Parking Lot v3 — Post-Classifier Flow Map

What happens AFTER Layer 3 (classification). Every category traced with color + preemption scoring.

Layer 4+: What Happens After Classification

Layer 3 output: category + subcategory + confidence. Layer 4 routes to one of 4 paths.

From Layer 3: Classified Issue

category + subcategory + confidence + skip_intake flag

▼

Layer 4: CATEGORY ROUTER

Routes to one of 4 paths based on category

▼

PATH A: BYPASS

No VP. No buffer. Respond and close.

question:factual → grep

question:metric → read file

question:status → read exec_log

question:cost → invoke analyst

question:process → VP Systems inline

report:session → SKIP intake

report:shipped → ack + archive

correction:skill → direct edit

correction:naming → direct edit

✓ Auto-closed. Free. No interruption.

PATH B: 2-STAGE SCOPE

Scope first. Diane approves. THEN buffer.

feature:*

improvement:*

question:architecture

Stage 1: /improve MINIMAL

Sanity + domain + analyst brief

Headless. ~Sonnet. ~5 min.

▼

Diane: "Worth exploring?"

[Yes --top X] or [No → close]

YES ▼

Stage 2: /improve FULL --top X

All 9 steps. Scout + Planner.

Headless. ~Opus. ~20 min.

▼

Diane: "Approve for VP?"

[Approve] [Defer] [Reject]

APPROVE ▼

BUFFER (delivery gate)

Wait for VP capacity. Batch related.

▼

VP receives pre-scoped work

PATH C: SCORED DELIVERY

Score decides: preempt, switch, or buffer.

bug:*

direction:task

proposal:*

correction:data/process

PREEMPTION SCORER

base + keywords + tone + age + task_penalty

▼

≥70

PREEMPT

Drop task

40-69

SWITCH

At break

<40

BUFFER

VP pulls

▼

VP acts

Proposals also need Diane approval

PATH D: AUTO-PROCESS

System handles. No VP. No Diane.

status update → Outstanding.md

direction:priority → P0 flag

process:* → VP Systems edit

✓ Auto-processed and closed.

bug feature improvement question correction direction proposal report process

Path A: Bypass — No VP, No Buffer, Respond and Close

Cheapest path. System answers directly. Zero interruption to VPs.

question:factual

e.g., "Where is the config file for reporter?"

grep/glob codebase

respond inline with file path

close issue

Cost: free | Time: <30s

question:metric

e.g., "What's the JPY Sharpe ratio?"

read tracker.csv / Views.md

respond with value + source

close issue

Cost: free | Time: <30s

question:cost

e.g., "How much did today's sessions cost?"

invoke analyst (Sonnet)

respond with cost breakdown

close issue

Cost: ~Sonnet | Time: ~2 min

correction:skill / correction:naming

e.g., "Rename the postmortem to include time"

VP Systems direct edit

reply: "Updated. Commit abc1234."

close issue

Cost: free | Time: ~5 min

report:session (VP reports)

e.g., "VP ML session: DirectLGBM shipped"

SKIP intake entirely

These are VP→Diane FYI. Not actionable issues. Filtered by classifier.

status update (blocker cleared)

e.g., "Registered in Play Store (F.2 done)"

detect milestone/blocker reference

update Outstanding.md + tracker.csv

reply: "F.2 marked done."

close issue

Cost: free | Milestone-specific update.

Path B: Two-Stage Scoping — Feature & Improvement Flow

Diane controls both gates. VP never sees unscoped work.

Issue classified as: feature or improvement or question:architecture

▼

STAGE 1: /improve MINIMAL (headless, no VP)

What runs:

✓ Sanity check (grep for duplicates)

✓ Domain research (WebSearch)

✓ Analyst brief (1-paragraph ROI)

What does NOT run:

✗ Scout (prior art deep dive)

✗ Planner (implementation plan)

✗ Full gap analysis (9 steps)

Cost: ~Sonnet | Time: ~5 min | Artifacts: minimal_brief

▼

GATE 1: Diane — "Worth exploring?"

"Exogenous variables for ML model"
Not duplicate. Domain research: macro indicators improve FX accuracy 15-30%.
Analyst: High confidence. Standard in FX literature.

Yes --top 5 No

Diane reviews async (phone/Forum Mobile). ~30 seconds.

NO ▼

CLOSED

No further cost.

Reason logged.

