FIN Agentic Harness and Knowledge Roadmap Audit Tracker

Audit date: 2026-05-20 Scope: roadmap accuracy, agentic harness recommendations, knowledge-system recommendations, intranet synchronization implications.

0. Executive Status

The roadmap family is directionally strong and matches the architectural law: four-table substrate, event-sourced write path, projection-backed read path, and intranet/markdown as the human-agent operating surface. The external recommendations are mostly valid, but three claims need correction before execution:

1. The Operational Corpus roadmap mislabels two Produce Pro reports. Current code verifies PD335 is vendor master, PD327 is customer list, and commodity master is the separate COMMODITY report.
2. The cited “gradeboard 4.01” baseline is not verified in this checkout. The checked-in gradeboard artifact is stale, generated 2026-04-07, and reports current overall grade 9.65.
3. Citation, freshness, schema, and per-surface model-routing recommendations are valid, but they are not yet fully implemented as inline harness gates.

This tracker converts the verified gaps into ranked execution work.

1. Verification Matrix

ID	Claim / recommendation	Verdict	Evidence	Execution consequence
V-001	Citation-validity verifier is the highest-leverage harness work.	VALID, NOT YET INLINE	Self-Improving Roadmap §3.2 specifies citation validity/freshness/schema verifiers. query_knowledge and query_operational_memory return source/freshness metadata, but no global egress verifier currently rejects unresolved citation ids or zero-citation answers.	Ship as P0 before prompt/model tuning.
V-002	Freshness should be per-source, not global.	VALID, PARTIAL SUBSTRATE	query_knowledge and query_operational_memory both expose freshness, but currently set stale: false in response builders. Produce Pro freshness tracking exists for report imports.	Add source-class SLA registry and per-answer freshness verifier.
V-003	Schema conformance should reject free strings for typed slots.	VALID, PARTIAL	Roadmap specifies schema verifier. Many tools/cards use Zod/structured schemas, but there is no global post-answer numeric/unit/currency verifier across QC/comms/intranet responses.	Add schema verifier after citation verifier.
V-004	Retrieval sets are probably too small.	VALID, MEASURE FIRST	query_knowledge.limit default is 5 and max 20. query_operational_memory.limit default is 5 and max 20. FINN context seed defaults to 5 and caps at 25.	Establish baseline, then tune top-K/context budgets by surface tier.
V-005	Model routing should differ by surface tier.	VALID, NOT CURRENTLY SURFACE-TIERED	Provider factory routes by TaskType (deterministic_batch, reasoning, eval_judge) and SMS has a narrow deterministic-model override. No S0/S1/S2/S3 routing policy is present.	Add surface-tier policy and telemetry.
V-006	Measure S0 latencies before optimizing.	VALID, OPEN	Reliability Roadmap defines S0 p50/p99 budgets, and trace/SLO code exists, but the checked-in gradeboard does not include fresh QC PWA, voice, and comms TTFT baselines.	Make baseline capture a P0 gate.
V-007	PP PO ingestion is a top strategic lift.	VALID, PARTIAL	Active PO Received/PD032 ingestion is already tested as event-store facts with retention/freshness metadata. RE101/PO-created plus master/profile integration still needs closure.	Execute corrected PP PO spine W1 scope.
V-008	Operational Corpus W1 report mapping is correct as written.	INVALID	Code maps PD335 to vendor, PD327 to customer, COMMODITY to commodity master. The roadmap says PD335 (commodities), PD327 (vendors).	Correct roadmap before handing tasks to agents.
V-009	”Every event carries po_id” is execution law.	VALID WITH SCOPE CLARIFICATION	Roadmap later narrows this to every event with a PO binding. Some source classes are master data, background knowledge, or pending signals that cannot carry a po_id at ingest time.	Enforce po_id on PO-bound operational events; route unbound signals to pending_po_resolution.
V-010	Taproot curation/drift event state machine is implemented.	PLANNED, NOT IMPLEMENTED	taproot.entry.* and taproot.drift.detected appear only in roadmap docs in this checkout.	Treat as roadmap task, not current capability.
V-011	Skill packaging discipline already covers QC/comms/intranet runtime skills.	NOT VERIFIED	Only packages/openclaw-skills/pp-knowledge-pack/SKILL.md appears under packages. No QC/comms/intranet SKILL.md + evals/ + verifiers/ runtime package set is present.	Package existing skills before adding new ones.
V-012	Gradeboard 4.01 is current.	NOT VERIFIED	docs/gradeboard/gradeboard.md is generated 2026-04-07 and reports 9.65; it may be stale and not representative of the current harness.	Regenerate before claiming gradeboard lift.
V-013	Several Produce Pro reports are already entering through intel@fintail.net / arrivals@fintail.net.	VERIFIED, PARTIAL PROOF	The email webhook decodes CSV attachments, detects Produce Pro report type, calls processReport, stores source artifacts, and writes freshness/import events. The bridge recognizes stock, inventory, vendor, PO, customer, arrival, commodity, sales, and physical inventory reports.	Update W1 from “build ingestion from scratch” to “prove and complete the existing mailbox-to-agent path.”
V-014	Produce Pro PDF intel is fully operational for every classified report class.	NOT VERIFIED	arrival_report has a routed RP581/arrival extraction path into pull_pp_inbound_arrivals; vendor_pricelist classification exists, but a structured price-list PDF extractor/adaptor is still open.	Require terminal proof for every classified PDF: extracted, failed, or explicit backlog/not-implemented event.
V-015	Only FIN Central Taproot gathers external sources; tenant taproots receive Central plus operational tenant data.	VALID ARCHITECTURE, NOT ENFORCED	Knowledge Forge owns scheduled external crawlers, but fin-agentic still has tenant-side USDA AMS, NWS/weather, and openFDA fetch paths.	Add P0 boundary guard and migrate tenant tools to Central-derived evidence.
V-016	AMS, PACA, weather, and other Central data are refreshing automatically.	MIXED	NOAA weather and active-intelligence v2 are fresh in KF D1; AMS crawl is green but extraction failed on a 401 raw body; PACA is cancelled/failing with no source docs; FDA/Federal Register/Drought crawls are green but extraction/quality is failed or absent.	Treat green crawl status as insufficient; freshness proof must cover crawl, raw artifact, extraction, quality, migration/serving, and tenant evidence.
V-017	Tenant freshness surface can tell the agent whether external evidence is current.	PARTIAL	Existing get_intel_signal_freshness proof exposes source-class SLA and reports usda_ams as unknown, vendor quotes stale, and Produce Pro/source artifacts fresh.	Wire Central freshness into MD evidence manifests and harness freshness verifier.

