FIN IPA, Harness, Data Spine Parallel Execution Plan

0. Mission

Execute the Intranet Projection Agentic Access program and the FIN Agentic Harness roadmap together, without losing the operational ingestion spine. The goal is not to close checklist rows. The goal is to prove FINN can use projected intranet docs, operational events, Produce Pro reports, email intelligence, pricing signals, and sales facts as a live, cited operating system.

Completion means a current staging candidate proves all four planes:

1. Human-visible intranet and workspace surfaces render the projection.
2. Click paths work through role-appropriate drill-downs.
3. Backend rows, events, projections, freshness, and tool permissions inspect cleanly under tenant scope.
4. Source artifacts, citations, content hashes, and proof JSON/screenshots make the answer replayable.

1. Current Verified Shape

1.0 Runtime update - 2026-05-20T02:15Z

Current staging proof is pinned to direct deploy dc311ae1-ipa-ds-20260520-r2 / Worker version b9307661-d85f-4142-aa3f-929856755d4d. The current proof index is:

• audit/ipa-harness-data-spine/2026-05-20/00-current-proof-index/proof-index.json

Verified closed or partially closed against that runtime:

• IPA-CTRL-P0-001 closed: current proof index separates historical proof from active staging proof.
• DS-P0-002 closed: Produce Pro sales parser/event payload now carries lot, unit cost, ship/order quantity, customer PO, and ship-via fields; staging sales/physical proof passed 13/13.
• DS-P0-003 closed: inventory_adjustment and inventory_snapshot import through import_pp_data, emit event-store rows, and serve fresh freshness state.
• IPA-P0-001 backend/agentic scope closed: po_timeline_projection runtime shape, replay, and get_po_timeline tool access are verified; UI drill-through proof remains open.
• Market-intel source-to-signal closed for the active tool path: query_market_signal now reads modern signal.price_observation rows as well as legacy vendor quote observations, and browser clickthrough renders the VendorQuoteSignal card with source attribution.

Still open after this update:

• Current real-mailbox terminal-state proof for newest intel@fintail.net and arrivals@fintail.net artifacts.
• customer_commodity_matrix, market_signal_rollup, and daily_sales_situation.
• Vendor/pricelist PDF structured extraction or explicit terminal backlog events.
• Harness verifier work: citation validity, gradeboard/S0 latency baseline, freshness verifier, schema verifier, and skill packaging.

1.1 IPA and projection law

The IPA source docs are still directionally correct:

• no runtime documents table;
• events, knowledge, actors, and operational_memory remain the substrate;
• markdown and vault content are projections, not the source of truth;
• agent-facing outputs must carry source attribution and citations;
• projection catalog entries require audience, cadence, renderer version, citation policy, and wave assignment.

The key gap is runtime coverage. A projection is not complete until it has:

• write path or replay path;
• JSONB or markdown output;
• MCP reader tool where agentic access is expected;
• role catalog / permission exposure;
• visible UI path;
• backend and artifact proof.

1.1.1 MD-first read-surface standard

For operational data, insight, and decision workflows, the primary human/agent read surface is the vault MD projection. Direct CardSpec/tool surfaces are not the final data-product UX unless they are embedded inside, linked from, or explicitly operating as fallback for an MD projection.

Required layering:

1. Event store remains the write path and source-of-truth plane.
2. JSONB projections compute replayable fast state and join-heavy rollups.
3. Vault/tenant MD projections are the canonical operator and agent read surface, with frontmatter that includes source_events[], source_artifacts[], projection_version, renderer_version, freshness, citations, content_hash, and Central-overlay inputs where relevant.
4. CardSpec components are allowed as embedded widgets, drilldowns, diagnostics, and fallback presentations only. A standalone card proof is data-path proof, not final UX proof, for sales/market/pricing workflows.
5. Direct MCP read tools remain useful for diagnostics, projector feeds, and fallback paths. Agent answers should prefer the MD projection when it is present and fresh; fallback to direct tools must be detectable in telemetry or proof artifacts.

