⚙️ Harness Engineering Sequence
The engineered loop that runs and evaluates the X-Ray Underwriting Agent — a disciplined sequence from objective to stored result, with a verifier, tiered human gates and dynamic feedback at every stage. Pick any of the 56 industry profiles and the Demo runs the whole thing live on that profile's own numbers (no backend required).
56 industry profiles7 archetypes10-stage sequenceColour-coded feedbackSuccess · Warning · Failure
CompletedIn progressWarningFailed / criticalNot started
1
🎯 Define the objective / task Pending
State the decision-relevant goal and success criteria for this run.
2
⚙️ Configure the harness environment Pending
Ephemeral sandbox, egress allowlist, scoped credentials, audit chain.
3
🧠 Select the model / agent under evaluation Pending
Pick the system-under-test and its role, model tier and permissions.
4
🧩 Prepare input, prompt, tools & test data Pending
Assemble the case: captured evidence, prompt, tool allowlist, test fixture.
5
▶️ Execute the harness Pending
Kick off the planner → executor → verifier loop against the case.
6
📡 Monitor execution in real time Pending
Stream planner decisions, tool calls and verifier verdicts as they happen.
7
📥 Capture output & intermediate actions Pending
Record the evidence graph, tool I/O and the final scenario produced.
8
🧪 Evaluate output against criteria Pending
Score against the gates: interval width ≤±20%, IRI floor, concordance.
9
💬 Generate feedback, errors & recommendations Pending
Produce scores, flag issues and propose concrete improvements.
10
🗄️ Store results for comparison & reporting Pending
Persist the run to the registry for A/B, dashboards and future iterations.
Execution panel
Empty — start the demo to stream the harness execution here.
Feedback & evaluation
Feedback appears when the run reaches the evaluation stage.
Harness coverage — all 56 industry profiles
Every profile is wired into the same harness: its own drivers, evidence ladder, benchmark floors and turnover band. Filter the list, then click a row (or Run) to load that profile into the Demo and watch the harness execute it.
4Food & Beverage
13Retail
10Manufacturing
19Services
4Transport
5Warehouse & Trade
1Scrap & Recycling
| Fixture | Profile | Type | Turnover band | Key driver | Primary evidence (Eq-7 next) | Interval | |
|---|---|---|---|---|---|---|---|
| P01-CAT-001 | Catering Services event/order batch, seasonal | Food & Beverage | ₹5–40 L/mo | Events / serving day | Event order/booking register (12 months) | ±34% → ±13% | |
| P02-RES-002 | Restaurant & Fast Food dine-in + takeaway + aggregator | Food & Beverage | ₹9–35 L/mo | Covers (seats) | Timed peak-hour occupancy count | ±34% → ±13% | |
| P03-TEA-003 | Tea Stall micro cart / kiosk, cash-heavy | Food & Beverage | ₹1–2 L/mo | Customers / spot / day (very high) | Timed cups-per-hour count (morning + evening peaks) | ±34% → ±13% | |
| P04-TIF-004 | Tiffin Services subscription meals, recurring | Food & Beverage | ₹1–9 L/mo | Subscribers / meals dispatched per day | Subscriber / route roster (current month) | ±34% → ±13% | |
| P15-AUT-015 | Auto Spare Parts Dealer automotive spares (2W/4W/CV) — counter retail + garage trade credit | Retail | ₹8–35 L/mo | Transactions / day | Counter bill/challan sequence sample | ±30% → ±13% | |
| P16-FOO-016 | Footwear Store footwear retail (family/branded/fashion), seasonal & discount-driven | Retail | ₹6–25 L/mo | Transactions / day | POS/Z-report day-total | ±30% → ±13% | |
| P17-GAR-017 | Garment – Readymade Cloth Store readymade apparel retail (family/ethnic/fashion), seasonal | Retail | ₹7–30 L/mo | Transactions / day | POS/Z-report day-total | ±30% → ±13% | |
