|worktree|convention|merge queue|noise reduction|cognitive complexity|jest assertion
+Compliance program	compliance|audit|remediat
+Release Strategy	release.?(strategy|train|notes|channel|cadence|automation|cut|soak|ring)|cut.?release|autocut|soak.?(promote|and-promote)|ring.?(rollout|promote|enrol|stag)|channel.?(promote|cut|soak|pin)|canary|version(ing)? (scheme|standard)|semantic versioning|tag protection|rollback|promote .*(channel|ring|stable)
+Initiatives Project	initiative|add.?to.?project|projects? v2|project board|roadmap view|initiatives project|issue.?fields|github projects
+Tooling	tooling|scratchpad|dev.?tool|utility script|template gallery
+Broodly	/broodly|broodminder|\bhive\b|honey|nectar|bloom|colony|\bmite|apiary|beekeep|hornet|treatment efficacy|nfc and qr|acoustic colony|zero-tap|off-season engagement
+Markets	/markets|snap/ebt|produce recognition|app-averse|failed-trip|vendor catalog|north-star|discover filter|sms check-in
+Google App Scripts	/google-app-scripts|apps? script|script_catalog|gmail|deployed-script
+TalkTerm	/talkterm|avatar|elevenlabs|lip-sync|screen companion|companion mode|3d avatar|spatial audio|byok|effort modes|voice (assistant|agent|interface|provider|transcription)|conversational (voice|oauth|permission|mcp|onboarding)|screenshot context|deliberation mode|emotion-aware|multi-language voice|infinite conversation|on-device (voice|ai)|vision-native
diff --git a/.github/scripts/add-to-project/initiative-taxonomy.tsv b/.github/scripts/add-to-project/initiative-taxonomy.tsv
new file mode 100644
index 000000000..9dd686865
--- /dev/null
+++ b/.github/scripts/add-to-project/initiative-taxonomy.tsv
@@ -0,0 +1,25 @@
+# Initiative <TAB> Theme  — canonical roll-up used by classify-initiative.sh.
+# Comments (#) and blank lines ignored. See standards/initiatives-project.md.
+dev-lead agent	Agentic Framework
+pr-review agent	Agentic Framework
+GH-AW	Agentic Framework
+Copilot Instructions	Agentic Framework
+Agent Shield	Agentic Framework
+Model Selection	Agentic Framework
+Fleet Monitor	Fleet Operations
+Daily Reports	Fleet Operations
+Org Standards	Fleet Operations
+Release Strategy	Fleet Operations
+Compliance program	Compliance
+Compliance Blitz	Compliance
+Self-healing	Compliance
+Auto-rebase	Compliance
+Initiatives Project	Tooling
+Tooling	Tooling
+Ad hoc	Ad hoc
+Business Analyst	Agentic Framework
+Cost Observability	Fleet Operations
+TalkTerm	Products
+Broodly	Products
+Markets	Products
+Google App Scripts	Products
diff --git a/.github/scripts/add-to-project/lib.sh b/.github/scripts/add-to-project/lib.sh
index 04d19baea..9f2bb86f4 100644
--- a/.github/scripts/add-to-project/lib.sh
+++ b/.github/scripts/add-to-project/lib.sh
@@ -260,3 +260,45 @@ delete_project_item() {
     return 1
   fi
 }
+
+# set_item_single_select_value <item_id> <field_id> <option_id>
+#   Set a single-select field on one project item to the given option.
+#
+# This is updateProjectV2ItemFieldValue — a per-item VALUE write. It is NOT
+# updateProjectV2Field (the schema mutation that full-replaces the option
+# list and can orphan every item's value; see
+# standards/initiatives-project.md → "Adding or modifying single-select
+# options safely"). Writing an item value here can never wipe the schema, so
+# a back-fill sweep over hundreds of items is inherently safe.
+#
+# Exit 0 on success; 64 on bad args. DRY_RUN=1 logs the intended write and
+# mutates nothing.