2. Priority-Ranked Remediation Tasks

Rank	Task ID	Severity	Status	Task	Acceptance proof
1	HKR-P0-001	P0	OPEN	Implement citation-validity verifier at harness egress. Every cited chunk_id / source ref must resolve to operational_memory, knowledge, or doc-store content; evidence-requiring answers with no citations return a structured insufficient-evidence response.	Unit tests for valid, missing, stale, and zero-citation paths; integration test through QC/comms agent response boundary; verifier.failed event proof.
2	HKR-P0-002	P0	OPEN	Regenerate gradeboard and capture S0 latency baseline before tuning. Measure p50/p95/p99 input-to-first-byte for QC PWA, voice agent, and comms reply.	Fresh docs/gradeboard/gradeboard.md/json or artifact bundle with generated timestamp; latency artifact with surface, route, model, retrieval budget, and TTFT distribution.
3	HKR-P0-003	P0	DOC_CORRECTION_AND_EXECUTION	Correct Produce Pro W1 report mapping and finish PO spine ingestion. Treat PD335 as vendor, PD327 as customer, COMMODITY as commodity master; complete RE101/PO-created + active PO received + supporting masters + watermelon operational profile.	Parser/bridge tests for corrected report mapping; PO-bound event proof with po_id; watermelon commodity_profile rendered from real operational data; agent answer with citations under target latency.
4	HKR-P0-004	P0	OPEN	Run mailbox-to-agent proof for Produce Pro reports already flowing through intel@ / arrivals@. Inventory the newest report attachments, prove source artifacts, CSV bridge imports, PDF arrival routing, normalized events/signals, freshness rows, Work Thread/load-context surfaces, and agent-cited answers.	Proof artifact for chain: channel.message_received -> ingest.artifact_stored -> bridge.report_imported or ingest.pdf_extracted -> signal.* -> knowledge/operational_memory -> agent response. Vendor-pricelist PDFs must either extract structured price observations or land in an explicit terminal backlog state.
5	EIF-P0-003	P0	OPEN	Build the intel@fintail.net classification taxonomy and artifact organizer before intel volume grows: closed source classes, classifier rules, extractor route, terminal states, freshness class, source license/class, confidence, and downstream consumers.	Current mailbox sample proves every artifact has classification version, source class, terminal state, content hash, extractor route, freshness class, downstream consumer, and owner for backlog/not-implemented states.
6	DS-P0-004	P0	OPEN	Build Produce Pro report utilization matrix across every report family and field. Every high-value field must map to an event payload, JSONB projection, MD evidence chunk, freshness row, entity-resolution candidate, or explicit parked/not-used rationale.	Matrix covers stock, inventory, vendor, PO, customer, arrival, commodity, sales, physical inventory, inventory adjustment, and snapshot reports; proof shows no high-value parsed field is silently ignored.
7	HKR-P0-005	P0	OPEN	Update FINN system prompt, Sales/Buyer/Executive personas, deterministic routing, synthesis, and evals for the MD-first projection law. Fresh MD projections must beat direct CardSpec/tool readers for operational data, insight, and decision prompts.	Unit tests prove fresh-MD-wins, stale/missing-MD-falls-back, direct fallback is logged, and sales/pricing answers cite MD projection evidence; persona prompt hash/manifest updates are inspected.
8	DSP-P0-001	P0	OPEN	Add agent evidence manifests and a projection chunk/index registry for primary MD projections. MD is the UX surface; stable chunks are the agent retrieval unit.	daily_sales_situation and one QC MD expose projection_doc_id, content_hash, section chunk_ids, source refs, freshness class, role scope, and citation tier; backend/index proof and agent citation proof pass.
9	HKR-P1-004	P1	OPEN	Tune retrieval budgets after baseline. Increase top-K/context budget where latency allows; preserve S0 budget with escalation/cached-context behavior and indexed projection evidence packs.	Before/after gradeboard and TTFT comparison; regression guard that S0 remains inside p99 target.
10	HKR-P1-005	P1	OPEN	Add surface-tier model routing policy. S0 fast/fallback, S1 higher quality/cached, S2 deterministic/batch, S3 cheap/batch/eval-safe.	Routing tests proving policy selection; telemetry row/event containing surface tier, selected model, fallback, timeout, and cost bucket.