| P18-HAR-018 | Hardware, Paint & Sanitaryware Store building-material trade + retail (hardware, paint tinting, sanitaryware) | Retail | ₹12–50 L/mo | Transactions / day | Counter bill/challan sequence sample | ±30% → ±13% | |
| P19-JEW-019 | Jewellery Making & Retail gold/silver jewellery retail + making, hallmarked | Retail | ₹30 L–1.5 Cr/mo | Average ticket | Making/wastage-charge bill sample | ±30% → ±13% | |
| P20-KIR-020 | Kirana Shop grocery / daily-needs | Retail | ₹5–20 L/mo | Transactions / day | POS / day-book Z-total | ±30% → ±13% | |
| P21-MOB-021 | Mobile & Electronics Dealer handsets, accessories & consumer electronics with EMI/finance mix | Retail | ₹25–90 L/mo | Transactions / day | Financier/EMI activation register | ±30% → ±13% | |
| P22-NUR-022 | Nursery & Floriculture plant nursery + on-site production | Retail | ₹1–8 L/mo | Sales / day | Grow-area / polyhouse extent photo | ±30% → ±13% | |
| P23-PAN-023 | Pan Shop micro-retail kiosk (pan / tobacco / confectionery) | Retail | ₹1–6 L/mo | Sales / day | Cigarette/pan-masala restock slip | ±30% → ±13% | |
| P24-PHA-024 | Pharmacy regulated chemist / medical store | Retail | ₹6–28 L/mo | Bills / day | Billing-software Z-report | ±30% → ±13% | |
| P25-RET-025 | Retail Store general small-format retail | Retail | ₹5–22 L/mo | Bills / day | POS Z-report / bill book | ±30% → ±13% | |
| P26-STA-026 | Stationery & Photocopy stationery retail + copy/print service | Retail | ₹1–6 L/mo | Transactions / day | Photocopier page-counter reading | ±30% → ±13% | |
| P27-VEG-027 | Vegetable Shop fresh produce / perishable | Retail | ₹2–9 L/mo | Sales / day | Mandi purchase slip (kacha) | ±30% → ±13% | |
| P05-AGR-005 | Agri Processing Unit (Dal / Spice / Oil Mill) dal / spice / edible-oil processing | Manufacturing | ₹50 L–1.5 Cr/mo | Finished output / hr | Mill/expeller nameplate rating | ±33% → ±13% | |
| P06-CAR-006 | Carpentry Business made-to-order wood furniture & fittings, semi-skilled labour | Manufacturing | ₹2–9 L/mo | Order-fill utilisation | Order pad / advance-booking register | ±33% → ±13% | |
| P07-ENG-007 | Engineering Works general fabrication / machining job-shop (lathes, CNC, milling) | Manufacturing | ₹5–15 L/mo | Order-fill utilisation of capacity | Open order book / job register | ±33% → ±13% | |
| P08-FAB-008 | Fabrication – Welding Works steel gates / grills / structures, per-kg or per-job, workshop + site | Manufacturing | ₹3–13 L/mo | Order-fill utilisation | Job order book / site work orders | ±33% → ±13% | |
| P09-FLO-009 | Flour Mill / Rice Mill grain milling (atta / rice) with by-products | Manufacturing | ₹60 L–2.0 Cr/mo | Milled output / hr | Roller-body/huller nameplate (TPH) | ±33% → ±13% | |
| P10-FOO-010 | Food Processing – Snacks / Namkeen namkeen / snacks batch manufacturing (FSSAI, branded) | Manufacturing | ₹30 L–1.1 Cr/mo | Packed output / hr | Auto-packing machine pouch counter | ±33% → ±13% | |
| P11-FUR-011 | Furniture Manufacturing batch + custom furniture, RM (wood/board/foam), showroom + FG inventory | Manufacturing | ₹8–35 L/mo | Capacity utilisation (batch + order) | Production plan / order + dealer book | ±33% → ±13% | |
| P12-GAR-012 | Garment Manufacturing – Job-work job-work stitching / CMT at scale (per-piece rate) | Manufacturing | ₹10–45 L/mo | Pieces stitched / hr (all lines) | Installed machines × occupied operator seats | ±33% → ±13% | |
| P13-PLA-013 | Plastic & Packaging Unit injection / extrusion / blown-film converting | Manufacturing | ₹35 L–1.2 Cr/mo | Converted output / hr | Injection/extruder shot/output counter | ±33% → ±13% | |
| P14-SPI-014 | Spinning & Weaving Plant textile spinning / weaving (spindles & looms) | Manufacturing | ₹80 L–2.2 Cr/mo | Yarn/fabric-equiv output / hr | Installed spindle/loom count + machine nameplate | ±33% → ±13% | |