+set_item_single_select_value() {
+  if [ "$#" -ne 3 ]; then
+    printf '[set_item_single_select_value] expected 3 args (item_id field_id option_id), got %d\n' "$#" >&2
+    return 64
+  fi
+  local item_id="$1"
+  local field_id="$2"
+  local option_id="$3"
+
+  if [ "${DRY_RUN:-}" = "1" ]; then
+    printf '[dry-run] would set item %s field %s = option %s\n' "${item_id}" "${field_id}" "${option_id}"
+    return 0
+  fi
+
+  # shellcheck disable=SC2016  # $projectId/$itemId/$fieldId/$optionId are GraphQL variables
+  gh api graphql \
+    -F projectId="${PROJECT_ID}" \
+    -F itemId="${item_id}" \
+    -F fieldId="${field_id}" \
+    -F optionId="${option_id}" \
+    -f query='mutation($projectId:ID!,$itemId:ID!,$fieldId:ID!,$optionId:String!){
+      updateProjectV2ItemFieldValue(input:{
+        projectId:$projectId, itemId:$itemId, fieldId:$fieldId,
+        value:{ singleSelectOptionId:$optionId }
+      }){
+        projectV2Item { id }
+      }
+    }' >/dev/null
+}
diff --git a/.github/workflows/add-to-project-reconcile.yml b/.github/workflows/add-to-project-reconcile.yml
index bbd5f324e..dcb94b0ae 100644
--- a/.github/workflows/add-to-project-reconcile.yml
+++ b/.github/workflows/add-to-project-reconcile.yml
@@ -9,19 +9,29 @@
 #
 # Runs centrally in petry-projects/.github (the scripts live here); mints the
 # petry-projects-planner App token for issues/PRs read + project write.
+#
+# Two phases run in order each cycle:
+#   1. reconcile — ensure every qualifying issue/PR/Idea is ON the board.
+#   2. classify  — set the Initiative (+ Theme) field on board items that
+#      don't have one yet (petry-projects/.github#415). Classify runs AFTER
+#      reconcile so items added this cycle are categorized in the same run.
 name: Add-to-Project Reconcile
 
 on:
   workflow_dispatch:
     inputs:
       dry_run:
-        description: "Log intended adds/removes without mutating the board"
+        description: "Log intended adds/removes/classifications without mutating the board"
         type: boolean
         default: true
       repos:
         description: "Optional space-separated owner/repo list (default: all App-installed repos)"
         type: string
         default: ""
+      reclassify_all:
+        description: "Re-evaluate the Initiative on items that already have one (default: only fill blanks)"
+        type: boolean
+        default: false
   schedule:
     - cron: "0 7 * * *" # daily 07:00 UTC
 
@@ -63,3 +73,15 @@ jobs:
           # Manual runs default to dry-run; scheduled runs are live.
           DRY_RUN: ${{ (github.event_name == 'workflow_dispatch' && inputs.dry_run) && '1' || '0' }}
         run: bash .github/scripts/add-to-project/reconcile-backlog.sh
+
+      - name: Classify Initiative on board items
+        env:
+          GH_TOKEN: ${{ steps.app-token.outputs.token }}
+          PROJECT_ID: PVT_kwDOD2inqs4BZq3-
+          PROJECT_URL: https://github.com/orgs/petry-projects/projects/1
+          # Same dry-run gating as reconcile: manual runs default to dry-run,
+          # scheduled runs are live.
+          DRY_RUN: ${{ (github.event_name == 'workflow_dispatch' && inputs.dry_run) && '1' || '0' }}
+          # Only fill blank Initiatives by default; opt in to re-evaluate all.