11	HKR-P1-006	P1	OPEN	Add freshness verifier with per-source SLA registry. USDA AMS, operational inventory, procedure docs, AGROVOC/public taxonomy, and curated Taproot entries must each have independent freshness windows.	Verifier tests for current/stale citations by source class; agent response exposes evidence staleness; stale evidence blocks or banners according to question type.
12	HKR-P1-007	P1	OPEN	Add schema conformance verifier for typed claims. Numeric claims, units, currency, precision, PO ids, lot ids, and quantities must serialize to expected schemas.	Tests for rejecting free-string typed slots and accepting conformant claims; failure events include field path and expected schema.
13	HKR-P1-008	P1	OPEN	Audit and enforce cite-then-claim at MCP boundary. Read tools return retrieval metadata; write tools only emit events; tool-returned content must carry source attribution/hash.	Boundary test matrix for empty evidence, invalid source refs, free-string typed slots, read/write separation, and tool content hashes.
14	DSP-P1-002	P1	OPEN	Add source-aware hybrid retrieval and query-planner telemetry over projection evidence chunks. Prefer compact, fresh, role-safe evidence packs over large markdown bodies or direct raw tools.	Retrieval tests combine keyword/vector/entity filters; telemetry shows selected evidence pack, stale/fallback reason, token budget, source classes, and latency.
15	ING-P1-001	P1	OPEN	Add ingestion watermarks, dedupe/idempotency, and entity-resolution coverage rollups for high-volume mailbox/report feeds.	Duplicate/replay tests, watermark inspection, unresolved-entity rollup, and projection lag metrics prove high-volume ingestion remains fast and replay-safe.
16	HKR-P2-009	P2	OPEN	Package existing QC, comms, and intranet-rendering skills before adding new runtime skills. Each skill gets SKILL.md, evals/, verifiers/, version, owner, and promotion gate.	Three package directories with manifests/evals/verifiers; CI/eval gate fails on skill change without verifier pass.
17	HKR-P2-010	P2	OPEN	Implement Taproot curation state machine and drift queue. taproot.entry.* and taproot.drift.detected become real events with intranet review queue projection.	Event contract tests; curation queue MD/intranet projection; drift fixture emits review item with diff context.
18	HKR-P2-011	P2	OPEN	Implement self-improvement layer after deterministic gates: procedure profiles, case memory, counterfactual replay, confidence calibration.	First watermelon procedure profile; case_memory accumulation proof; counterfactual replay on a real PO chain; confidence calibration artifact.
19	CTR-P0-001	P0	OPEN	Enforce Central-only external-source gathering. Tenant runtime must not call USDA, PACA, FDA, NOAA/NWS, AGROVOC, Federal Register, Google Trends, or market/news public APIs for Taproot evidence.	Static guard and runtime tests prove tenant-side tools consume Central signal projections, Central mirror docs, operational tenant data, or caller-supplied payloads only.
20	CTR-P0-002	P0	OPEN	Build Central Taproot freshness dashboard/reader.	API/backend proof joins GitHub workflow status, KF D1 crawler_health, source_documents, extraction jobs, R2 hash/existence, migration status, tenant mirror age, SLA, and failure stage.
21	CTR-P0-003	P0	OPEN	Repair USDA AMS auth/extraction and make green crawls fail when raw artifacts are 401/error pages.	Latest AMS run produces valid raw artifact, extraction/quality success, citable Central chunks, and tenant MD evidence inside SLA.
22	CTR-P0-004	P0	OPEN	Repair PACA refresh. Align weekly config with workflow cadence and stop the 3-hour cancellation path.	PACA source docs and license-status entities appear in Central D1 and tenant compliance projections with source license/fetched timestamp.
23	CTR-P0-005	P0	OPEN	Repair Market News/terminal-market workflow with partial-source success and structured gap events.	Terminal-market DNS failures no longer block valid AMS artifacts; failed markets emit freshness/gap rows.
24	CTR-P0-006	P0	OPEN	Repair FDA, Federal Register, Drought Monitor, and Google Trends post-crawl extraction/quality/queue publication.	No source is considered citable from green crawl alone; each source has extraction success or explicit terminal gap state.
25	CTR-P0-007	P0	OPEN	Feed Central external signals into MD-first projection evidence manifests.	Sales, QC/weather, and compliance MD projections cite Central chunks with source, source license, fetched/extracted timestamps, confidence, content hash, and freshness SLA.

