Field notes
Operating notes, reproducible comparisons, and migration decisions. Claims are dated; unknowns stay visible. Each note is grouped by the kind of evidence it is, so you know what it can and cannot prove before you open it.
An architectural comparison of TypeSafe AI's Jev System One model against LLM structured JSON output: latency, cost per 100k decisions, and strict failure boundaries.
Read the field note →Repeated or blinded runs on a fixed fixture with a verifier. They show what happened on that task, on that date. They do not name a universal winner.
You leave with The audit failures, the exact packaging check, two negative-test results, and a free sampler vault to rerun them
The audit failures, exact packaging check, two negative tests, and the limits that remain before a paid pack can open.
Read the field note →You leave with Median time, token, and code-size deltas across twelve runs, with the fixture and verifier linked, and a better add-on test to copy
Twelve passing runs produced different time, token, and code-size tradeoffs—evidence that skills must be tested per harness and job.
Read the field note →You leave with Raw runs for both arms, the fixture and verifier, and per-metric medians for time, tokens, commands, and implementation lines
Both arms passed; Ponytail's median implementation was smaller, while its median run took longer and used more tokens on the same fixed task.
Open the comparison →You leave with Raw runs for both arms, the fixture and verifier, and per-metric medians for time, turns, tokens, list cost, and implementation lines
A blinded three-run-per-arm comparison with the exact fixture, verifier, implementation artifacts, time, turns, tokens, and reported list-cost estimates.
Open the comparison →You leave with A cross-harness baseline with elapsed time, cost boundaries, verifier results, one failure, and one blocked client, plus the pattern to reuse
Real elapsed time, cost boundaries, and verifier behavior across Codex, Claude Code, OpenCode with local Qwen3 Coder and Devstral, and Gemini CLI.
Read the field note →Plans, worksheets, and verifiers assembled from primary sources. They tell you what to check and in what order. They are not evidence that a migration has been run.
You leave with A 3-year TCO ledger comparing SaaS vs self-hosted across infrastructure, maintenance labor, upgrade risk, and exit cost, plus a downloadable calculation worksheet
Compare self-hosted open-source software against SaaS: price the invisible operator labor, backup restoration drills, security patching, and exit costs.
Read the field note →You leave with A free CSV with five fictional client records, a daily follow-up routine, and a clear spreadsheet-to-CRM decision.
Download a simple client follow-up spreadsheet for Excel or Google Sheets, with five fictional examples, a daily routine, and clear signs you need a CRM.
Read the field note →You leave with A five-agent price check from current vendor pages, the AGPLv3 self-hosted path, an administrator ledger, and a six-step export test
Current first-party pricing for five agents, the AGPL and MIT self-hosted options, the administrator work nobody quotes, and what each export leaves behind.
Read the field note →You leave with A twenty-seat price check, the seat minimum that breaks the comparison, export findings per product, and a conditional decision path
A twenty-person scenario priced from current vendor pages, the GPL and AGPL self-hosted path, and what each export omits.
Read the field note →You leave with Six versioned, pullable kits, each with a starter ledger, a shell verifier, and a stated evidence boundary
Versioned kits for specifications, agent evaluation, document parsing, browser performance, and dependency upgrades—with explicit evidence boundaries.
Read the field note →You leave with A migration ledger: authority record, source inventory, conversion and access matrix, pilot, reconciliation, cutover, and rollback checks
An export-authority, conversion-fidelity, access-reconstruction, operating-ownership, backup, reconciliation, and rollback plan.
Read the field note →You leave with A migration ledger with the license boundary, source-scope inventory, destination pilot matrix, hosting gate, and rollback checks
A license, source-scope, import-fidelity, operating-ownership, reconciliation, and rollback plan for a real team workspace.
Read the field note →You leave with A secret-free migration kit: access-design worksheet, synthetic pilot, reconciliation table, cleanup checklist, and a runnable verifier
A safe team plan for access design, a one-time pilot, plaintext-export controls, reconciliation, cleanup, and rollback.
Read the field note →You leave with A billing-unit comparison plus a kit with the workflow inventory, failure drills, clean-instance exit test, and one verifier
Current billing units, license boundaries, self-hosting ownership, and a failure-and-reconstruction test for one real workflow.
Read the field note →You leave with A ten-seat base-price check, an exit test to run before signing, and a downloadable administrator-and-exit cost kit
Current official pricing, operating responsibility, migration proof, and an open-source path—without declaring a universal winner.
Read the field note →You leave with A fixed 25-person cost scenario, a migration checklist, an editable ownership worksheet, and a dated list of unknowns
A fixed 25-person cost scenario, migration checklist, editable worksheet, and honest list of unknowns.
Open the decision kit →Steps to run before you buy or connect something. They catch known failure modes. Passing one is not a guarantee of safety or fit.
You leave with A downloadable ten-check scorecard, a five-client test fixture, and a decision rule that keeps unknowns visible.
Use a free ten-check CRM trial scorecard to test follow-ups, permissions, handoffs, exports, and plan limits with five fictional clients.
Read the field note →You leave with The five requirements of a defensible add-on test, both measured results side by side, and the protocol to rerun on your own repository
Two blinded three-run-per-arm comparisons of the same add-on reached opposite conclusions. What a defensible add-on test requires, and how to run it yourself.
Read the field note →You leave with A six-step preflight worksheet with a provenance record, permission ledger, negative-test table, and an approve, pilot, or reject gate
A source-led preflight for provenance, dependency signals, least privilege, disposable execution, negative tests, and revocation.
Read the field note →You leave with A five-step exit test (export, open, restore, operate, reconcile) and export findings for five popular products
What five popular products actually export—and a five-step test for distinguishing an archive from a working migration path.
Read the field note →Source reviews and anonymized operating notes. They report what we read or ran in one environment. Nothing here was re-measured on a shared fixture.
You leave with A source review plus two pullable workflows: a three-role code-change gate and a controlled one-agent-versus-multi-agent experiment
What the 264-role repository gets right, where orchestration claims outrun evidence, and two smaller workflows with explicit checks and stop conditions.
Read the field note →You leave with A source review of five credibility decisions, what the repository proves and does not, and the field test still to run
What Blake McCarn's Paperless knowledge graph gets right about evidence, freshness, retrieval modes, and fail-closed answers—and what still needs a reproducible test.
Read the field note →You leave with Four durable findings from two real environments and an ownership worksheet for pricing the operator, not just the license
What two real operating environments taught us about control, maintenance, CI runners, review, and the cost hidden outside the license.
Read the field note →