The QC workflow is the exemplar: event writes create replayable projections, the workflow surface renders the rich MD/projection context, embedded widgets support focused action, and source drilldowns make the evidence inspectable.

1.1.2 Agent harness and prompt implication

This standard requires a small harness policy change. It is not enough to update docs or renderer UX: the agent must be told the rule, routed through it, and tested against it.

Required harness changes:

1. Add the MD-first rule to FINN_PERSONA_PROMPT and the Sales, Buyer, and Executive role personas.
2. Add deterministic routing so operational data, insight, and decision prompts prefer fresh MD projection readers before direct CardSpec/tool readers.
3. Add deterministic synthesis for MD projection readers that surfaces content_hash, source_events[], source_artifacts[], freshness, and embedded widget/source drilldowns.
4. Keep direct tools such as query_market_signal available for diagnostics, projector inputs, and logged fallback when the MD projection is missing, stale, or out of scope.
5. Add evals proving fresh-MD-wins, stale-MD-fallback, zero-citation rejection, and sales/pricing answers grounded in MD projection evidence.

1.1.3 Data-shape performance assessment

The current storage pattern is directionally right for both auditability and agent leverage:

• append-only events preserve write truth and replay;
• JSONB projections provide fast deterministic state;
• vault MD projections provide the standardized human/agent read surface;
• MCP tools provide scoped access and controlled writes.

It is not yet optimal for agent performance until every primary MD projection also exposes machine-readable evidence units. Markdown alone is good for humans, but the agent should not have to infer source boundaries from rendered prose. Each MD projection should therefore publish an agent evidence manifest, either in frontmatter or a sidecar artifact, containing:

• stable projection_doc_id, projection_version, renderer_version, and content_hash;
• section-level chunk_ids with chunk_hash, title, entity keys, role scope, and freshness class;
• source_events[], source_artifacts[], source_projection_rows[], and Central-overlay pins per chunk;
• citation tier, source license/class, fetched/extracted timestamps, model or extractor run id, and confidence where applicable;
• fallback state when a chunk is absent, stale, or degraded.

That manifest becomes the agent’s high-performance retrieval unit. It lets the harness retrieve by hybrid keyword/vector/entity filters, verify citations, compare freshness by source class, and answer from a compact evidence pack instead of re-reading large markdown bodies or falling back to raw tools.

1.1.4 Data foundation hardening principle

The data foundation is now strong enough that the next risk is uncontrolled growth, not lack of data. Before mailbox/report volume grows much further, every source family needs a stable classification, utilization, and proof contract.

Required discipline:

• every intel@fintail.net artifact lands in a closed-vocabulary source class with terminal state: extracted, routed, deferred, failed-resumable, or explicitly backlog/not-implemented;
• every Produce Pro report family has a report-to-event-to-projection mapping;
• every parsed field is either used, deliberately parked, or explicitly marked as not useful yet with rationale;
• every source artifact carries a content hash, source mailbox, classification version, parser/extractor version, extraction run id, and freshness class;
• every high-value field has a downstream consumer: event payload, JSONB projection, MD projection evidence chunk, freshness row, entity-resolution candidate, or operator workflow surface.

This is the structure freeze before scale. New ingestion should not add silent raw blobs unless the artifact also has a terminal-state event and a backlog owner.

1.2 Sales and market architecture continuity

The Sales & Market Projection Architecture was present in the intranet mirror and an adjacent arrivals-storage worktree, but was missing from this active checkout. It has been restored to:

• docs/architecture/SALES_MARKET_PROJECTION_ARCHITECTURE_2026-05-18.md

That architecture defines:

• 4 JSONB projections: market_signal_rollup, po_timeline_projection, customer_commodity_matrix, and an extended daily_inventory_snapshot_projection;
• 5 vault MD projectors: daily_sales_situation, customer_situation_brief, po_intelligence, sight_picture.weekly, and vendor_scorecard;
• a live-source gap for Gary Blair / Hunts Point intel@ reports;
• sales, buyer, and exec work-rail integration requirements.