3. Execution Guidance

Do these in waves:

1. Wave A - Measurement and gates: HKR-P0-001, HKR-P0-002.
2. Wave B - Operational truth: HKR-P0-003 and HKR-P0-004.
3. Wave C - Budget/routing optimization: HKR-P1-004, HKR-P1-005.
4. Wave D - Evidence discipline: HKR-P1-006, HKR-P1-007, HKR-P1-008.
5. Wave E - Compounding loop: HKR-P2-009 through HKR-P2-011.

Avoid prompt-only tuning until Wave A gates exist. Avoid adding new skills until HKR-P2-009 packages the existing high-value skills. Avoid TURVO/Samsara optimization until the PP PO spine has real po_id gravity.

4. Intranet Sync Requirements

These docs carry intranet-style frontmatter and should be synced into fin-platform/fin-central-intranet when promoted:

• docs/architecture/FIN_PLATFORM_NORTH_STAR_ROADMAP_2026-05-19.md
• docs/architecture/FIN_RELIABILITY_AND_OPS_ROADMAP_2026-05-19.md
• docs/architecture/FIN_SELF_IMPROVING_SYSTEM_ROADMAP_2026-05-19.md
• docs/architecture/FIN_TAPROOT_KNOWLEDGE_CURATION_ROADMAP_2026-05-19.md
• docs/architecture/FIN_OPERATIONAL_CORPUS_INGESTION_ROADMAP_2026-05-19.md
• docs/programs/FIN_AGENTIC_HARNESS_KNOWLEDGE_ROADMAP_AUDIT_TRACKER_2026-05-20.md

Promotion is not complete until the intranet repo has the same corrected markdown and a fetch-back proof confirms the files are present under docs/architecture/ and docs/programs/.