YES --top X ▼

STAGE 2: /improve FULL --top X

All 9 /improve steps run headless:
✓ Smoke test ✓ Domain deep dive ✓ KPI gap analysis
✓ Scout (prior art, build vs buy) ✓ Planner (milestones, tasks, estimates)
✓ Top X recommendations ranked by ROI

Cost: ~Opus | Time: ~20 min | Artifacts: improve_report + scout_report + implementation_plan

Diane controls X — how many results to produce (default 5, can set 3, 7, 10...)

▼

GATE 2: Diane — "Approve for VP?"

Top 5 recommendations:
1. Add DXY index (est 2h, +12% MAPE) ★
2. Add VIX volatility (est 1h, +8% MAPE)
3. Add yield spreads (est 3h, +5% MAPE)
Plan: 3 milestones, 8 tasks, est 6h total
Scout: yfinance free tier, Alpha Vantage API

Approve Defer Reject

APPROVE ▼

NOW ENTERS BUFFER

state: approved

Linked to milestone + KPI in tracker

Waits for VP capacity via delivery gate

▼

VP receives pre-scoped work

All artifacts attached. No in-session scoping.

Picked up via /next-task Level 2.5

Path C: Scored Delivery — Bugs, Directions, Proposals, Corrections

Preemption scorer decides WHEN the VP sees it. Buffer batches low-urgency items.

Issues routed to Path C:

bug:* direction:task proposal:* correction:data/process

▼

PREEMPTION SCORER (5 components → 0-100)

base
(category)

keywords
(urgency)

tone
(escalation)

age
(staleness)

-20

penalty
(task progress)

▼

≥70 PREEMPT

VP drops current task. Direct push.

When this fires:

Diane confirmed bug by testing (base=40)

Data loss/corruption detected (base=40)

"Stop!" + role violation (kw=30)

3+ rapid messages same issue (tone=20)

Real examples (from data):

• "Stop! Why are you updating skills??" → 80
• Email scanner 2nd bug + escalation → 95
• "very very important to me" (Discord) → 80

40-69 SWITCH

VP finishes current sub-task, then switches.

When this fires:

Stale state / redundancy catch (base=20+kw=20)

Domain knowledge injection (base=20+bonus)

"did you check..." (kw=15)

Prior decision override (base=30)

Real examples:

• "You only worked for 2 hours" → 65
• "M9-M12 are already done" → 40
• "did you check forum?" → 45

<40 BUFFER

Parked. VP pulls via /next-task when ready.

When this fires:

New bug report (not confirmed) (base=15)

Direction without urgency (base=15)

Proposal for later (base=15)

Buffer behavior:

• Batched by VP × priority × category
• 3 UI bugs → delivered as 1 batch
• Age escalation: +10 if open 2+ sessions
• VP pulls at /next-task Level 2.5

Real examples:

• Q.58b "Play Store next steps" → 25
• "I'd like diversity of datasets" → -2

Scorer Retraining

After each preemption decision, log: {issue_id, score, band, diane_agreed: yes/no}
Weekly: retrain component weights from Diane's labels. Same approach as classifier retraining.
Target: <2 misclassified preemptions per week.

Path D: Auto-Process — System Handles, No Human Needed

Status Update (blocker cleared)

Diane: "Registered in Play Store (F.2)"

Detect milestone reference (F.2)

Update Outstanding.md: F.2 → [x]

Update tracker.csv: milestone status

Reply: "F.2 marked done. Milestone updated."

Close issue

direction:priority (P0 flag)

Diane: "Stop work on X, focus on Y"

Always preempts (score auto ≥70)

Reprioritize VP tracker

Reply: "Priority changed. VP notified."

process:* (rule changes)

Diane: "All VPs must check Forum before ending"

VP Systems direct edit to SKILL.md

Reply: "Rule encoded. Commit abc1234."

Close issue

Preemption Scorer — Full Component Breakdown

Component	Signal	Score
Base Category (0-40)
	QA failure confirmed by Diane using product	40
	Data loss / corruption detected	40
	Infrastructure failure blocking comms	35
	Prior decision being violated	30
	Role/process boundary violation	30
	Cost waste at calculable rate	25
	Domain knowledge injection	20
	Stale state / redundancy catch	20
	Feature priority pivot	15
	New unconfirmed bug report	15
	Design iteration feedback	5
Urgency Keywords (0-30)
	"Stop!" / "very very" / "this is critical"	30
	"I found a bug" / "X is broken" / two bugs same session	25
	"based on earlier conversations we decided..."	20
	"why is X..." (active use, something missing)	20
	"did you check..."	15
	"how about..." (domain question)	10
	"I was thinking..." / "I'd like..."	3
Tone (0-20)
	3+ messages rapid succession about same issue	20
	Past-tense confirmation ("I tested it, it...")	15
	Explicit cost calculation presented	15
	Escalation from earlier in same session	10
	Single informational message	0
Age (0-10)
	Issue raised 2+ sessions ago, still open	10
	Issue raised earlier same session	5
	New issue this session	0
Current Task Penalty (0 to -20)
	Task >80% complete (about to commit)	-20
	Task 50-79% complete	-10
	Task <50% / debugging / stuck	0
	Task is meta/process (not user-facing)	+5

