Get the app

Journal

Sourced evidence and practical notes about company records, importers, data grooming, and local-first software.

Subscribe with the Atom feed.

Evidence

47 publications

Read sourced evidence about company records and local-first software.

Product

1 publication

Notes about muniment’s local app, graph, importers, and grooming report.

Guides

26 publications

Read practices for keeping your company’s record straight, from importers to data grooming.

Announcements soon

No publications yet

Major news about muniment’s local app and paid console will appear here, including releases.

A browser window of blank controls sits under a strip of round holes, and a verdigris line runs from the one filled hole to the one filled button while a pen nib points at an empty field.

evidence ·

Browsing by multiple choice

In Muniment Desktop a decision model picks each browser step from numbered controls. The chat model writes text, confirms risky clicks and checks results.

product ·

Two decision models, two jobs

Choosing the answer model and choosing tools fail differently. Muniment Desktop gives each its own decision model, and the tool picker can pick nothing.

guides ·

Models read names

An internal ID in an MCP server key used 43 of a 64-character tool name and left the model five letters of the tool it meant.

evidence ·

Half the download was never the app

Our Linux package shipped 1.3 GB of Chromium debug symbols. Stripping, better compression, and fewer languages cut each download by at least a third.

Publications

DateCategoryTitle
evidenceBrowsing by multiple choice
productTwo decision models, two jobs
guidesModels read names
evidenceHalf the download was never the app
evidenceRouting needs more than a classifier
evidenceA control hook needs a guardian outside the agent
guidesHTML tables in a native chat renderer
guidesExtract workflow heredoc parsers into scripts
· Amended guidesLocal mode keeps the control plane out
evidenceMonitoring records stay in the customer’s account
guidesA launch smoke cannot prove a service serves
guidesA password hash is a pool outage in disguise
evidenceRegistry approval is not permission to run
evidenceThe driver states the limit beside the command
evidenceA confident answer is not a checked answer
evidenceA faster return can hide a later discovery
· Amended guidesA retry classifier read the wrong file
evidenceRequest routing needs host monitoring
guidesA security rule can hide an API contract
evidenceAgent access follows the employment record
evidenceA policy compiler checks structure, not intent
evidenceSeparate audit paths from adaptive investigation
evidenceLocal model reach is not the same as Hub use
guidesTwo green branches can merge red
· Amended evidenceA fleet of task agents needs shared controls
· Amended evidenceThe model picker is not the traffic path
· Amended evidenceScope in the task is not a control
evidenceAgent access should expire with the task
evidenceEach governance layer needs its own test
evidenceA tool request now carries its own routing record
guidesA document hash cannot localize a contract change
guidesA hand-listed lint shrinks the rule it guards
evidenceAn average score hides the governance check
guidesAn append-only table rejects an unplanned write
evidenceA filing count outranks a poll headline
evidenceAn approval holds only against the state it read
evidenceA token count does not measure a task
evidenceA cleared step does not clear the sequence
guidesAudit a written rule before you enforce it
evidenceAudit the backend from the text alone
guidesHigh coverage still hides untested guards
guidesA completion-cache hit is not a cheap model call
guidesAndroid connected is not Android ready
guidesImmutable chat history without copying content
guidesLocal durability is not model delivery
guidesResume each artifact, publish the set
guidesInstall the speech model when you need it
evidenceAn activity log is not a release control
guidesPure functions make image zoom geometry testable
guidesFail-open classifiers can fail closed downstream
guidesA unit test cannot find a contract mismatch
guidesTests without layout cannot see the fold
evidenceWhat it took to fine-tune a small router
evidenceAgent oversight loosens only when records hold
guidesAccessibility preferences are a live signal
· Amended evidenceA rented operating layer needs portable records
· Amended evidenceProduction capability rose and return did not
evidenceCloud versus local cost is a workload equation
evidenceOpen weights require an operating layer
· Amended evidenceCost receipts belong to the request
· Amended evidenceEU model provider records become enforceable
evidenceThe operating layer is becoming an owned asset
· Amended evidenceSmall models belong at the front door
· Amended evidenceAgent governance is a tiered operating model
· Amended evidenceInference routing is a systems problem
evidenceMeasure agent adoption where work ships
evidenceThe agent platform is a control plane
evidenceAgent teams move inside the system of record
· Amended evidenceRouting each reasoning step needs its own receipt
· Amended evidenceGuarantees belong in the audit harness
evidenceMobile inference needs phase-aware scheduling
guidesVirtualizing chat without losing the conversation
· Amended evidenceLocal inference is becoming a real routing tier
· Amended evidenceMeasure discovery and verification separately

Join the waitlist

Get desktop release updates.

We will email you about desktop releases and new features. muniment is a desktop workspace for your models, tools, and files.