5. Integrated IPA / Harness / Data-Spine Addendum - 2026-05-20

The intranet projection work, harness verifier work, operational ingestion spine, email intel fusion, and sales/pricing projection architecture are now tracked as one execution program:

• docs/programs/FIN_IPA_HARNESS_DATA_SPINE_PARALLEL_EXECUTION_PLAN_2026-05-20.md

New ranked rows opened by this reconciliation:

Rank	Task ID	Severity	Status	Task	Acceptance proof
13	IPA-CTRL-P0-001	P0	CLOSED_VERIFIED	Build a current proof index across IPA projections, data spine, email intel, and harness gates. Separate historical proof from current staging/runtime proof.	Closed in audit/ipa-harness-data-spine/2026-05-20/00-current-proof-index/proof-index.json; names runtime version dc311ae1-ipa-ds-20260520-r2, worker version b9307661-d85f-4142-aa3f-929856755d4d, proof artifact paths, and current pass/fail per lane.
14	IPA-P0-001	P0	BACKEND_AND_AGENTIC_VERIFIED_UI_DRILL_OPEN	Complete po_timeline_projection runtime and agent access: handler, dispatcher, replay descriptor, approved projection registry, MCP reader, permissions, role catalog, tests.	Backend/agentic closed: replay processed 573 events into 9 rows; get_po_timeline returned PO 00549622 via /api/tools/call. UI-specific PO drill-through remains open.
15	DS-P0-002	P0	CLOSED_VERIFIED	Make Produce Pro sales rows reconciliation-ready by promoting LOT NUMBER, UNIT COST, SHIP QTY, ORDER QTY, CUSTOMER P.O. #, and SHIP VIA into first-class parsed fields and emitted event payloads.	Closed by parser/bridge tests and staging produce-pro-sales-physical-r2-2026-05-20 proof, 13/13 assertions passed.
16	DS-P0-003	P0	CLOSED_VERIFIED	Verify or complete inventory_adjustment and inventory_snapshot ingestion as the live inventory variance anchors.	Closed by staging import_pp_data proof: 2 inventory.adjustment_posted rows, 2 inventory.snapshot rows, freshness for both domains is fresh.
17	IPA-P0-002	P0	OPEN	Complete customer_commodity_matrix runtime and agent access with safe role exposure.	Sales/admin can call get_customer_commodity_matrix; customer/consumer roles denied; projection replay and customer drill-through proof pass.
18	EIF-P0-002	P0	OPEN	Add terminal structured handling for vendor_pricelist PDFs/images/spreadsheets.	Representative vendor price sheet is extracted into price observations or lands in explicit resumable failure/not-implemented event; no silent PDF limbo.
19	MDP-P0-001	P0	OPEN	Make MD projections the primary read surface for agentic operational data, insight, and decision interactions; demote direct tool/CardSpec data cards to embedded, drilldown, diagnostic, or logged fallback paths.	Market/pricing proof shows agent/UI opens daily_sales_situation or equivalent MD projection first; backend proof inspects MD frontmatter source_events[], source_artifacts[], content_hash, freshness, and renderer/projection version; standalone CardSpec/direct-tool proof is accepted only as data-path or fallback proof.
20	MDP-P0-002	P0	OPEN	Patch FINN system prompt/personas, deterministic router, synthesis, and evals so MD-first projection access is enforced by the harness.	Fresh daily_sales_situation routes before query_market_signal; stale/missing MD falls back with telemetry; eval proves sales/pricing answer cites MD projection evidence.
21	DSP-P0-001	P0	OPEN	Add agent evidence manifest and chunk/index registry to primary MD projections.	daily_sales_situation chunks expose stable ids, hashes, source refs, freshness, role scopes, and citation tiers; harness citation verifier can resolve the chunk ids.
22	EIF-P0-003	P0	OPEN	Build intel@fintail.net classification taxonomy and artifact organizer.	Every current intel artifact has source class, classifier version, terminal state, hash, freshness class, downstream consumer, and backlog owner when applicable.
23	DS-P0-004	P0	OPEN	Build Produce Pro report utilization matrix across every report family and field.	Every high-value PP field is used downstream or deliberately parked with rationale; matrix links fields to events, projections, MD evidence chunks, freshness, and entity resolution.
24	SM-P0-001	P0	OPEN	Build market_signal_rollup from price observations, vendor quotes, market chatter, and availability observations.	Replayed rollup rows inspect cleanly and feed daily_sales_situation with source attribution.
25	SM-P0-002	P0	OPEN	Build daily_sales_situation MD projection and agent reader by reconciling daily sales, pricing intel, inventory, lifecycle variance, and customer demand.	Pixel/click proof for sales role MD surface; backend proof of MD/frontmatter/source links; agent answer cites sales + pricing + inventory sources; embedded VendorQuoteSignal is a widget/drilldown, not the primary surface.
26	IPA-P1-001	P1	OPEN	Add default/specialist/drill role-mode tool-budget policy before expanding projection tools.	Runtime catalogs stay within default budget; drill/specialist mode exposes extra projection tools with telemetry.
27	DSP-P1-002	P1	OPEN	Add source-aware hybrid retrieval and query-planner telemetry over projection evidence chunks.	Retrieval tests show keyword/vector/entity filtering over MD projection chunks; telemetry records evidence-pack choice, fallback reason, token budget, and latency.
28	ING-P1-001	P1	OPEN	Add ingestion watermarks, dedupe/idempotency, and entity-resolution coverage rollups for mailbox/report feeds.	Replay duplicates do not double-count; unresolved vendor/customer/commodity keys are visible; projection lag and import watermarks are inspectable.
29	OBS-P1-001	P1	OPEN	Persist denied-tool telemetry and add bounded admin rollup.	Durable system.tool_error or equivalent denial events inspectable by day/role/tool/code; admin reader returns grouped counts.
30	RP-P1-001	P1	OPEN	Add central MD pins to proof-package deterministic rendering before expanding recall/proof MDs.	Frozen central_md_pins[] appear in frontmatter/body and changed pin hash changes content hash while shuffled pin order does not.