Runtime update: query_market_signal now reads both modern signal.price_observation rows and legacy signal.vendor_quote_observed events, normalizes them into vendor quote signal rows, and the browser proof shows the VendorQuoteSignal card with source attribution. That leverages pricing intel and proves the fallback/widget path, but it still does not prove the MD-first sales/pricing UX. The remaining work must reconcile pricing intel with daily sales, inventory snapshots, customer demand, and PO/lot lifecycle facts inside market_signal_rollup and daily_sales_situation, then expose that MD projection as the primary agent and operator surface.

1.3 Data spine and email intel

The prior readiness ledgers show several staging-proven slices:

• Produce Pro sales report ingestion emits sales.report_ingested.
• Buyer-office/Gary pricing replay produced signal.email_extract, extractor.pricing_rail_extracted, signal.price_observation, signal.vendor_quote_observed, and query_market_signal output.
• Vendor/entity/commodity enrichment produced commodity mentions, market chatter, routing decisions, and source-backed market_signal knowledge.
• Data-spine freshness and gap monitoring exists.

However, those are historical proofs. Before further buildout, the current program must rerun current staging proof for:

• newest intel@fintail.net and arrivals@fintail.net traffic;
• every Produce Pro report family now arriving by mailbox;
• terminal PDF states, especially vendor_pricelist;
• Gary Blair pricing reports through source artifact → normalized signals → market tool → agent answer;
• source artifact family maintenance after later deploys.

1.4 Current execution blockers to treat as real

The following are not theoretical. They block the desired expert-system lift:

• vendor_pricelist PDFs classify, but do not yet have full structured price-list extraction/adaptation proof.
• The sales parser field gap for LOT NUMBER, UNIT COST, SHIP QTY, ORDER QTY, CUSTOMER P.O. #, and SHIP VIA is closed in the active branch and staging proof. The next sales gap is reconciliation: sales facts must be joined with pricing intel, inventory, customer demand, and PO/lot lifecycle projections.
• inventory_adjustment and inventory_snapshot ingestion is now verified through staging import, event-store inspection, and freshness serving. The next gap is adjustment-aware inventory state and UI/agent surfacing.
• po_timeline_projection backend/runtime access is now verified through replay and get_po_timeline; UI drill-through remains open.
• customer_commodity_matrix still needs runtime handlers, replay descriptor, MCP reader, role catalog, approval registry entry, and proof before it is counted as accessible to agents.
• Role catalogs are close to the tool-budget cap. New projection tools need default vs specialist/drill-mode policy, not blind catalog expansion.

1.5 Central Taproot freshness and external-source boundary audit

New verified rule: FIN Central Taproot / Knowledge Forge is the only external source gatherer. Tenant taproots receive Central Taproot mirrors, Central-derived signal projections, and operational tenant data. Tenant taproots must not independently fetch USDA, PACA, FDA, NOAA/NWS, AGROVOC, Federal Register, Google Trends, or market-news sources.

Runtime verification on 2026-05-20 shows this rule is not fully enforced yet:

• Knowledge Forge active-agent is deployed and NOAA weather alert staging is fresh: D1 has active-agent:weather_alert source docs with latest crawled at 2026-05-19T22:25:00Z and extracted at 2026-05-20T03:01:37Z.
• Active intelligence v2 is fresh in Central staging: latest quality run completed 2026-05-20T03:02:55Z.
• USDA AMS is not citable/fresh end to end. The GitHub crawl is green, but D1 extraction failed for raw/usda-ams-market-news/2026-05-19/3324.json; the inspected R2 object contains a 401 body (User is not found).
• Composite USDA has completed extraction rows, but its quality stage is failed with a foreign-key constraint error.
• Market News is broken for terminal markets: the latest scheduled run failed all terminal markets with DNS resolution errors and skipped extraction queue publication.
• PACA is broken: source config says weekly, workflow is monthly, the latest scheduled run was cancelled after three hours, and Central D1 has no PACA source_documents.
• FDA recalls and Federal Register crawls are green, but extraction jobs fail with invalid extraction JSON.
• Drought Monitor crawl is green, but extraction jobs are failed with zero entities.
• Google Trends crawl is green, but its workflow has no extraction queue publish step in the inspected file.
• Tenant fin-agentic still has direct external fetch paths for USDA AMS, NWS weather, and openFDA. These must be disabled, migrated behind Central signal projections, or constrained to caller-supplied-payload evaluation.

