Mode · In-VPC
Models, Policy Engine and ledger run in your cloud tenancy. Inference makes no external calls.
GroundSetThe platform
One engine. Four layers. One record.
Models read messy reality. Your rules decide. Every outcome routes with its justification and is set in a ledger that lives inside your perimeter. The four verbs from the homepage — Understand, Decide, Route, Set — run on the four layers below.
01Architecture
Perception → Context → Judgement → Execution.
- Perception01
Reads the evidence.
Domain-trained SLMs and VLMs read documents, images, voice and video — handwritten ledgers to camera feeds.
in scans · IDs · voice · video out facts + confidence
- Context02
Grounds the facts.
Joins what was read to the case it belongs to: the account, the policy terms, the site, the prior decisions.
in facts out grounded case
- Judgement03Policy Engine
Applies your rules.
The Policy Engine. Your policies, thresholds and SOPs, encoded as versioned rules and evaluated deterministically.
in grounded case out outcome + justification
- Execution04
Acts, with reasons.
Approves, rejects, holds or escalates to a named human role, and writes back to your core systems and queues.
in outcome out route + action
Decision ledgerevery layer writes here · immutable · replayable
The Policy Engine is the moat.
Models read; rules decide. Your policies are encoded, versioned and owned by you, and every decision is recorded against the exact version that made it.
Change a threshold and the ledger shows every decision on either side of the change. That is what makes judgement reviewable — by your risk committee, your auditors and your regulator.
02The ledger
Anatomy of one decision. This is the product.
- record_id
- dl_2026_09_28_000417
- case
- MSME-LN-88213
- evidence[0]
- handwritten_ledger_0447.jpg · sha256:9c1e…a04b
- evidence[1]
- id_scan.png · sha256:41f7…0c2d
- facts
- monthly_revenue ₱412,300 (0.987) · debtors 14 (0.962)
- rule
- MSME-CREDIT-114.v12 · dscr ≥ 1.25 · PASS
- outcome
- APPROVE · limit ₱180,000
- route
- auto
- confidence
- 0.981
- override
- none
- latency
- 47ms
- prev_hash
- 0x77d0…19be
- hash
- 0x8f3a…c41d
- 01Evidence
- Every input, fingerprinted. The document itself is purged after processing; its hash stays, so the record proves what was seen without keeping it.
- 02Rule version
- The exact rule and version that decided. Rules are versioned like code — nothing decides against an unrecorded rule.
- 03Outcome
- The decision and its route, with the justification that travels to whoever acts next.
- 04Confidence
- Per-fact extraction confidence. Below your threshold, the case routes to a person instead of guessing.
- 05Override
- When a person changes the outcome, who, when and why is set in the same record. Judgement stays attributable.
- 06Hash
- Each record is chained to the one before it. Alter one and the chain breaks. Any decision replays end-to-end.
03Deployment
Inside your perimeter. In three weeks.
GroundSet deploys in-VPC or fully air-gapped. The first production workflow is live three weeks from kickoff. Here is what your InfoSec team will review, and what they will find.
Mode · Air-gapped
On your hardware, with no network path out. Updates enter through your own change process.
What your InfoSec team reviews
01Network
No egress at inference. The only inbound path is the one you open.
02Data handling
In-memory processing, purge on completion, zero retention.
03Access
Who can read the ledger, change rules and override decisions — enforced inside your environment.
04Model provenance
Open-weight base models, post-trained in-house. Composition disclosed under NDA.
05Testing
VAPT results and the security annex, shared under NDA.
04Model doctrine
Open-weight base models, post-trained in-house, running entirely inside your perimeter. Composition disclosed under NDA.
05Objections, answered
Plain answers to fair questions.
- 01Isn’t this an LLM with a prompt?
- No. Models read; they do not decide. Extraction produces facts with confidence scores, and your rules — not a model — produce the outcome. Same input, same output, every time. There is no temperature in the decision path.
- 02Why not a copilot?
- A copilot suggests and a person decides, so the record is still whatever that person remembers. GroundSet decides within rules you set, routes exceptions to people with the reasoning attached, and records both. Accountability lives in the ledger, not in an inbox.
- 03Why does the record compound?
- Every decision is set against a rule version with its evidence. Over time the ledger becomes the institution’s memory of judgement: which rules held, where people overrode, what changed when a threshold moved. It grows with every decision, and it is yours.
- 04What happens when the evidence is unclear?
- Confidence below your threshold routes the case to a person, with the evidence and the partial reading attached. The engine never guesses its way into an approval.
06Next step
See one real workflow become infrastructure.
We map one decision workflow from your operation and show it running end-to-end — understood, decided, routed, and set in the ledger. Thirty minutes, your data patterns, no slideware.