v2.4.3

Maintainer integrity and plugin onboarding

• Enforces the documented public-layer ceilings and records the template exception.
• Restores compact front-door language and the factual reframing guardrail.
• Adds a clearer Codex-only local plugin starting path while keeping host-model usefulness explicitly unqualified.
• Repairs the former GitHub Pages base with a minimal canonical redirect stub.

Validation: full repository suite, plugin validation, installed STDIO read/write smoke, GitHub Actions validation, and CodeQL are green.

v2.4.2

Discovery evidence limits

• Adds a bounded checkpoint before a workflow redesign hypothesis: trace a normal case and material exception, separate active work from wait and rework time, test upstream and downstream constraints, and record evidence limits.
• Clarifies that timestamp gaps describe elapsed time between recorded events, not necessarily touch time or the whole workflow.
• Refreshes workflow qualification guidance and the local FDE plugin snapshot. No runtime behavior, schema, control family, connector, process-mining service, or automated-discovery capability has been added.

Verification

npm test, packaged-plugin validation, repository validation, and GitHub CodeQL all passed.

v2.4.1

Derived business-state governance

A stored CRM field, extracted relationship, score, or operational summary is now explicitly treated as a revisable claim rather than source truth. The Guide adds a candidate-versus-current review path that preserves raw evidence, prior revisions, producer configuration, time semantics, lifecycle status, and owner disposition.

Evidence-graph retrieval remains task-scoped, progressive derived context. It does not become a causal model, a company-wide context dump, or an authorization surface.

This patch also refreshes the local Applied AI Field Guide plugin snapshot and release bindings. It adds no control family, schema, hosted service, graph runtime, or production claim.

v2.4.0

Practical entry points for applied-AI work\n\n- Add a short first-workflow investigation that surfaces systems, handoffs, workarounds, and unanswered questions without claiming population-level data quality.\n- Add a recurring working adoption loop and a one-page results walkthrough that distinguishes demonstrations, estimates, and observed outcomes.\n- Add a fictional finance variance-commentary walkthrough: deterministic numbers and policy, attributable owner explanations, a reviewable model draft, and finance-controlled approval.\n- Refresh the local Applied AI Field Guide plugin snapshot and rebind the teaching release evidence to the v2.4.0 dependency lock.\n\nNo new controls, schemas, hosted service, executable finance lab, or production-performance claim.

v2.3.0 — Retrieval comparison and durable recovery

Highlights

• Adds an optional, provider-neutral embedding/hybrid comparison to the retrieval lab. The deterministic BM25 baseline remains the default and makes no network call.
• Adds a durable-recovery practice lab for persisted intent, stable idempotency, source-of-truth readback, policy drift, and escalation after an interrupted effect.
• Adds public practice routing, browser coverage, release bindings, and a synchronized local Applied AI Field Guide plugin snapshot.

Boundaries

Both exercises use public synthetic fixtures. They are teaching and evaluation practice, not deployment designs, production evidence, or provider performance claims.

Verification

added 130 packages, and audited 131 packages in 1s

10 packages are looking for funding
run npm fund for details

found 0 vulnerabilities,

applied-ai-field-guide@2.3.0 test
npm run validate && npm run test:artifact-validator && npm run test:document-review && npm run test:retrieval-evaluation && npm run test:durable-recovery && npm run test:copilot && npm run test:markdown && npm run test:prose && npm run test:paths && npm run test:repository && npm run test:contracts && npm run test:tool-security && npm run test:telemetry && npm run test:governance && npm run test:release-integrity && npm run test:release-gates && npm run test:solutions && npm run test:value-framework && npm run test:data-readiness && npm run test:field-practice && npm run test:production-readiness && npm run test:enterprise-reality && npm run test:production-ai-operations && npm run test:factory-improvement && npm run test:skills && npm run test:guide-plugin && npm run test:policy && npm run test:reference && npm run test:evals && npm run test:hybrid && npm run test:site

applied-ai-field-guide@2.3.0 validate
node scripts/validate-repository.mjs

{
"status": "passed",
"files": 363,
"markdown_files": 158,
"json_files": 82,
"schemas": 22,
"controls": 69,
"evaluation_cases": 11,
"catalog_artifacts": 142
}

applied-ai-field-guide@2.3.0 test:artifact-validator
node --test tests/artifact-validator.test.mjs

TAP version 13

Subtest: every governed JSON type supports bounded complete validation outside the clone

ok 1 - every governed JSON type supports bounded complete validation outside the clone

duration_ms: 454.705333
type: 'test'
...

Subtest: evaluation, tool and release checks retain shared semantic invariants

ok 2 - evaluation, tool and release checks retain shared semantic invariants

duration_ms: 53.844958
type: 'test'
...