Proof file:

• audit/ipa-harness-data-spine/2026-05-20/04-central-taproot-freshness-audit/central-taproot-freshness-audit.md

2. Proof Contract

Every task in this program must close with:

Proof class	Required evidence
Pixel proof	Authenticated screenshot of the relevant intranet/workspace/FIN Central surface.
Click proof	Playwright or browser proof that the user can navigate from summary to drill-down and back.
Backend proof	Tenant-scoped SQL/API inspection of events, projection rows, knowledge rows, operational memory, freshness, and permissions.
Artifact proof	Source artifact row, storage path/checksum/content hash, extraction run, and terminal disposition.
Agentic proof	Tool call or agent answer with citations, source freshness, and verifier output.
Replay proof	Idempotent replay or recompute from source events without duplicate output.

No row moves to FIXED_VERIFIED without proof artifacts stored under one of:

• .artifacts/staging-verification/<program-id>-<sha>-<timestamp>/
• output/playwright/<program-id>/<timestamp>/
• audit/<program-id>/<YYYY-MM-DD>/<slice>/
• docs/handoffs/<YYYY-MM-DD>-<program-id>-handoff.md

3. Parallel Workstreams

Lane	Owner model	Primary result	Write scope	First proof
A. IPA projection access	subagent A	Projection access matrix and missing runtime wiring	db/migrations, packages/mcp-server/src/event-handlers, scripts/cli, packages/mcp-server/src/tools, audit/card-window-system	po_timeline_projection replay + get_po_timeline tool proof
B. Harness verifiers	subagent B	Citation, freshness, schema, and egress gates	packages/agent, packages/mcp-server/src/tools, tests/evals	zero-citation answer rejected with verifier.failed event
C. Operational data spine	subagent C	PP mailbox-to-event-to-agent chain current	parsers, bridge, batch insert, freshness, proof scripts	newest arrivals@ PP reports inspectable and agent-citable
D. Email intel fusion	subagent D	intel@ source-artifact and pricing signal continuity	classifiers, extractor consumers, PDF adapters, source artifacts	Gary Blair + vendor_pricelist terminal proof
E. Sales/pricing reconciliation	subagent E	pricing intel reconciled with sales/inventory/PO/customer facts	sales parser, market rollup, customer matrix, daily sales renderer	daily_sales_situation cites sales + pricing + inventory
G. MD-first projection routing	subagent G	MD projection is preferred read surface for data/insight workflows	projection readers, agent routing, workspace rail integration, proof scripts	market/pricing prompt opens daily_sales_situation; CardSpec appears only as embedded/drill/fallback
H. Projection evidence/indexing	subagent H	MD projections become high-performance agent evidence units	MD renderer frontmatter/sidecars, projection index, retrieval tests	stable chunk ids, hashes, source refs, freshness, and role scopes inspect in backend and agent answers
I. Central Taproot freshness	subagent I	Central-only external-source gathering and fresh Central-to-tenant evidence	Knowledge Forge crawlers, central freshness reader, tenant fetch retirement, mirror/evidence contracts	AMS/PACA/weather/FDA/Federal/Register/Google Trends freshness dashboard and tenant MD evidence proof
F. Proof captain	subagent F	staging proof bundle, screenshots, click-through, backend inspection	proof scripts and handoff docs only	one JSON proof index linking all artifacts

Current environment note: the active desktop session has an agent-thread limit, so new subagents may not spawn until existing agents are closed or reused. The work is still decomposed so existing agents or the next session can pick up disjoint write sets safely.