+          RECLASSIFY: ${{ (github.event_name == 'workflow_dispatch' && inputs.reclassify_all) && 'all' || '' }}
+        run: bash .github/scripts/add-to-project/classify-initiative.sh
diff --git a/standards/initiatives-project.md b/standards/initiatives-project.md
index 80b27d4ea..2e6e32d7f 100644
--- a/standards/initiatives-project.md
+++ b/standards/initiatives-project.md
@@ -109,8 +109,8 @@ Theme. Roadmap-view grouping uses Initiative; cross-cutting filters
 
 | Theme | Initiatives |
 |---|---|
-| **Agentic Framework** | `dev-lead agent`, `pr-review agent`, `GH-AW`, `Copilot Instructions`, `Agent Shield`, `Model fallback` |
-| **Fleet Operations** | `Fleet Monitor`, `Daily Reports`, `Org Standards` |
+| **Agentic Framework** | `dev-lead agent`, `pr-review agent`, `GH-AW`, `Copilot Instructions`, `Agent Shield`, `Model Selection` |
+| **Fleet Operations** | `Fleet Monitor`, `Daily Reports`, `Org Standards`, `Release Strategy` |
 | **Compliance** | `Compliance program`, `Compliance Blitz`, `Self-healing`, `Auto-rebase` |
 | **Tooling** | `Initiatives Project`, `Tooling` |
 | **Ad hoc** | `Ad hoc` |
@@ -195,12 +195,15 @@ configured manually in the UI):
 
 ```text
 .github/workflows/add-to-project.yml             # Workflow (events → script call)
-.github/workflows/add-to-project-reconcile.yml   # Scheduled/manual backlog reconcile (#518)
+.github/workflows/add-to-project-reconcile.yml   # Scheduled/manual backlog reconcile (#518) + classify (#415)
 .github/scripts/add-to-project/
-    lib.sh                                       # find/add/draft/delete helpers (DRY_RUN-aware)
+    lib.sh                                       # find/add/draft/delete/set-field helpers (DRY_RUN-aware)
     add-issue-or-pr.sh                           # Noise gate + addProjectV2ItemById
     reconcile-discussion.sh                      # Paginated find + 4-state reconciler
     reconcile-backlog.sh                         # Scans open issues/PRs + Ideas, reconciles via the above
+    classify-initiative.sh                       # Rule-driven Initiative/Theme back-fill (#415)
+    initiative-rules.tsv                         # Ordered "Initiative <TAB> regex" match rules
+    initiative-taxonomy.tsv                      # "Initiative <TAB> Theme" roll-up
 .github/workflows/add-to-project-tests.yml       # shellcheck + bats CI gate
 test/workflows/add-to-project/                   # bats tests, gh stub, fixtures
 ```
@@ -219,6 +222,46 @@ discussion number) serialize via concurrency group — `created` and
 **Project ID:** `PVT_kwDOD2inqs4BZq3-` (hardcoded in the workflow env).
 Multi-Project consumers of the same scripts would parameterize this.
 
+## How Initiative classification works
+
+Adding an item to the board and *categorizing* it are separate concerns. The
+add path (above) only sets board membership; it leaves the **Initiative**
+field blank. `classify-initiative.sh` (#415) fills it.
+
+- **Deterministic, not AI.** Each item is flattened into a lowercase
+  signature `title | labels | owner/repo` and matched against ordered regex
+  rules in `initiative-rules.tsv` — **first match wins**, so specific
+  initiatives sit above the generic bucket they'd otherwise be swallowed by
+  (Compliance Blitz before Compliance program; Org Standards before
+  Compliance program). A matched item also gets the **Theme** that Initiative
+  rolls up to, from `initiative-taxonomy.tsv`.
+- **Gate labels are stripped from the signature.** Every board item carries
+  the `dev-lead` required label, so if it stayed in the signature the
+  `dev-lead agent` rule would match *every* item. `SIGNATURE_IGNORE_LABELS`
+  (default = the required + excluded noise-gate labels) drops it; only labels
+  that actually signal an initiative survive. This is the concrete mechanism
+  behind the "`dev-lead` is a work-assignment signal, not a classification
+  signal" note in **Fields** above.