Calibration: 9 Real Events

Event	Base	Kw	Tone	Age	Pen	Total	Band	Match?
"Stop! Why are you updating skills??"	30	30	15	5	0	80	PREEMPT	✓
Email scanner 2nd bug + escalation	40	25	20	5	+5	95	PREEMPT	✓
"very very important to me" (Discord)	35	30	15	0	0	80	PREEMPT	✓
"You only worked for 2 hours"	30	20	15	0	0	65	SWITCH	✓
"M9-M12 are already done"	20	20	0	0	0	40	SWITCH	✓
"did you check forum?"	20	15	10	0	0	45	SWITCH	✓
"how about exogenous variables?"	20	10	0	0	-10	20	QUEUE	~
Q.58 "registered + need breakdown"	15	10	0	0	0	25	QUEUE	✓
"I'd like diversity of datasets"	5	3	0	0	-10	-2	QUEUE	✓

8/9 correct. 1 edge case (~): Diane's domain questions during ML work occasionally have paradigm-level ROI. Needs VP-ML context bonus.

Full Category × Path × Action Matrix

Category	Subcategory	Path	Scorer?	Buffer?	VP?	Diane?	Action
bug	ui	C	Yes	Score	vp-mobile	No	Fix + verify build
bug	infra	C	Yes	Score	vp-systems	No	Fix daemon/service
bug	data	C	Yes	Score	owner	No	Fix data pipeline
bug	crash	C	Yes (≥70)	No	owner	No	P0 — preempt always
bug	regression	C	Yes (≥70)	No	owner	No	P0 — preempt always
feature	new	B	No	After approve	After approve	2x	/improve min → full → VP
feature	enhancement	B	No	After approve	After approve	2x	/improve min → full → VP
feature	infra	B	No	After approve	vp-systems	2x	/improve min → full → VP
feature	ux	B	No	After approve	vp-mobile	2x	/improve min → full → VP
improvement	*	B	No	After approve	After approve	2x	/improve min → full → VP
question	factual	A	No	No	No	No	grep → respond → close
question	metric	A	No	No	No	No	read file → respond → close
question	status	A	No	No	No	No	read exec_log → respond → close
question	cost	A	No	No	No	No	invoke analyst → respond → close
question	process	A	No	No	No	No	VP Systems inline → close
question	architecture	B	No	After approve	VP	1x	/improve min → Diane → VP
correction	skill	A	No	No	No	No	VP Systems direct edit → close
correction	naming	A	No	No	No	No	VP Systems direct edit → close
correction	data	C	Yes	Score	owner	No	Fix data
correction	process	C	Yes	Score	vp-systems	No	Fix process/rule
direction	task	C	Yes	Score	owner	No	Create task → VP
direction	priority	D	Auto ≥70	No	Yes	No	Reprioritize immediately
proposal	architecture	C	Yes	Score	VP	Yes	Score → VP + Diane approval
proposal	org	C	Yes	Score	—	Yes	Diane only
proposal	process	C	Yes	Score	vp-systems	Yes	Score → VP + Diane approval
report	session	SKIP	No	No	No	No	Filtered out of intake
report	shipped	A	No	No	No	No	Ack + archive → close
process	*	D	No	No	vp-systems	No	Direct edit → close
status update	*	D	No	No	No	No	Outstanding + tracker → close

Parking Lot v3 — Full Post-Classifier Flow Map

Layer 4+: What Happens After Classification

Path A: Bypass — No VP, No Buffer, Respond and Close

Path B: Two-Stage Scoping — Feature & Improvement Flow

Path C: Scored Delivery — Bugs, Directions, Proposals, Corrections

When this fires:

Real examples (from data):

When this fires:

Real examples:

When this fires:

Buffer behavior:

Real examples:

Scorer Retraining

Path D: Auto-Process — System Handles, No Human Needed

Preemption Scorer — Full Component Breakdown

Calibration: 9 Real Events

Full Category × Path × Action Matrix