Subtest: starter workflow validation requires decision-bearing fields without forking the canonical schema

ok 3 - starter workflow validation requires decision-bearing fields without forking the canonical schema

duration_ms: 24.731834
type: 'test'
...

Subtest: the documented twelve-field starter is copyable without advertising an unpublished global binary

ok 4 - the documented twelve-field starter is copyable without advertising an unpublished global binary

duration_ms: 10.721916
type: 'test'
...

Subtest: CLI help leads with commands available from a private repository clone

ok 5 - CLI help leads with commands available from a private repository clone

duration_ms: 36.152334
type: 'test'
...

Subtest: progressive validation keeps collaborative Markdown plans human-readable and structurally guarded

ok 6 - progressive validation keeps collaborative Markdown plans human-readable and structurally guarded

duration_ms: 0.831625
type: 'test'
...

Subtest: starter validation retains closed-object and type rules from the canonical schema

ok 7 - starter validation retains closed-object and type rules from the canonical schema

duration_ms: 11.0325
type: 'test'
...

Subtest: complete validation applies canonical structure and semantic invariants

ok 8 - complete validation applies canonical structure and semantic invariants

duration_ms: 20.995708
type: 'test'
...

Subtest: starter engagement validation requires two cited claims and a bounded next move

ok 9 - starter engagement validation requires two cited claims and a bounded next move

duration_ms: 11.076041
type: 'test'
...

Subtest: starter data validation accepts one representative source and names failure behavior

ok 10 - starter data validation accepts one representative source and names failure behavior

duration_ms: 14.058125
type: 'test'
...

Subtest: unknown external artifacts fail with an actionable type message

ok 11 - unknown external artifacts fail with an actionable type message

duration_ms: 0.80475
type: 'test'
...
1..11

tests 11

suites 0

pass 11

fail 0

cancelled 0

skipped 0

todo 0

duration_ms 685.290417

applied-ai-field-guide@2.3.0 test:document-review
node --test examples/invoice-exception/document-review/document-review.test.mjs

TAP version 13

Subtest: real baseline exposes format limitations and does not draft known unsafe cases

ok 1 - real baseline exposes format limitations and does not draft known unsafe cases

duration_ms: 2.831167
type: 'test'
...

Subtest: a polished wrong amount passes shape but fails the independent source grader

ok 2 - a polished wrong amount passes shape but fails the independent source grader

duration_ms: 0.30475
type: 'test'
...

Subtest: business record rejects hallucinated quotes, extra effects, fractional cents and missing fields

ok 3 - business record rejects hallucinated quotes, extra effects, fractional cents and missing fields

duration_ms: 0.083
type: 'test'
...

Subtest: review acceptance needs explicit source check and records no production authority

ok 4 - review acceptance needs explicit source check and records no production authority

duration_ms: 0.224209
type: 'test'
...

Subtest: candidate cannot emit a passing grade and candidate failure retains manual fallback

ok 5 - candidate cannot emit a passing grade and candidate failure retains manual fallback

duration_ms: 2.056084
type: 'test'
...

Subtest: live adapter fixes destination, bounds completion and rejects incomplete outputs without retry

ok 6 - live adapter fixes destination, bounds completion and rejects incomplete outputs without retry

duration_ms: 0.21025
type: 'test'
...
1..6

tests 6

suites 0

pass 6

fail 0

cancelled 0

skipped 0

todo 0

duration_ms 36.410625

applied-ai-field-guide@2.3.0 test:retrieval-evaluation
node --test examples/invoice-exception/retrieval-evaluation/retrieval-evaluation.test.mjs

TAP version 13

Subtest: BM25 output is deterministic and one vocabulary mismatch stays visible

ok 1 - BM25 output is deterministic and one vocabulary mismatch stays visible

duration_ms: 3.445666
type: 'test'
...

Subtest: tenant, scope, revision and time filters run before ranking

ok 2 - tenant, scope, revision and time filters run before ranking

duration_ms: 0.3975
type: 'test'
...

Subtest: results bind exact source revisions and never inherit instruction authority

ok 3 - results bind exact source revisions and never inherit instruction authority

duration_ms: 0.3785
type: 'test'
...

Subtest: grader ignores candidate claims and catches forged evidence

ok 4 - grader ignores candidate claims and catches forged evidence

duration_ms: 0.102666
type: 'test'
...

Subtest: grader rejects malformed and internally inconsistent candidate results

ok 5 - grader rejects malformed and internally inconsistent candidate results

duration_ms: 0.073542
type: 'test'
...