+- **No guessing.** An item that matches no rule is left blank and printed as
+  `UNMATCHED` for triage — the classifier never assigns "Ad hoc" as a
+  catch-all. Coverage is reported in the job summary on every run. Tune
+  coverage by adding rows to `initiative-rules.tsv`.
+- **Safe by construction.** It writes only per-item field *values*
+  (`updateProjectV2ItemFieldValue`), never the field *schema*
+  (`updateProjectV2Field`), so it cannot trip the option-wipe footgun in
+  "Adding or modifying single-select options safely" above. It resolves live
+  option ids by name at runtime; a rule naming an initiative that isn't a
+  live option is skipped with a warning rather than failing the run.
+- **Non-destructive.** By default it only fills items whose Initiative is
+  empty, so a human's manual assignment is never overwritten. Set
+  `RECLASSIFY=all` (workflow input `reclassify_all`) to re-evaluate every
+  item.
+
+It runs as the second step of `add-to-project-reconcile.yml`, right after the
+backlog reconcile, so items added in a cycle are categorized in the same run.
+Manual dispatch defaults to `dry_run` (preview coverage + the UNMATCHED list
+before mutating). The first live run back-fills the existing blank items.
+
 ## Deferred work (not in the pilot)
 
 Tracked in [#415](https://github.com/petry-projects/.github/issues/415).
diff --git a/test/workflows/add-to-project/classify-initiative.bats b/test/workflows/add-to-project/classify-initiative.bats
new file mode 100644
index 000000000..a014b5fe4
--- /dev/null
+++ b/test/workflows/add-to-project/classify-initiative.bats
@@ -0,0 +1,314 @@
+#!/usr/bin/env bats
+# Tests for classify-initiative.sh — the deterministic Initiative/Theme
+# back-fill classifier:
+#   - pure classification (signature build, rule match, taxonomy roll-up)
+#   - the gate-label strip that stops the universal `dev-lead` label from
+#     making every item look like the "dev-lead agent" initiative
+#   - the sweep: resolve fields → page items → set values (DRY_RUN + apply),
+#     skipping already-associated items and leaving unmatched ones blank.
+#
+# Classification cases run against the REAL rules/taxonomy TSVs, so they also
+# guard those data files against regressions.
+
+bats_require_minimum_version 1.5.0
+
+load 'helpers/setup'
+
+setup() {
+  tt_make_tmpdir
+  tt_install_gh_stub
+  export PROJECT_ID="PVT_test_project"
+  export PROJECT_URL="https://example.invalid/projects/1"
+  export GH_TOKEN="t_test"
+  export GH_STUB_LOG="${TT_TMP}/gh.log"
+  # shellcheck source=/dev/null
+  . "${TT_SCRIPTS_DIR}/classify-initiative.sh"
+}
+
+teardown() {
+  tt_cleanup_tmpdir
+}
+
+# One-page items response. Each arg is a compact JSON node object.
+write_items_page() {
+  local out_path="$1"; shift
+  local nodes='[]'
+  local n
+  for n in "$@"; do
+    nodes=$(jq --argjson node "$n" '. + [$node]' <<<"$nodes")
+  done
+  jq --argjson nodes "$nodes" \
+     '{data:{node:{items:{pageInfo:{hasNextPage:false,endCursor:""},nodes:$nodes}}}}' \
+     <<<"{}" >"$out_path"
+}
+
+# Field-schema response used by resolve_fields.