Execution order: complete IPA-CTRL-P0-001 first, then run IPA-P0-001, DS-P0-002, DS-P0-003, and EIF-P0-002 in parallel lanes where agent slots are available. MDP-P0-001, MDP-P0-002, and DSP-P0-001 are gating standards for Sales/Market work: do not accept direct CardSpec/tool proof as final UX proof, and do not force the agent to parse long markdown when stable projection chunks should exist. Do not treat Sales/Market projection work as complete until SM-P0-001 and SM-P0-002 prove pricing intel is reconciled with sales, inventory, PO, and customer-demand facts through the MD-first read surface and the harness can cite the projection evidence chunks.

6. Runtime Reconciliation - 2026-05-20T02:15Z

Current staging proof is now pinned to deploy version dc311ae1-ipa-ds-20260520-r2 and Worker version b9307661-d85f-4142-aa3f-929856755d4d.

Closed in this pass:

• IPA-CTRL-P0-001: current proof index created and reconciled.
• DS-P0-002: sales parser/event payload fields promoted and staging sales + physical inventory proof passed.
• DS-P0-003: inventory adjustment/snapshot import, event-store write, and freshness serving passed.
• IPA-P0-001 backend/agentic scope: projection table shape synced, replay applied, and get_po_timeline returned projection data through /api/tools/call.
• Market-intel source-to-signal runtime: modern signal.price_observation rows now feed query_market_signal; browser clickthrough renders the VendorQuoteSignal card with median price and source attribution.