Subtest: grader binds authority metadata and rejects duplicate evidence without inflating ranking

ok 6 - grader binds authority metadata and rejects duplicate evidence without inflating ranking

duration_ms: 0.057708
type: 'test'
...

Subtest: the bundled BM25 candidate has no evaluator, network, environment or filesystem dependency

ok 7 - the bundled BM25 candidate has no evaluator, network, environment or filesystem dependency

duration_ms: 2.405875
type: 'test'
...

Subtest: an opt-in hybrid candidate embeds only admitted evidence and keeps the same independent safety grade

ok 8 - an opt-in hybrid candidate embeds only admitted evidence and keeps the same independent safety grade

duration_ms: 1.190584
type: 'test'
...

Subtest: the explicit provider adapter is unavailable without valid configuration and reports provider usage only when supplied

ok 9 - the explicit provider adapter is unavailable without valid configuration and reports provider usage only when supplied

duration_ms: 1.346417
type: 'test'
...

Subtest: unsafe candidates fail even when they add a relevant result

ok 10 - unsafe candidates fail even when they add a relevant result

duration_ms: 0.263542
type: 'test'
...

Subtest: grader independently rejects every inadmissible source even when a case does not name it

ok 11 - grader independently rejects every inadmissible source even when a case does not name it

duration_ms: 0.241959
type: 'test'
...

Subtest: no-evidence cases stay out of ranking averages

ok 12 - no-evidence cases stay out of ranking averages

duration_ms: 0.401291
type: 'test'
...

Subtest: custom candidates are not mislabeled as the bundled baseline

ok 13 - custom candidates are not mislabeled as the bundled baseline

duration_ms: 1.208125
type: 'test'
...

Subtest: candidate input is isolated from frozen evaluation fixtures

ok 14 - candidate input is isolated from frozen evaluation fixtures

duration_ms: 0.715917
type: 'test'
...
1..14

tests 14

suites 0

pass 14

fail 0

cancelled 0

skipped 0

todo 0

duration_ms 44.662792

applied-ai-field-guide@2.3.0 test:durable-recovery
node --test examples/invoice-exception/durable-recovery/durable-recovery.test.mjs

TAP version 13

Subtest: an ...

v2.2.0 — Start with the work, test the evidence

A broader front door and a retrieval lab you can break

• Start with the job: choose where AI could help, build or repair a feature, evaluate retrieval, test a material change, launch a service, or rescue a bad brief without learning repository taxonomy first.
• Test retrieval separately: run a dependency-free BM25 baseline against ten synthetic invoice-policy questions. Admission, ranking, conflict, citation, abstention, latency, and cost remain separate evidence.
• Try adversarial cases: tenant leaks, stale and superseded sources, forged authority, duplicate evidence, malformed results, invalid dates, fixture mutation, conflicts, and missing evidence fail visibly. One deliberate vocabulary miss keeps the baseline honest.
• Improve discovery: 66 canonical web pages now include dedicated human-review, context-and-retrieval, and retrieval-practice routes with corrected structured data, task-led navigation, sitemap coverage, and project-scoped crawler guidance.
• Keep the writing human: the full Markdown corpus runs through a deterministic prose audit adapted from Peter Yang’s MIT-licensed no-ai-slop rubric. It flags review targets without pretending to detect authorship.
• Refresh the local copilot: the v2.2.0 plugin carries the complete Guide snapshot and 16 skills. Its 12 local tools retain bounded search, immutable engagement artifacts, dependency-aware decisions, and fail-closed workspace controls.

Verification and honest limits

The clean npm ci --ignore-scripts, npm test, and git diff --check gate passed locally and on the PR and merged main SHA. CodeQL, workflow linting, staged-patch secret scanning, desktop/mobile Playwright checks, the official Codex plugin validator, and an installed-plugin STDIO smoke test also passed. The smoke test used a disposable workspace; the existing local engagement workspace remained unchanged.

The retrieval lab is deterministic teaching evidence over fictional data. It does not prove semantic-retrieval quality, production readiness, customer value, or search ranking. GitHub Pages crawler files under a project path cannot set host-wide crawler policy. Passing CI never supplies customer authority or release approval.

Full change: #49

v2.1.0 — Practice the work, strengthen the copilot

A guide you can practice, and a copilot that meets you where the evidence is

• Try the work: a self-contained invoice practice packet with a messy brief, difficult documents, staged evidence, a flawed submission and a review rubric. Extend the existing worked engagement; no second framework.
• Run and review: compare a rules baseline with optional live model proposals, then inspect sources, correct fields, reject or escalate in a local browser exercise. Nothing posts to a ledger or makes a payment.
• Use your existing evidence: map reviewed team records to the copilot's stages without duplicating templates. Exact revisions, stale dependencies, conflicting decisions and operating ownership still matter.
• Find and validate more: search uncataloged canonical chapters with sections, lines and digests; validate all 22 governed JSON types against fixed local schemas and shared single-artifact invariants.
• Stay in the guide: 65 canonical web pages, supporting study routes, portable Markdown/JSON downloads and recoverable search errors.
• Updated local plugin: all 16 skills, refreshed canonical snapshot, routing regressions and a separate host-model evaluation protocol.

