Public ledger · Alpha Covenant Holdings LLC (PA 0016032097)

Where a model is used here, and where it never is

A statement of facts for the reviewer who has to inventory AI. Live figures from the call ledger; laws named with what each turns on and what we do not claim.

Position

the record	Receipts, rosters, Receipt Watch diffs, held-file reads and the Audit Handover file are produced by deterministic code that fetches a named official source and copies what it printed. No language model reads a source, decides what a source printed, builds a receipt, computes a hash, signs or timestamps anything, or decides whether two names are the same party. Every receipt body carries this statement with model_calls_during_build = 0, inside the hashed and signed body.
the guard	Every model call on this platform passes through one ledger function (aillog.record). That function refuses, and writes the refusal down, if it is reached while a receipt is being built. The ledger is public in aggregate at /api/public/ai-calls and per receipt at /api/public/ai-calls?receipt_id=.
no decisions	The platform makes no determination about any person or organization — no score, flag, rating, clearance, eligibility, compliance, credit, employment, tenant, insurance or risk conclusion — by model or otherwise. Model output here is explanation or draft, labelled as such, optional, and never an input the platform acts on.
ai off mode	Any account can switch AI off (/account/ai-mode or POST /api/account/ai-mode). In AI-off mode every AI surface returns 403 before any model is touched for requests bearing that account's session; the receipt product is unchanged because it never involved a model. Buyers who must avoid procuring an 'AI system' can take the product this way and the contract says so.

Every surface where a model is called

explain	visitor or customer · plain-English explanation of a receipt or record already fetched; the model sees only numbered facts from that record · route /api/explain calls in the last 30 days: 0 · refused inside a receipt build: 0 · models seen: none in window
record reader	visitor or customer · reading and Q&A over numbered lines lifted from a receipt or Audit Handover file; determinations are refused before the model is called · route /api/ai calls in the last 30 days: 0 · refused inside a receipt build: 0 · models seen: none in window
forms desk	visitor or customer · draft answer to a form question from the register, the eCFR text and the filer's own public filing; each sentence cited · route /api/forms-desk calls in the last 30 days: 0 · refused inside a receipt build: 0 · models seen: none in window
free guide	visitor · plain-words help and translation pointing to the official free sources on /free; never fills or submits a form · route /api/free-guide calls in the last 30 days: 0 · refused inside a receipt build: 0 · models seen: none in window
advisory desk	operator only · internal consultants briefed on a standard's public text and our own SOPs; nothing customer-facing · route /api/advisors calls in the last 30 days: 0 · refused inside a receipt build: 0 · models seen: none in window
daily replay summary	operator only · plain-English summary written FROM the daily self-check results after they are computed · route /api/audit calls in the last 30 days: 2 · refused inside a receipt build: 0 · models seen: claude-sonnet-5-5
source scout	operator only · memo per Federal Register item for the owner's decision; recommends, never publishes · route /api/audit/sentinel calls in the last 30 days: 0 · refused inside a receipt build: 0 · models seen: none in window
social caption	operator only · caption draft for our own posts; a person reviews before anything is published · route /api/social calls in the last 30 days: 0 · refused inside a receipt build: 0 · models seen: none in window

Ledger began 2026-10-25T10:05:35 · 2 calls in the last 30 days · 0 refusals inside a receipt build, all time. JSON: /api/public/ai-calls; per receipt: /api/public/ai-calls?receipt_id=…

AI use-case inventory (the shape an M-25-21 inventory asks for)

AI-001 · explain	Plain-English explanation of a receipt or record already fetched Determination: None. Output is a labelled explanation of facts already printed; the platform acts on nothing in it. Human role: A reader reads it; nothing is issued, changed or decided from it. Off switch: AI-off mode (account) / EMERGENT_LLM_KEY unset (global)
AI-002 · record_reader	Reading aloud and Q&A over numbered lines lifted from a receipt or Audit Handover file Determination: None. Questions that ask for a determination are refused before the model is called. Human role: Same as AI-001. Off switch: same
AI-003 · forms_desk	Draft answer to a form question from the register, eCFR text and the filer's own public filing Determination: None. A draft; the filer fills and signs their own form. Human role: The filer edits and submits; the draft is logged with its citations. Off switch: same
AI-004 · free_guide	Plain-words help and translation pointing to official free sources Determination: None. Never fills or submits a form. Human role: Visitor reads; nothing is filed. Off switch: same
AI-005 · advisory_desk	Operator-only internal consultants on a standard's public text and our SOPs Determination: None. Advice to the operator about our own documents. Human role: Operator decides; nothing customer-facing. Off switch: global
AI-006 · daily_replay_summary	Summary written FROM the daily self-check results after they are computed Determination: None. The checks are computed first; the model narrates them. Human role: Operator reads. Off switch: global
AI-007 · source_scout	Memo per Federal Register item for the owner's source decision Determination: None. Recommends a source to read; the owner adds or declines. Human role: Owner decides. Off switch: global
AI-008 · social_caption	Caption drafts for our own posts Determination: None. Human role: A person reviews before anything is published. Off switch: global

Proposed to us and declined

Reviewers have proposed these; each would make the platform a decision engine. Listed so the line is visible.

Candidate match triage ranked by a deterministic fuzzy score	RULES.md forbids fuzzy, probabilistic or similarity matching and any score. Receipts match on exact key or exact printed name only; two rows under the same name are never asserted to be the same party. A ranked candidate list is an identity determination with a number on it.
Draft narrative on whether a candidate match is the same entity	That is an identity determination, which no surface here makes. The record reader refuses determination questions before the model is called.
Recommendation: clear / escalate / request_more_info	'Clear' is on the forbidden-words list of the Receipt Format, and a recommendation is a decision input. The platform issues receipts, not recommendations; decisions belong to the customer's named officer, who records them on a decision record we countersign for hash and time only.
Anomaly flagging in transaction clusters	'Anomaly' and 'red flag' are forbidden words in the Receipt Format; a flag is a verdict. We print what a source printed and when.
Tracking divergence between AI recommendations and officer decisions	There are no recommendations to diverge from. The measurable things are published: calls per surface, model per call, prompt and output hashes, refusals inside a receipt build (zero).