4. Wave Plan

Wave 0 - Control, reconciliation, and current truth

Priority: P0.

Tasks:

1. Reconcile active checkout, prior integration worktree, and intranet mirror for IPA docs and Sales/Market architecture.
2. Rerun current staging status for API SHA, worker version, migrations, projection tables, role catalogs, and mailbox freshness.
3. Produce a proof index that separates historical proof from current proof.
4. Verify whether direct deploy remains the approved path for staging proof, while preserving branch-protection discipline for protected branches.

Acceptance:

• Tracker rows show current_verified_at, runtime_sha, and proof path.
• Missing docs from intranet are restored or explicitly linked.
• No historical proof is treated as current without a rerun.

Wave 1 - Projection agentic access

Priority: P0.

Tasks:

1. Implement or complete po_timeline_projection runtime wiring first: handler, dispatcher, replay descriptor, approval registry, tests, and reader get_po_timeline.
2. Implement or complete customer_commodity_matrix second: reconcile schema with architecture, handler, replay, reader get_customer_commodity_matrix, tests, and role-safe exposure.
3. Add catalog mode policy before adding more default tools if budget would exceed 30 plus runtime memory tool.
4. Add projection access matrix covering every IPA projection: projection row, MD/vault output, MCP reader, role exposure, UI route, proof.

Acceptance:

• Buyer/admin can call get_po_timeline; denied roles cannot.
• Sales/admin can call get_customer_commodity_matrix; customer/consumer roles cannot.
• Pixel/click proof shows a role can navigate to the projection and drill into source events.
• Backend proof shows projection rows are replay-supported and tenant-scoped.

Wave 2 - Operational ingestion and data spine

Priority: P0.

Tasks:

1. Rerun mailbox-to-agent proof for newest arrivals@ Produce Pro attachments.
2. Finish or verify first-class support for:
   • PO created and PO received;
   • sales report with lot/cost/quantity/customer PO/ship-via fields;
     • inventory adjustment;
     • inventory snapshot;
     • physical inventory only as real-data flow if Nathel actually sends it.
3. Build the Produce Pro report utilization matrix: every report family, attachment pattern, parsed field, event type, projection consumer, MD evidence consumer, freshness SLA, entity-resolution dependency, and parked/not-used rationale.
4. Build or complete lot lifecycle reconciliation: ordered + received + adjustments + sales + on-hand snapshot → variance.
5. Expose data freshness and reconciliation in agent tools and operator UI.

Acceptance:

• At least one real Nathel lot proves full lifecycle reconciliation.
• Agent answer to “what is actually in stock and why?” cites the projection, events, and source artifact.
• Work Thread/load-context surfaces show the same lot facts.
• Produce Pro report utilization matrix proves no high-value field is silently ignored.

Wave 3 - Email Intel Fusion and pricing reports

Priority: P0.

Tasks:

1. Rerun current intel@ proof: channel.message_received -> ingest.artifact_stored -> operational_memory -> signal.email_extract -> extractor.* -> signal.* -> knowledge/tool.
2. Build the intel classification taxonomy and organizer: source class, sender/domain rule, subject/body classifier, expected artifact types, extractor route, terminal states, freshness class, source license, confidence, and downstream consumers.
3. Implement terminal structured handling for vendor_pricelist PDFs/images/ spreadsheets: extracted, failed with resumable code, or explicit not-implemented backlog event.
4. Add Gary Blair / Hunts Point live consumer if not already active in current runtime.
5. Verify source-artifact family maintenance and content-hash instrumentation.

Acceptance:

• Gary Blair and at least one representative vendor price sheet produce normalized price observations with artifact links and freshness windows.
• Agent answers cite source basis and stale warnings.
• PDF backlog has no silent limbo state.
• intel@ artifacts inspect with classification version, terminal state, source class, downstream consumer, and owner for any backlog state.

Wave 4 - Sales and market insight projection

Priority: P0/P1.