+write_fields_schema() {
+  local out_path="$1"
+  cat >"$out_path" <<'JSON'
+{"data":{"node":{
+  "initiative":{"id":"F_INIT","options":[
+    {"id":"o_orgstd","name":"Org Standards"},
+    {"id":"o_auto","name":"Auto-rebase"},
+    {"id":"o_devlead","name":"dev-lead agent"}
+  ]},
+  "theme":{"id":"F_THEME","options":[
+    {"id":"t_fleet","name":"Fleet Operations"},
+    {"id":"t_comp","name":"Compliance"},
+    {"id":"t_agentic","name":"Agentic Framework"}
+  ]}
+}}}
+JSON
+}
+
+gh_script_line() { printf '%s\t%s\t%s\n' "$1" "$2" "$3"; }
+
+assert_invocation_count() {
+  local expected="$1" actual=0
+  [ -f "${GH_STUB_LOG}" ] && actual=$(wc -l <"${GH_STUB_LOG}" | tr -d ' ')
+  [ "$actual" -eq "$expected" ] || {
+    printf 'expected %d gh invocations, got %d\n' "$expected" "$actual" >&2
+    [ -f "${GH_STUB_LOG}" ] && cat "${GH_STUB_LOG}" >&2
+    return 1
+  }
+}
+
+assert_log_contains() {
+  local blob; blob=$(sed 's/\\//g' "${GH_STUB_LOG}")
+  for needle in "$@"; do
+    [[ "$blob" == *"${needle}"* ]] || {
+      printf 'expected gh log to contain %q\nlog: %s\n' "$needle" "$blob" >&2
+      return 1
+    }
+  done
+}
+
+# ---------------------------------------------------------------------------
+# Arg validation
+# ---------------------------------------------------------------------------
+
+@test "normalize_signature: rejects wrong arg count" {
+  run normalize_signature "t" "[]"
+  [ "$status" -eq 64 ]
+}
+
+@test "classify_by_rules: rejects wrong arg count" {
+  run classify_by_rules
+  [ "$status" -eq 64 ]
+}
+
+@test "theme_for: rejects wrong arg count" {
+  run theme_for a b
+  [ "$status" -eq 64 ]
+}
+
+@test "decide_for_signature: rejects wrong arg count" {
+  run decide_for_signature
+  [ "$status" -eq 64 ]
+}
+
+# ---------------------------------------------------------------------------
+# normalize_signature — lowercasing + gate-label strip
+# ---------------------------------------------------------------------------
+
+@test "normalize_signature: lowercases and joins title | labels | repo" {
+  run normalize_signature "Fix The Thing" '[{"name":"auto-rebase"}]' "petry-projects/.github"
+  [ "$status" -eq 0 ]
+  [[ "$output" == "fix the thing | auto-rebase | petry-projects/.github" ]]
+}
+
+@test "normalize_signature: strips the dev-lead gate label but keeps real labels" {
+  run normalize_signature "some title" '[{"name":"dev-lead"},{"name":"auto-rebase"}]' "o/r"
+  [ "$status" -eq 0 ]
+  [[ "$output" != *"dev-lead"* ]]
+  [[ "$output" == *"auto-rebase"* ]]
+}
+
+@test "normalize_signature: strips dev-lead:* and initiative:* routing variants" {
+  # These colon-variants contain dev-lead/initiative and would otherwise drive
+  # a spurious dev-lead-agent / Initiatives-Project match.
+  run normalize_signature "some title" '[{"name":"dev-lead:needs-human"},{"name":"initiative:auto"},{"name":"scorecard"}]' "o/r"
+  [ "$status" -eq 0 ]
+  [[ "$output" != *"dev-lead"* ]]
+  [[ "$output" != *"initiative"* ]]
+  [[ "$output" == *"scorecard"* ]]
+}
+
+@test "normalize_signature: non-array labels degrade to no labels (no crash)" {
+  run normalize_signature "title" "null" "o/r"
+  [ "$status" -eq 0 ]
+  [[ "$output" == "title |  | o/r" ]]
+}
+
+# ---------------------------------------------------------------------------
+# classify_by_rules — real-rules regression guard + ordering
+# ---------------------------------------------------------------------------
+
+@test "classify_by_rules: the bare dev-lead label does NOT match dev-lead agent" {
+  # Signature as normalize_signature would produce it: gate label already gone.