| P29-ARC-029 | Architect Firm small architecture/design practice (residential + small commercial) | Services | ₹4–20 L/mo | Projects billed / month | Signed engagement letters / proposals | ±30% → ±13% | |
| P30-BEA-030 | Beauty Salon unisex / ladies salon & grooming | Services | ₹2–8 L/mo | Clients served / day (chairs × turns) | Chair/station photo (installed capacity) | ±30% → ±13% | |
| P31-CAC-031 | Chartered Accountant / Company Secretary Practice CA/CS professional practice (compliance retainers + audit/tax/filing fees) | Services | ₹2–12 L/mo | Clients billed / month | Client / filing register (portal login count) | ±30% → ±13% | |
| P32-COA-032 | Coaching / Training Institute competitive-exam / skills coaching (classroom) | Services | ₹5–30 L/mo | Enrolled students (classrooms × batch size) | Admission / enrolment register (term) | ±30% → ±13% | |
| P33-CON-033 | Consultancy management/IT/engineering advisory (retainers + projects) | Services | ₹5–22 L/mo | Active engagements / month | Retainer agreements / SOWs on file | ±30% → ±13% | |
| P34-DIA-034 | Diagnostic Lab / Pathology pathology / diagnostic collection & testing | Services | ₹3–20 L/mo | Tests / day | Test register / LIS day-count | ±30% → ±13% | |
| P35-EVE-035 | Event Management & Decoration weddings / corporate events / decor & production | Services | ₹8–50 L/mo | Avg event value | Signed event contracts (last 6) | ±30% → ±13% | |
| P36-FIT-036 | Fitness Center / Gym membership gym & group fitness | Services | ₹2–12 L/mo | Active members | Active-member register / app dashboard | ±30% → ±13% | |
| P37-GEN-037 | Generic G2 — Self-Employed Professional fallback: professional not otherwise listed (lawyer, small consultant, freelance pro) | Services | ₹2–12 L/mo | Clients / month | Appointment/client register (3 mo) | ±30% → ±13% | |
| P38-GEN-038 | Generic G3 — Self-Employed Non-Professional fallback: small non-professional self-employed (operator/vendor not otherwise listed) | Services | ₹1–5 L/mo | Jobs / day | Cash day-book / job tally (2 wk) | ±30% → ±13% | |
| P39-INS-039 | Insurance Agency commission-based agency (life / health / motor / general) | Services | ₹2–7 L/mo | Policies issued / month | Policy issuance register / portal count | ±30% → ±13% | |
| P40-INT-040 | Interior Designer / Decorator turnkey interior design + execution studio | Services | ₹4–25 L/mo | Projects handed over / month | Signed project contracts + advance receipts | ±30% → ±13% | |
| P41-MOB-041 | Mobile & Appliance Repair mobile / electronics / small-appliance repair counter | Services | ₹1–7 L/mo | Jobs / day | Repair job-ticket register (7-day) | ±30% → ±13% | |
| P42-REA-042 | Real Estate Agency property brokerage (resale + rental + primary-sales channel) | Services | ₹2–18 L/mo | Deals closed / month | Deal register / agreement-to-sell records | ±30% → ±13% | |
| P43-REG-043 | Registered Medical Practitioner / Clinic solo/small OPD clinic (GP/specialist) with dispensing add-on | Services | ₹3–9 L/mo | Consults / day | Appointment/OP register day-count | ±30% → ±13% | |
| P44-TAI-044 | Tailoring Services stitching / alterations / boutique job-work | Services | ₹1–5 L/mo | Garments / day | Order/job-ticket book tally | ±30% → ±13% | |
| P45-TOU-045 | Tour & Travel Agency air / hotel / package booking & ticketing | Services | ₹15–80 L/mo | Bookings / day | GDS/PNR issuance summary (monthly) | ±30% → ±13% | |
| P46-TUI-046 | Tuition Agency school tuition / home & small-premises coaching | Services | ₹1–4 L/mo | Enrolled students (batches × students) | Enrolment / attendance register | ±30% → ±13% | |