Tasks:

1. Build market_signal_rollup from signal.market_chatter, signal.price_observation, signal.vendor_quote_observed, and signal.availability_observation.
2. Build daily_sales_situation MD renderer using: daily sales, inventory snapshot, lifecycle variance, market signal rollup, open commitments, customer demand, and pricing intel.
3. Build sight_picture.weekly after daily sales is proven.
4. Add get_sales_situation and role-appropriate rail integration only after default/drill tool-budget policy is settled.
5. Route sales/market/pricing agent reads through the MD projection first; use VendorQuoteSignal and direct market-signal tools as embedded widgets, drilldowns, diagnostics, or logged fallback only.

Acceptance:

• Sales role sees the daily_sales_situation MD projection as the primary surface with pixel/click proof.
• Agent answer combines daily sales, inventory, Gary Blair/vendor pricing, and customer demand into a recommendation with citations and confidence.
• The same source facts are inspectable in MD frontmatter, backend rows, and artifacts.
• CardSpec-only proof is insufficient; embedded card and direct-tool fallback states must be click-proven or logged.

Wave 5 - Harness gates and measurable performance

Priority: P0/P1.

Tasks:

1. Ship citation-validity verifier at harness egress.
2. Regenerate gradeboard and S0 latency baseline.
3. Patch FINN system prompt/personas, deterministic router, synthesis, and evals for MD-first projection access.
4. Add projection evidence manifests and a projection chunk/index registry for primary MD outputs.
5. Tune retrieval budgets only after baseline and only against indexed evidence units.
6. Add per-source freshness verifier.
7. Add schema conformance verifier for typed claims.
8. Add model routing by surface tier, answer-path telemetry, and denied-tool telemetry.

Acceptance:

• Evidence-required answer with zero citations is rejected.
• Stale evidence is surfaced or blocked by source class.
• Numeric/currency/unit claims are schema checked.
• Fresh MD projection evidence wins over direct tool/CardSpec fallback in deterministic routing tests.
• Projection evidence chunks expose stable ids, hashes, source refs, freshness, and role scopes.
• Gradeboard and latency artifacts show before/after movement.