+  run classify_by_rules "some unrelated title |  | petry-projects/.github"
+  [ "$status" -eq 0 ]
+  [ -z "$output" ]
+}
+
+@test "classify_by_rules: sonarcloud title → Org Standards" {
+  run classify_by_rules "sonarcloud: resolve remaining s7635 stubs |  | petry-projects/.github"
+  [ "$output" = "Org Standards" ]
+}
+
+@test "classify_by_rules: auto-rebase title → Auto-rebase" {
+  run classify_by_rules "auto-rebase: unguarded conflict-comment aborts |  | x/y"
+  [ "$output" = "Auto-rebase" ]
+}
+
+@test "classify_by_rules: pr-review title → pr-review agent" {
+  run classify_by_rules "pr-review advisory bot review gate |  | x/y"
+  [ "$output" = "pr-review agent" ]
+}
+
+@test "classify_by_rules: a dev-lead-agent title still matches via the word in the title" {
+  run classify_by_rules "dev-lead deep-impl engine-error timeout |  | x/y"
+  [ "$output" = "dev-lead agent" ]
+}
+
+@test "classify_by_rules: ordering — Compliance Blitz beats generic Compliance" {
+  run classify_by_rules "compliance blitz 2026 backlog |  | x/y"
+  [ "$output" = "Compliance Blitz" ]
+}
+
+@test "classify_by_rules: cross-cutting infra WINS over the product repo (SonarCloud in a product repo → Org Standards)" {
+  # The key ordering guarantee: a SonarCloud sweep inside a product repo is
+  # Org Standards work, not product work.
+  run classify_by_rules "sonarcloud: shell script hygiene cleanup |  | petry-projects/talkterm"
+  [ "$output" = "Org Standards" ]
+}
+
+@test "classify_by_rules: a genuine product feature in the product repo → the product initiative" {
+  run classify_by_rules "add spatial audio to the voice provider pipeline |  | petry-projects/talkterm"
+  [ "$output" = "TalkTerm" ]
+}
+
+@test "classify_by_rules: avatar idea draft (no repo) → TalkTerm" {
+  run classify_by_rules "3d avatar evolution with lip-sync |  | draft"
+  [ "$output" = "TalkTerm" ]
+}
+
+@test "classify_by_rules: beekeeping idea → Broodly, not TalkTerm (hive wins)" {
+  run classify_by_rules "conversational voice assistant grounded in hive history |  | draft"
+  [ "$output" = "Broodly" ]
+}
+
+@test "classify_by_rules: feature-ideation → Business Analyst" {
+  run classify_by_rules "feat(feature-ideation): add curated source list for mary |  | petry-projects/.github"
+  [ "$output" = "Business Analyst" ]
+}
+
+@test "classify_by_rules: token cost report → Cost Observability, not Daily Reports" {
+  run classify_by_rules "org-wide weekly token cost observatory report |  | petry-projects/.github"
+  [ "$output" = "Cost Observability" ]
+}
+
+@test "classify_by_rules: no match → empty output, exit 0" {
+  run classify_by_rules "zzq nonsense placeholder widget |  | x/y"
+  [ "$status" -eq 0 ]
+  [ -z "$output" ]
+}
+
+@test "classify_by_rules: missing rules file → 65" {
+  RULES_FILE="${TT_TMP}/nope.tsv" run classify_by_rules "anything"
+  [ "$status" -eq 65 ]
+}
+
+# ---------------------------------------------------------------------------
+# theme_for + decide_for_signature
+# ---------------------------------------------------------------------------
+
+@test "theme_for: Auto-rebase rolls up to Compliance" {
+  run theme_for "Auto-rebase"
+  [ "$output" = "Compliance" ]
+}
+
+@test "theme_for: unknown initiative → empty" {
+  run theme_for "Nonexistent"
+  [ "$status" -eq 0 ]