| P47-VEH-047 | Vehicle Service car/two-wheeler workshop & garage | Services | ₹3–18 L/mo | Jobs / day (bays × throughput) | Workshop bay & lift photo (capacity) | ±30% → ±13% | |
| P48-CAB-048 | Cab / Taxi Operator (aggregator) owned cars on Ola/Uber aggregator platforms | Transport | ₹2–9 L/mo | Trips / vehicle / day | Aggregator earnings dashboard (weekly trips) | ±35% → ±13% | |
| P49-COM-049 | Commercial Vehicles Business fleet operator running commercial vehicles on hire / contract | Transport | ₹5–50 L/mo | Trips / vehicle / day | Trip sheet / contract log | ±35% → ±13% | |
| P50-COU-050 | Courier / Last-mile Logistics bike/van fleet + riders on hub / franchise model | Transport | ₹2–25 L/mo | Parcels / rider / day | Hub manifest / parcels-scanned report | ±35% → ±13% | |
| P51-GOO-051 | Goods Transport Operator owned trucks / tempos on B2B freight | Transport | ₹5–40 L/mo | Loads / vehicle / day | Trip register / lorry receipts (LR) | ±35% → ±13% | |
| P52-COL-052 | Cold Storage refrigerated storage rental (agri / perishables) | Warehouse & Trade | ₹4–25 L/mo | Effective rent yield / sq ft / day (net of vacancy) | DISCOM bill (12-month kWh profile) | ±30% → ±13% | |
| P53-FMC-053 | FMCG Distributor – Stockist fast-moving stockist / super-distributor | Warehouse & Trade | ₹25–90 L/mo | Dispatch value / sq ft / day | Daily beat / van load-out sheet | ±30% → ±13% | |
| P54-GOD-054 | Godown Owner dry storage rental / warehousing | Warehouse & Trade | ₹1–7 L/mo | Effective rent yield / sq ft / day (net of vacancy) | Occupancy read (bays/racks filled vs total) | ±30% → ±13% | |
| P55-TRA-055 | Trader (General Goods) wholesale / buy-sell trading | Warehouse & Trade | ₹6–30 L/mo | Sales / sq ft / day | Dispatch register / outward invoices | ±30% → ±13% | |
| P56-WHO-056 | Wholesale Business C&F carrying & forwarding agent / super-stockist | Warehouse & Trade | ₹20–90 L/mo | Throughput value / sq ft / day | Dispatch / GRN register | ±30% → ±13% | |
| P28-SCR-028 | Scrap Dealer Shop mixed-scrap buy-sell (shop scale) | Scrap & Recycling | ₹5–28 L/mo | Tonnage / day | Weighing-scale / kanta slips | ±36% → ±13% |
What the harness is
The harness wraps the model in eight cooperating components, each with one job, a typed contract and its own failure behaviour — so a fault is contained and observable, not silently corrupting the case. It acquires, estimates, reconciles and recommends across all 56 profiles; a human decides.
🧭Planner
Chooses the next objective by expected uncertainty reduction per unit cost/time (Eq-7).
⚡Executor
Runs one planned step via the Tool Router or the capture sub-agent. Deliberately thin.
🛡️Verifier
Independently checks every output — schema, business rules, guardrails — before it enters memory.
🧠Memory
Four tiers: working state, human-confirmed facts, append-only audit log, cross-case priors.
🔀Tool Router
The only path out. Allowlist, scoped short-lived creds the model never sees, retries, logging.
✋Approval Gate
Blocks human-required steps until a named human approves. Tiered by risk × impact.
📝Logger / Tracer
One trace_id per case, span_id per step — every prompt, tool call and decision.
🔗X-Ray Integrator
Typed adapter to the Business X-Ray engine: pulls diagnostics per profile, proposes new signals.
Governing objective. Minimise credit risk while keeping TAT under 48 hours, with full regulator-grade auditability and human-only sanction. Every stage below is justified against it — for every industry profile.
The 10-stage workflow
A short, mostly-fixed plan made adaptive: the loop chooses whether to acquire more evidence, ask, pull an external signal or stop — by what most reduces decision-relevant uncertainty, given the TAT budget. The plan is identical across profiles; the profile decides which evidence and floors bind.