5. Ranked Task Queue

Rank	ID	Severity	Lane	Task	Why now
1	IPA-CTRL-P0-001	P0	F	Build current proof index across IPA, data spine, email intel, and harness.	Prevents stale proof from masking runtime gaps.
2	IPA-P0-001	P0	A	Complete po_timeline_projection runtime and get_po_timeline.	Deterministic, unlocks recall/proof and PO spine.
3	DS-P0-001	P0	C	Verify newest arrivals@ PP report chain, including source artifacts and freshness.	Operational truth must be current before agent lift.
4	EIF-P0-001	P0	D	Verify newest intel@ chain and Gary Blair source-to-agent path.	Pricing intel is a live-risk lane.
5	EIF-P0-003	P0	D/H	Build intel@ classification taxonomy and artifact organizer.	Prevents intel mailbox growth from becoming an unstructured archive.
6	DS-P0-004	P0	C/H	Build Produce Pro report utilization matrix across every report family and field.	Ensures every useful PP field has a downstream use or explicit parked rationale.
7	DS-P0-002	P0	C/E	Make sales report fields queryable: lot, cost, ship/order qty, customer PO, ship-via.	Required to reconcile sales against lots and pricing.
8	DS-P0-003	P0	C/E	Verify or complete inventory adjustment and inventory snapshot ingestion.	Required for real stock variance.
9	IPA-P0-002	P0	A/E	Complete customer_commodity_matrix runtime and reader.	Required for demand insight and sales outreach.
10	EIF-P0-002	P0	D	Add terminal vendor_pricelist extractor/adaptor or explicit backlog event.	Ends silent PDF limbo for pricing reports.
11	MDP-P0-001	P0	E/G	Enforce MD-first read routing for operational data, insight, and decision workflows.	Prevents direct CardSpec/tool paths from becoming the primary UX and keeps the harness standardized.
12	MDP-P0-002	P0	B/G	Patch FINN prompt/personas, deterministic router, synthesis, and evals for MD-first projection law.	Makes the architecture enforceable by the harness instead of relying on doc intent.
13	DSP-P0-001	P0	H/B	Add projection evidence manifest and chunk/index registry for primary MD projections.	Gives the agent stable, fast, citation-valid evidence units instead of parsing long markdown or falling back to raw tools.
14	SM-P0-001	P0	E	Build market_signal_rollup.	Converts raw price/chatter into reusable projection input.
15	SM-P0-002	P0	E/G/H	Build daily_sales_situation MD and agent reader.	First operator-visible sales/pricing intelligence artifact and MD-first market/pricing proof target.
16	HKR-P0-001	P0	B	Citation-validity verifier.	Makes agent outputs measurable and citable.
17	HKR-P0-002	P0	B/F	Fresh gradeboard and S0 latency baseline.	Optimization without measurement is noise.
18	DSP-P1-002	P1	H/B	Add source-aware hybrid retrieval and query-planner telemetry over projection evidence chunks.	Maximizes agent performance by retrieving compact, fresh, role-safe evidence packs.
19	IPA-P1-001	P1	A/B	Default vs specialist/drill role-mode tool budgets.	Prevents projection tools from blowing up role catalogs.
20	HKR-P1-006	P1	B	Per-source freshness verifier.	Pricing and inventory evidence decay quickly.
21	HKR-P1-007	P1	B	Schema conformance verifier.	Prevents vague numeric/quantity claims.
22	OBS-P1-001	P1	B/F	Denied-tool durable telemetry and rollup.	Shows where agents are blocked from needed projection access.
23	RP-P1-001	P1	A	Proof-package central MD pins.	Deterministic legal-grade projection path.
24	ING-P1-001	P1	C/D/H	Add ingestion watermarks, dedupe/idempotency, and entity-resolution coverage rollups for high-volume mailbox/report feeds.	Keeps projections fast and trustworthy as report volume grows.
25	IPA-P2-001	P2	A/E	sight_picture.weekly and vendor/customer MD projectors.	Compound after daily sales and matrix are proven.
26	CTR-P0-001	P0	I/B	Enforce Central-only external-source gathering. Retire or gate tenant-side USDA, NWS/weather, and FDA live fetch paths; tenant tools must consume Central-derived signals or caller-supplied payloads.	Static guard + tests prove tenant runtime cannot call external source APIs for Taproot evidence; agent/tool telemetry records Central-source evidence use.
27	CTR-P0-002	P0	I/F	Build Central Taproot freshness dashboard/reader across GitHub workflow status, KF D1 crawler_health, source_documents, extraction jobs, R2 artifact existence/hash, migration status, and tenant mirror freshness.	Backend/API proof returns per-source status, freshness age, SLA, failure stage, last good artifact hash, and tenant mirror freshness.
28	CTR-P0-003	P0	I	Repair USDA AMS crawl/extraction/auth. A green crawl must not publish 401 bodies as raw artifacts, and AMS evidence must land as citable Central chunks within SLA.	Latest AMS run has valid raw artifact, successful extraction/quality, source docs, migration/serving proof, and tenant MD evidence citation.
29	CTR-P0-004	P0	I	Repair PACA license refresh. Align config/workflow cadence, stop the 3-hour cancellation path, and prove staged PACA source docs.	Weekly PACA run succeeds, D1 source docs exist, extracted entities include license status, and tenant compliance surfaces cite Central PACA evidence.
30	CTR-P0-005	P0	I	Repair Market News terminal-market workflow. Isolate source failures, allow partial success, and publish only valid source prefixes.	AMS success is not blocked by terminal-market DNS failure; terminal-market failures emit structured gap events and freshness state.
31	CTR-P0-006	P0	I	Repair post-crawl extraction/quality for FDA recalls, Federal Register, Drought Monitor, and Google Trends queue publication.	Green crawl implies either successful extraction/quality or a terminal gap event; no green-only raw crawl is counted as citable evidence.
32	CTR-P0-007	P0	I/H/G	Feed Central external signals into MD-first projection evidence manifests.	daily_sales_situation and relevant QC/weather/compliance MD projections cite Central chunks with source license, fetched/extracted timestamps, confidence, SLA, and content hash.

