Contact

Principles

Enscrive is built around a small number of commitments that are older than any feature and outlast any release. When we have to choose between making something easy for us and making something trustworthy for you, the principles below tell us which way to resolve the conflict.

These aren’t aspirations. They’re ratified as a founding document in our governance repository and bind every design decision going forward. If any part of Enscrive doesn’t currently meet them, our job is to close the gap — not to lower the bar.

1. API-first, always

The public /v1/* API is the complete authoritative surface for every capability Enscrive provides. The web portal and the CLI are peer clients of that API. Neither depends on the other. Every capability you can reach by clicking in the portal, you can script through the CLI. Every capability you can script through the CLI, a human on your team can reach through the portal.

What this means for you.

  • Your automation will never lag our UI. When a new capability ships, it’s available to agents the same hour it’s available to humans.
  • A workflow you built through the portal can be handed to a teammate who uses only the CLI, with zero translation required.
  • Our own tests run against the same API you do. We don’t get shortcuts we haven’t given you.
  • AI agents driving your Enscrive workflow have the same surface as your developers, without special “agent tools” or hidden endpoints.

What it disallows. Enscrive will never ship a feature available only in the portal. We will never hold back a CLI command to “keep the UI differentiated.” We will never design functionality that requires a human browser session to access.

2. A Voice is one coherent thing

Most vector databases let you pick an embedding model, pick a chunker, and pick a query processor — three independent knobs with no enforced coherence. Enscrive takes the opposite position: a Voice owns chunking, embedding, query processing, and ranking as a single semantic identity. The voice that reads your corpus is the voice that answers your questions.

This is the product thesis. It’s why “voice” is a meaningful noun for Enscrive customers and not marketing dressing on a pipeline config. A legal-research voice thinks about legal documents coherently end-to-end, rather than chunking legal text one way and querying it with an unrelated ranking strategy.

What this means for you.

  • Voice versions are meaningful milestones. “My voice v7, tuned on my eval dataset” is a real artifact that represents a deliberate iteration history.
  • Voice changes have honest costs. We’ll tell you before you commit whether a change requires re-embedding your corpus and what that will cost. You can experiment on query-side behavior for free and commit to chunking changes when you’re ready.
  • Your voice’s performance is reproducible. Given the same voice, same corpus, and same query, you get the same result.

What it disallows. We won’t split voice into fragmented sub-concepts to save compute. We won’t silently auto-apply a chunking change without cost confirmation.

3. We bill on what crosses your corpus boundary

Enscrive bills on three deterministic dimensions: vectors written, vector comparisons read, and bytes stored. Embedding and chunking tokens from external providers (OpenAI, Nebius, Voyage) are billed as published passthrough with a transparent markup on a two-line invoice.

We do not bill on compute-time, network throughput, or anything else you can’t directly observe. Those are recorded for platform stability — we monitor them, alert on them, capacity-plan around them — but they never appear on your invoice.

What this means for you.

  • Predictable pricing. If your workload produces 10,000 writes and 50,000 reads this month, you know what that costs. No surprises from our infrastructure jitter.
  • Transparent markup. When you see Embedding (OpenAI): $10.00 / Enscrive markup: $2.00 on your invoice, you know exactly what went to OpenAI and what went to us. No hidden layers.
  • Comparable to alternatives. Pinecone, Turbopuffer, Weaviate, MongoDB Atlas, and Qdrant Cloud bill on similar dimensions. Your cost estimates transfer across platforms; ours are not mystery-priced.

What it disallows. We will not bundle unrelated costs into a “compute unit” we define. We will not charge for internal data transfer you cannot see. We will not adjust markup without published notice.

4. Numbers you can verify

Every number Enscrive publishes — benchmark scores, metered quantities, wallet debits — is produced by deterministic closed-form computation against reproducible inputs. You can recompute any benchmark score from our published scorer formula and our published qrels and get the same answer we got. You can recompute any wallet debit from our metering event log and our published pricing rates.

What this means for you.

  • Benchmark results from Enscrive are comparable to any other benchmark computed the same way. No fuzzy model-as-judge scores masquerading as objective measurement.
  • If you dispute a metered quantity, we can walk you through exactly how it was derived, event-by-event, against authoritative upstream references.
  • Your monthly invoice reconciles to the penny against our operational metering events, which reconcile to the penny against the relevant provider invoices we received.

What it disallows. We will not publish benchmark numbers generated by a language model’s subjective scoring. We will not use “approximately” in any billing calculation where exact computation is possible. Where LLM-judged scoring is offered as an advanced feature for qualitative evaluation on data without ground truth, it is separately priced and unambiguously labeled.

5. Your agent does the reasoning

Enscrive is deterministic infrastructure. When you need an LLM to reason — to propose voice improvements, to interpret eval results, to write custom recipes — that reasoning runs in your agent’s LLM context, against your agent’s LLM subscription, using your agent’s prompt engineering. We surface structured primitives through the CLI (enscrive evals diagnose, enscrive datasets describe, enscrive voices diff-cost) that give your agent everything it needs to reason well.

What this means for you.

  • No double-billing. You already have an LLM subscription; we don’t re-sell you tokens.
  • No lock-in to our opinions. If Claude reasons differently than GPT-4 about voice improvement, you choose which agent you prefer; we don’t pick for you.
  • Your agent integrates with Enscrive naturally. We’re a tool in your stack, not a competitor to it.
  • Our surface is auditable. If an agent makes a questionable decision, the decision was made with data you can inspect — not by a black-box on our servers.

What it disallows. We will not ship a “proprietary reasoning engine” as a paid add-on. We will not bury prompt design choices behind our API. Our CLI surface will remain complete enough that any agent can drive any workflow without special integrations.

6. Your billing ledger reconciles to the penny

Enscrive’s metering system is append-only, idempotent, and double-entry, cross-checked daily against upstream provider invoices and our own authoritative Qdrant state. Billing numbers are derived from deterministic events with authoritative source references. Historical billing records are never silently corrected; when a correction is needed, it’s a compensating entry that links to the event it corrects, and the correction is visible on your invoice.

We don’t ship customer billing until internal reconciliation has run green with zero drift for seven consecutive days. Not faster, not with “we’ll fix it after launch” exceptions.

What this means for you.

  • You can audit any invoice. Every debit links to a metered event, which links to an authoritative upstream reference — an OpenAI batch ID, a Qdrant transaction, a storage snapshot.
  • You won’t be silently reconciled. If we discover a billing error after the fact, you’ll see the correction on your next invoice, not invisibly adjusted in our database.
  • Our discipline is your insurance. The same reconciliation that catches our bugs before you notice them also catches yours — the first time your usage exceeds expectations, we’ll tell you, not surprise you at month-end.

What it disallows. Misbilling a customer is an incident, not a bug-fix. Deferring metering correctness to “ship now, fix later” is foreclosed as a decision path.

How we enforce this

These principles are binding. Every design document at enscrive-governance is reviewed against them. Every code change that conflicts with them requires ratified amendment, not silent exception.

When a customer asks “why does Enscrive do X this way?” the answer routes back to one of these six. When a customer asks “will Enscrive start doing Y?” the answer is constrained by whether Y is consistent with these.

A fuller internal version exists in our governance repository, adding the technical-implementation commitments that matter to how we build but not to what we promise you. That repository is not public today, so the six commitments above — the ones we make to you — are stated here in full rather than by reference.