1
Define the objective / task
State the decision-relevant goal and success criteria for this run.
2
Configure the harness environment
Ephemeral sandbox, egress allowlist, scoped credentials, audit chain.
3
Select the model / agent under evaluation
Pick the system-under-test and its role, model tier and permissions.
4
Prepare input, prompt, tools & test data
Assemble the case: captured evidence, prompt, tool allowlist, test fixture.
5
Execute the harness
Kick off the planner → executor → verifier loop against the case.
6
Monitor execution in real time
Stream planner decisions, tool calls and verifier verdicts as they happen.
7
Capture output & intermediate actions
Record the evidence graph, tool I/O and the final scenario produced.
8
Evaluate output against criteria
Score against the gates: interval width ≤±20%, IRI floor, concordance.
9
Generate feedback, errors & recommendations
Produce scores, flag issues and propose concrete improvements.
10
Store results for comparison & reporting
Persist the run to the registry for A/B, dashboards and future iterations.
Harness configuration
The environment, permissions and gates this run executes under. Read-only sample; production values come from the versioned policy registry at decision time.
| Profile suite | 56 industry profiles across 7 archetypes — each with its own drivers, evidence ladder & floors |
| Execution context | Ephemeral sandbox container · one per case · destroyed at close |
| Isolation | Network egress allowlist · Router-issued scoped, short-lived credentials · tenant isolation |
| System under test | X-Ray Underwriting Agent · role underwriting.first_pass |
| Model | claude-opus-4-8 · reasoning: high |
| Tool allowlist | xray.assess · gstn.get_filings · udyam.lookup · aa.fetch_inflows · bureau.pull · los.attach_package (+5) |
| Stop / quality gates | interval half-width ≤ ±20% · IRI ≥ profile floor · concordance C · no unresolved conflict |
| Approval tiers | T0 auto · T1 review · T2 maker-checker · T3 committee (by risk × impact) |
| Governing objective | Minimise credit risk · TAT < 48h · full auditability · human-only sanction |
| Posture | Sandbox / mock external pulls · redacted · no funds movement · abstention is valid |
Test cases
A representative, blinded slice of the 56-profile fixture set, with expected criteria and the latest actual result. The full suite runs one fixture per profile.
| Fixture | Profile | Expected | Actual | Result |
|---|---|---|---|---|
| P24-PHA-024 | Pharmacy | ±20% width, IRI≥0.75 | ±18%, IRI 0.71 | Warn |
| P02-RES-002 | Restaurant & Fast Food | ±20% width | ±12% | Pass |
| P28-SCR-028 | Scrap Dealer Shop | cash-share flagged | flagged ✓ | Pass |
| P56-WHO-056 | Wholesale Business C&F | concordance ≥0.8 | 0.86 | Pass |
| P47-VEH-047 | Vehicle Service | tool-use ≥90% | 95% | Pass |
| P12-GAR-012 | Garment Manufacturing – Job-work | ±20% width | ±14% | Pass |
| P51-GOO-051 | Goods Transport Operator | utilisation observed | FASTag ✓ | Pass |
| P18-HAR-018 | Hardware, Paint & Sanitaryware Store | consent valid | EXPIRED → escalate | Fail |
Feedback & logs
Append-only, timestamped trace of the most recent run. Colour-coded by severity. Run the Demo to populate this live.
CompletedIn progressWarningFailed / criticalNot started
| Time | Level | Stage | Event |
|---|---|---|---|
| No events yet — start the Demo. | |||
Reports
Profiles wired56 / 56
Pass rate88%
Avg quality89%
Median TAT2.6 s
Output quality — last 5 runs
Recent runs
| Run | Profile | Result | Quality |
|---|---|---|---|
| #241 | Pharmacy | warning | 86% |
| #240 | Restaurant & Fast Food | success | 93% |
| #239 | Scrap Dealer Shop | success | 90% |
| #238 | Hardware, Paint & Sanitaryware Store | failure | — |
| #237 | Wholesale Business C&F | success | 91% |
Illustrative sample data for demonstration. Production reports draw from the run registry (per-run trace, scores, gate outcomes) for A/B tests, drift and fairness slices — sliced by industry profile and archetype.