6. Subagent Dispatch Pack

Use these as bounded assignments when agent slots are available.

Agent A - IPA projection access

Own:

• po_timeline_projection;
• customer_commodity_matrix;
• MCP reader tools;
• approval registry and replay descriptors.

Do not touch:

• harness verifier code;
• email intel extraction;
• UI proof scripts except reader smoke.

Agent B - Harness gates

Own:

• citation verifier;
• freshness verifier;
• schema verifier;
• model routing and role-mode policy;
• FINN prompt/persona MD-first rule and deterministic eval gates.

Do not touch:

• Produce Pro parsers;
• projection migrations except denied-tool telemetry if explicitly assigned.

Agent C - Operational data spine

Own:

• Produce Pro parsers and bridge;
• mailbox-to-event proof;
• Produce Pro report utilization matrix and field-to-consumer mapping;
• inventory adjustment/snapshot/sales field completeness;
• lifecycle reconciliation.

Do not touch:

• intranet MD renderers except data contracts they consume.

Agent D - Email Intel Fusion

Own:

• intel@ / arrivals@ current proof;
• intel@ classification taxonomy and artifact organizer;
• PDF terminal state;
• vendor_pricelist extractor/adaptor;
• Gary Blair / Hunts Point live consumer.

Do not touch:

• customer matrix renderer or role catalogs.

Agent E - Sales/pricing intelligence

Own:

• market_signal_rollup;
• daily_sales_situation;
• get_sales_situation;
• sales-role workspace/rail surface.
• embedded VendorQuoteSignal and source drilldowns inside the MD-first sales situation surface.

Do not touch:

• citation verifier internals;
• raw mailbox routing.

Agent G - MD-first projection routing

Own:

• agent/tool router preference for fresh MD projections on operational data, insight, and decision prompts;
• deterministic synthesis for MD projection readers;
• fallback logging when direct MCP/CardSpec paths are used because the MD projection is missing, stale, or out of scope;
• workspace click path that opens the MD projection first and embedded widgets second.

Do not touch:

• event payload schemas;
• raw ingestion/parsing logic.

Agent H - Projection evidence and indexing

Own:

• agent_evidence frontmatter/sidecar contract for primary MD projections;
• projection chunk registry/index with stable chunk ids, hashes, entity keys, source refs, freshness, and role scopes;
• hybrid retrieval/query-planner tests over projection chunks;
• answer-path telemetry showing whether the agent used MD evidence, direct fallback, or raw tool output.

Do not touch:

• role persona tone;
• source extractor heuristics except metadata fields required by the manifest.

Agent F - Proof captain

Own:

• staging proof index;
• Playwright screenshot/click suites;
• backend inspection SQL/API scripts;
• final handoff.

Do not touch:

• production code unless proof scripts need a narrow assertion helper.

7. First Execution Move

Start with Wave 0 and the first two P0 slices:

1. Current proof index:
   • run data-spine freshness proof;
   • run source-to-signal proof;
   • run browser click-through proof;
   • inspect active staging SHA and migrations.
2. po_timeline_projection:
   • verify DDL/schema shape;
   • implement or complete runtime handler/replay/reader;
   • prove role-safe tool access and click-through.

These two moves are intentionally paired: the proof index tells us what is real today, while po_timeline_projection gives the agentic system a deterministic operational spine it can use immediately.

8. Stop Conditions

Stop and escalate if:

• a proof target is not the intended Nathel staging tenant;
• any parser extracts pricing or sales facts without source artifact lineage;
• a new table appears necessary without an approved projection registry entry, RLS, replay descriptor, and migration review;
• a customer/external surface can see vendor identity, margin, other customers, internal notes, or unrestricted inventory details;
• the agent answers with uncited operational or pricing claims after the citation verifier is in place;
• a deployment overwrites a required cron/handler without a proof rerun.