Primary current proof artifacts:

• audit/ipa-harness-data-spine/2026-05-20/00-current-proof-index/proof-index.json
• .artifacts/staging-verification/ipa-harness-data-spine-current-2026-05-20/deploy-metadata-after-r2-deploy.json
• .artifacts/staging-verification/market-intel-source-to-signal-r2-2026-05-20/2026-05-20T02-07-30-488Z/market-intel-source-to-signal-proof.json
• .artifacts/staging-verification/market-intel-browser-clickthrough-r2-2026-05-20/2026-05-20T02-14-52-095Z/buyer-market-intel-clickthrough-proof.json
• .artifacts/staging-verification/operational-data-spine-r2-2026-05-20/operational-data-spine-freshness-gap-staging-proof.json
• .artifacts/staging-verification/produce-pro-sales-physical-r2-2026-05-20/produce-pro-sales-physical-staging-proof.json
• audit/ipa-harness-data-spine/2026-05-20/02-po-timeline-backend-proof/get-po-timeline-tool.json
• audit/ipa-harness-data-spine/2026-05-20/03-inventory-adjustment-snapshot-proof/event-store-and-freshness-inspection.json

Open items after this pass:

• Current real-mailbox terminal-state inspection for newest intel@fintail.net and arrivals@fintail.net artifacts, especially PDFs and vendor/pricelist inputs.
• UI-specific PO drill-through proof for po_timeline_projection.
• customer_commodity_matrix, market_signal_rollup, and daily_sales_situation runtime/projection work.
• MDP-P0-001: current VendorQuoteSignalCard proof is valid fallback/widget and data-path proof only; market/pricing workflows still need MD-first surface proof.
• MDP-P0-002: FINN prompt/persona, deterministic router, synthesis, and evals still need to enforce the MD-first rule.
• DSP-P0-001: MD projections still need machine-readable evidence manifests and stable chunk/index registry entries for fast, citation-valid agent retrieval.
• ING-P1-001: high-volume mailbox/report ingestion still needs explicit watermark, idempotency, dedupe, projection-lag, and unresolved-entity rollup proof before it can be called optimal.
• EIF-P0-003: intel@fintail.net still needs a closed classification taxonomy and artifact organizer so every artifact has a source class, terminal state, downstream consumer, and owner.
• DS-P0-004: Produce Pro report ingestion still needs a field utilization matrix proving every high-value parsed field is used or deliberately parked.
• Citation verifier, gradeboard/S0 latency baseline, freshness verifier, schema conformance verifier, and skill packaging discipline.

7. Data And Harness Performance Assessment - 2026-05-20

Current judgment: the data architecture is strong but not yet optimal.

Strong and retained:

• Event store as source of truth.
• JSONB projections for replayable, fast operational state.
• Vault MD projections as the standardized human/agent surface.
• MCP tools as scoped read/write boundaries.

Missing for maximum agent performance:

• Projection evidence manifests and stable chunk ids so the agent retrieves compact evidence units instead of entire MD bodies.
• Source-aware hybrid retrieval over projection chunks, knowledge, and operational_memory, with tenant/role/entity filters.
• Answer-path telemetry showing whether an answer used MD evidence, JSONB projection rows, operational memory, knowledge, or fallback tools.
• Ingestion idempotency/watermarks/entity-resolution rollups so replay and mailbox volume do not silently distort projections.
• Intel mailbox classification taxonomy so external market intel does not become a loosely organized archive.
• Produce Pro field utilization matrix so report data does not stop at ingestion and every useful field contributes to a projection, evidence chunk, entity-resolution candidate, freshness signal, or workflow.
• Per-source freshness and schema verifiers connected to the same chunk/source metadata the agent cites.

Therefore the storage layout should not be replaced. It should be hardened with agent evidence manifests, retrieval indexes, idempotent ingestion controls, and deterministic harness routing.

8. Central Taproot Freshness Reconciliation - 2026-05-20T03:10Z

Proof file:

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

Verified current state:

• Central weather is fresh through Knowledge Forge active-agent.
• Central active-intelligence v2 is fresh.
• AMS is not fresh/citable end to end because the latest green crawl produced a 401 raw artifact and failed extraction.
• PACA is not refreshing into staged source docs.
• Market-news terminal markets are failing before queue publication.
• FDA recalls and Federal Register are crawling but failing extraction.
• Drought Monitor is crawling but not producing successful extraction jobs.
• Google Trends crawls weekly but lacks queue publication in the inspected workflow.
• Tenant fin-agentic still contains direct external source fetchers, so the Central-only external-source boundary is not yet enforced.

Execution implication: Central freshness and boundary enforcement now sit ahead of sales/market intelligence expansion. daily_sales_situation and market rollups should consume Central-derived external evidence only after CTR-P0-001 through CTR-P0-007 are underway or explicitly waived for a bounded staging experiment.