Platform

One line. Three steps.
One receipt per change.

Haltere is the line your agents’ work runs on. A change cannot reach production without passing every step, and the record is written in the same commit.

The line

Build it. See it. Control it.

01 · The map

The system telling the truth about itself

Services, operations, permissions, contracts, with live traffic and evidence on the same screen. Drift fails the build, so agents read one record and act.

The Haltere live overlay on the system map — edge pipeline, entrypoints, commands, handlers, events and tables, with live traffic and rejections drawn on every pipe
02 · The gate

Enforcement, not alerting

Checks run before state changes and approvals survive restarts. An unauthorized change doesn’t get a warning comment. It doesn’t merge. Nothing enters the line anonymously: every action carries the named requester it acts on behalf of, and the grant that authorised it.

A Haltere gate blocking a change — a cross-domain write is BLOCKED, shown with its evidence and a re-scope prompt
04 · The receipt

Prove it.

Two receipts, one shape: how the software was built, and what it decides once it runs. Click any line.

haltere · operation receiptrecording
request
CHG-2417 · workers’-comp rating change · Meridian Casualty MGA
“update WC loss-cost multipliers per carrier bulletin” · requested by program manager, role rating:request · scope granted: rating.client-config.write, nothing else — what the order doesn’t grant, the agent can’t touch
one query: the requester, the authority — intent to production, unbroken
retrieved
program WC-MGA-2024 · config v7 · filed rate FR-TX-2024-118
context pinned before drafting: config v7 @ e41c9b · filed rate FR-TX-2024-118 with its approved band 1.38–1.61 · the last 6 changes to this program, with their receipts · every reference frozen — the model worked from reality, not a guess
one query: exactly what the AI saw, at the versions it saw them
drafted
claude-opus · prompt rating_change_v4 @a3f9c2
model claude-opus · prompt rating_change_v4 @a3f9c2 — the exact rendered prompt is on the record
one query: the instruction, not a recollection of it
checked
✓ stay-in-scope  ✓ core-boundary  ✓ filed-rate-conformance
✓ stay-in-scope — touched only rating_rules; nothing outside the work order’s grant
✓ core-boundary — no writes across the domain boundary
✓ filed-rate-conformance — 1.55 sits inside the filed band 1.38–1.61; a value outside it fails here, and no human is ever asked
enforced in CI, not asserted in a PDF — each check names the policy revision it ran against
approved
Dana Reyes · underwriting operations lead, Meridian Casualty · with the why, on the record
Dana Reyes · underwriting operations lead, Meridian Casualty · Jul 24, 9:14 AM CT · the why, verbatim: “loss-cost multiplier +9%, within the filed band; hazard group C unchanged” · approval is durable — it survives restarts, redeploys, and years
one query: the human name the examiner asks for, with the reasoning attached
recorded
wrote rating_rules v7 → v8 · emitted the event · wrote the receipt in the same commit
rating_rules v7 → v8, diff on the record · event emitted downstream · evidence pack drawn: security-review exhibit · DOI exam evidence · AI-controls matrix — every exhibit drawn from receipts the system already emits
downstream of here, nothing is generated. Only read.
✓ on the record · one query reconstructs it all

Click any line.

Fixture data; fields the spec marks proposed are shown as they will appear.

The format is public · The sample evidence pack (PDF)

An orchestration layer records its own view of a call, and the trail stops at the boundary of whatever it calls. Haltere is what the systems are built out of, so the record covers the whole transaction you build on it, not one layer’s view of it.

The line today

What ships.

On the line today

Built, and hard to fake.

  • Denials recorded with the rule that matched. Not a warning comment: no merge, no approval requested, and a row that names the rule. See it refuse
  • The map is derived from code, and drift fails the build.
  • Rails and conformance suites, enforced in CI.
  • One door for every action.
  • The append-only record, enforced at the database by role.
  • The same keys on every row.
  • Agents are principals: on behalf of a named requester, with a grant id.
  • The exact rendered prompt on the record. The format, and the schema
  • Durable approvals that survive restarts, redeploys, years.
  • Dry run with blast radius: what a change would touch, and how far, previewed before it is allowed to happen.
The limits, stated here

Where the record stops.

  • The record is append-only, enforced below the application: the database refuses UPDATE and DELETE, and the role that serves the application cannot lift them. The chain that would make the sequence self-checking is not built, and we are not going to call this tamper-evident.
  • Seven capabilities on this line are not built. They are listed with their status, dated, on the capability-status page.

What is not in the record: systems that predate the line, and they cannot be. A rule can only be enforced inside software built to run it; a layer outside it records that a call was made, not what the call did. So enforcement starts at what you build next, and existing systems enter the record as they move onto it. We would rather you read that here than find it on exam.

Security & trust

Inside your walls.

Your boundary

It runs where your risk sits.

Self-hosted or private cloud, air-gap on request. Nothing you run trains us, or anyone.

Your identity

Your permissions, on the record.

Your identity provider and authorization configuration plug in from day one. Least privilege by default.

Your models

The port is the contract.

Anthropic ships today; other providers and local models drop in behind the same port.

The questionnaire

We answer it in writing, with the unbuilt items named, before the engagement begins.

Architecture, data flows, subprocessors, deployment boundary, in writing.

Questions

Asked on every first call.

Does Haltere replace our agents or our tools?

No. Keep Claude Code, Cursor, Codex, GitHub, and your cloud. Haltere is the line their work runs on: the map they read, the gate they can’t bypass, the receipt they leave.

Where does it run?

Inside your boundary: self-hosted or your private cloud, air-gap on request. Your identity provider and authorization configuration plug in from day one. Nothing you run trains us or anyone.

What happens to code an agent wrote?

It carries the model, the exact rendered prompt at its SHA, the frozen commit, the policy checks that ran, and the human who approved it. A change that fails policy never merges.

What ships today, and what is next?

Shipped: the derived map, the rails that fail the build, one door, the append-only record, agents as principals, durable approvals, denials recorded with the rule that matched, and dry run with blast radius. Every capability is listed with its status and its basis, dated, on the capability-status page.

Do we own what gets built?

Yes. The software, the map, and the receipts are yours, in open formats, in your environment. Walk at renewal with everything.

Start with the workflow you already have to build.

Thirty minutes, with the engineers who build the line. You name the workflow and the examiner you answer to; we tell you what a governed build of it looks like, what it costs, and when it can start. Bring your compliance lead.

Inquiries are answered within one business day, by someone who can answer them.