+  [ -z "$output" ]
+}
+
+@test "decide_for_signature: matched → 'initiative<TAB>theme'" {
+  run decide_for_signature "sonarcloud ruleset drift |  | x/y"
+  [ "$status" -eq 0 ]
+  [[ "$output" == "Org Standards"$'\t'"Fleet Operations" ]]
+}
+
+@test "decide_for_signature: unmatched → empty" {
+  run decide_for_signature "zzq nonsense placeholder |  | x/y"
+  [ "$status" -eq 0 ]
+  [ -z "$output" ]
+}
+
+# ---------------------------------------------------------------------------
+# sweep_project — DRY_RUN: resolve + page only, correct tallies
+# ---------------------------------------------------------------------------
+
+setup_sweep_stub() {
+  local fields="${TT_TMP}/fields.json" items="${TT_TMP}/items.json"
+  write_fields_schema "$fields"
+  write_items_page "$items" \
+    '{"id":"PVTI_A","initiative":null,"content":{"__typename":"Issue","title":"SonarCloud: fix S7635 stubs","labels":{"nodes":[{"name":"dev-lead"}]},"repository":{"nameWithOwner":"petry-projects/.github"}}}' \
+    '{"id":"PVTI_B","initiative":{"name":"Auto-rebase"},"content":{"__typename":"PullRequest","title":"auto-rebase tweak","labels":{"nodes":[{"name":"dev-lead"}]},"repository":{"nameWithOwner":"petry-projects/.github"}}}' \
+    '{"id":"PVTI_C","initiative":null,"content":{"__typename":"DraftIssue","title":"zzq nonsense placeholder"}}'
+  local script="${TT_TMP}/script.txt"
+  {
+    gh_script_line 0 "$fields" "-"   # resolve_fields
+    gh_script_line 0 "$items" "-"    # paginate items
+    gh_script_line 0 "-" "-"         # (apply) set Initiative on A
+    gh_script_line 0 "-" "-"         # (apply) set Theme on A
+  } >"$script"
+  export GH_STUB_SCRIPT="$script"
+}
+
+@test "sweep_project DRY_RUN: tallies scanned/already/matched/unmatched, no mutations" {
+  setup_sweep_stub
+  DRY_RUN=1 run sweep_project
+  [ "$status" -eq 0 ]
+  [[ "$output" == *"board items scanned : 3"* ]]
+  [[ "$output" == *"already associated  : 1"* ]]
+  [[ "$output" == *"newly matched       : 1"* ]]
+  [[ "$output" == *"unmatched (blank)   : 1"* ]]
+  [[ "$output" == *"MATCH  Org Standards"* ]]
+  [[ "$output" == *"UNMATCHED"* ]]
+  # Only the two read calls (resolve + paginate); the mutations are dry-run.
+  assert_invocation_count 2
+}
+
+@test "sweep_project apply: sets Initiative and Theme on the matched item" {
+  setup_sweep_stub
+  run sweep_project
+  [ "$status" -eq 0 ]
+  # resolve + paginate + set-initiative + set-theme
+  assert_invocation_count 4
+  assert_log_contains "updateProjectV2ItemFieldValue" "PVTI_A" "F_INIT" "o_orgstd"
+  assert_log_contains "F_THEME" "t_fleet"
+  # Never touches the already-associated or unmatched items.
+  run grep -c PVTI_B "${GH_STUB_LOG}"
+  [ "$output" = "0" ]
+  run grep -c PVTI_C "${GH_STUB_LOG}"
+  [ "$output" = "0" ]
+}
+
+@test "sweep_project: fails fast (75) when the project node is null" {
+  local nullresp="${TT_TMP}/null.json"
+  printf '%s' '{"data":{"node":null}}' >"$nullresp"
+  local script="${TT_TMP}/script.txt"
+  gh_script_line 0 "$nullresp" "-" >"$script"
+  export GH_STUB_SCRIPT="$script"
+  run sweep_project
+  [ "$status" -eq 75 ]
+}
+
+@test "sweep_project: fails fast without GH_TOKEN" {
+  unset GH_TOKEN
+  run --separate-stderr sweep_project
+  [ "$status" -eq 64 ]
+  [[ "$stderr" == *"GH_TOKEN is empty"* ]]
+}