Verification and honest limits

360 tests across 27 lanes passed, alongside repository validation, CodeQL, official plugin validation, workflow linting, staged-content secret scanning and desktop/mobile browser checks. The installed plugin was smoke-tested with fictional data; the existing local engagement workspace and configuration were preserved.

The rules baseline gets 6/8 public synthetic cases correct, with zero unsafe drafts on that set. This is teaching evidence, not a reliability claim. The optional live model attempt returned HTTP 401; no model-quality result is claimed. Human effectiveness and host-model judgment remain not evaluated. Passing CI never supplies customer authority or production approval.

Full change: #48

v2.0.0 — The Applied AI Field Guide

The project is now The Applied AI Field Guide: fieldwork, value, engineering, and operations for AI that works beyond the demo.

This release makes the existing method explicit for internal applied-AI teams, product and workflow owners, service operators, and FDEs. It keeps the field discovery and reframing work, smallest-sufficient-mechanism decisions, evidence-backed evaluations, operating ownership, and canonical production controls.

What changed

• Updated short and complete guides, lifecycle navigation, operating-model guidance, public metadata, citation, and editable social artwork.
• Renamed selected source paths and three public web routes, with compatibility aliases under the new site base.
• Renamed the personal Codex plugin to applied-ai-field-guide and conductor to $run-ai-engagement; preserved all sixteen skills and local STDIO workspace capabilities.
• Added a tested migration that retains the existing engagement workspace and legacy configuration. Invalid or overlapping paths fail before replacement.
• Refreshed the scorecard PDF and bundled it in the offline plugin snapshot.
• Fixed a lifecycle diagram node collision, made reframing conditional on material contradiction, and corrected mobile overflow and search Escape behavior.

Migration

Repository: https://github.com/davidahmann/applied-ai-field-guide

Web guide: https://davidahmann.github.io/applied-ai-field-guide/

GitHub redirects the old repository/Git URLs, but not the old GitHub Pages base. Update web bookmarks. Schema URI/title annotations changed; validation semantics and independent schema versions did not. Existing control/artifact identifiers and immutable historical records remain intact.

Read the 2.0 migration note before updating an installed plugin or URI-pinned consumer.

Verification and scope

Complete repository tests, deterministic reference evaluations, all sixteen skill metadata checks, plugin/package parity and migration tests, workflow linting, and a redacted secret scan passed locally. Independently reviewed content, plugin migration, and site behavior; exercised all 24 canonical pages at desktop and mobile widths and visually inspected the regenerated exports.

The release adds no method, lifecycle, control family, or net new file in guide/library/templates/blueprints. Synthetic tests remain teaching evidence, not production certification. This release also includes the previously untagged 1.34 evaluation-qualification and proof-loop changes recorded in the changelog.

v1.33.0 — Discovery and evaluation practice

What changed

• Added a two-aperture discovery loop that reconciles strategic intent with observed operating and technical evidence.
• Added explicit candidate value surfaces without making customer visibility a universal priority.
• Added a protected long-horizon optimization pattern with durable hypotheses, results, insights, revisions, and a separately controlled promotion evaluator.
• Added pinned Arbor attribution and bounded practitioner evidence notes.
• Updated fast-uri from 3.1.5 to 3.1.7 in both lockfiles, rebound release evidence, and synchronized the complete FDE plugin at v1.33.0.

Validation

The release passed the clean-install repository suite, contract and release-integrity tests, site generation, plugin tests, official plugin validation, CodeQL, dependency audit, workflow lint, spelling review, and patch secret scan.

v1.32.0

Company agent operating model

• Centralizes reusable identity, evaluation, telemetry, routing, security, cost, and delivery rails while preserving workflow-local authority.
• Adds explicit company/program, workflow, metric/verifier, service, operational, platform, risk, delivery, and operator decision rights.
• Adds a fail-closed proof-to-operation gate with separate technical, operator, adoption, value, economics, risk, and production-readiness evidence.
• Makes autonomy earned per segment and effect class, with receiving-team capability required before bounded production.
• Uses the production-system layer view as an end-to-end completeness check without adding another framework or fixed topology.
• Updates the FDE local plugin and canonical operating/system-design skills to version 1.32.0.

See the changelog for the complete release record.