NIST AI RMF — the control behind each function

GOVERN	RULES.md in the repository and the AI & Automation Policy (/legal/ai-usage) state what no model may touch; this inventory names a surface, a route and an owner for every model use; the owner is the operator of record.
MAP	Eight surfaces, each tied to one route and one purpose (table above). None sits in the receipt build path: aillog.ReceiptBuildGuard makes a model call inside a build a logged, raised error.
MEASURE	Every call is one ledger row: surface, provider, model, time, subject, prompt SHA-256, output SHA-256 and length, or the error class. Aggregates are public (/api/public/ai-calls); the texts live in the surface's own log. No recommendation-vs-decision metric exists because no recommendation exists.
MANAGE	AI-off mode per account (403 before any model call); global off by removing the model key; the deterministic record path and every issued receipt are unaffected either way.

OMB M-25-21 topics — our status, in our words

AI use case inventory	Published on this page (AI-001…AI-008) with route, purpose and off switch; JSON at /api/public/ai-position for a customer's own documentation.
High-impact / consequential-decision determination	No surface makes or materially shapes any decision about a person or organization; the platform makes none. Reasoning: each row's 'determination' column. We do not grade this as a risk level — there is no scored risk here.
Risk management (NIST AI RMF)	Controls per function listed above; no certification claimed.
Human oversight	Receipts need no signature from anyone — they are deterministic reads, and that is the point: there is no AI step for a human to oversee in the record path. Where a customer records a decision, the customer's named officer types and signs it; we countersign hash and time.
Audit logging	Model identifier, prompt hash, output hash and length, time, subject and receipt linkage (calls ABOUT a receipt) are retained in the ledger; prompt and output texts in the surface logs. No model output contributes to a receipt, so 'officer decision vs. AI recommendation' is not a field.
AI can be disabled	Yes — per account (AI-off mode) and globally. The deterministic path is untouched.

The instruments a reviewer will ask about

EU AI Act, Regulation (EU) 2024/1689 — Article 50 (transparency) and Annex III (high-risk)	Turns on: Art. 50: a person interacting with an AI system, or reading AI-generated text, must be told. Annex III: systems that make or materially shape decisions on credit, employment, essential services, law enforcement and similar. Our facts: Every model output is labelled 'explanation', 'reading' or 'draft' with the provider and model named. No system here makes or shapes an Annex III decision; the Terms and Acceptable Use Policy prohibit using anything here for such decisions about individuals. We do not claim: That we have conducted an EU conformity assessment, or that a customer's own use is outside the Act — the customer's use is the customer's. official page
Colorado Artificial Intelligence Act (SB 24-205, as amended by later session bills)	Turns on: Developers and deployers of 'high-risk' AI systems — systems that make or are a substantial factor in a consequential decision about a consumer (education, employment, financial or lending, essential government services, health care, housing, insurance, legal services). Our facts: The receipt is not an AI system. The AI surfaces here explain or draft text and are not a factor in any consequential decision the platform makes, because the platform makes none; the Acceptable Use Policy bars using the service for such decisions about a consumer. We do not claim: An exemption. We state facts; whether a customer's workflow is a 'deployment' is the customer's determination. The compliance date has moved more than once — read the Attorney General's page for the current one. official page
Texas Responsible Artificial Intelligence Governance Act (HB 149, TRAIGA), effective 2026-01-01	Turns on: Disclosure when a government agency's AI interacts with a consumer; prohibited uses (manipulation, social scoring, certain biometric identification, discrimination with intent). Our facts: Every AI output is labelled as model output at the point it appears. None of the prohibited uses is present: no scoring of persons, no biometric identification, no profiling. We do not claim: To be a government agency or to speak for one; a Texas agency that buys the product with AI surfaces on must make its own disclosure, or buy it in AI-off mode. official page
California AI Transparency Act (SB 942, as amended by AB 853) and AB 2013 (training-data transparency)	Turns on: SB 942: 'covered providers' of generative AI systems with over one million monthly users must offer detection tools and latent disclosures. AB 2013: developers of generative AI systems must publish training-data documentation. Our facts: We do not develop or train a generative AI system and do not have one million monthly users; we call third-party models (OpenAI, Anthropic, Google) and label the output. We do not claim: That the model providers' own obligations are ours, or that our labels are the providers' latent disclosures. official page
Utah Artificial Intelligence Policy Act (SB 149, 2024, as amended)	Turns on: Disclosure that a person is interacting with generative AI when asked, and always for regulated occupations. Our facts: Disclosure is not on request here; every AI output carries the label and model name every time. We do not claim: That the labels satisfy any regulated-occupation duty of a customer who is a licensed professional. official page
OMB Memoranda M-25-21 (agency use of AI) and M-25-22 (acquisition of AI) — bind federal agencies	Turns on: Agencies must inventory AI use cases, manage 'high-impact' AI, and apply acquisition terms when they buy AI. Our facts: The receipt product contains no AI and is available with every AI surface off (AI-off mode), so an agency can procure it as software. If an agency uses the AI surfaces, the facts an inventory needs — surfaces, providers, models, logging, what the model sees — are on this page and in the ledger. We do not claim: That this page satisfies any agency's AI governance plan; those duties bind the agency. official page
GDPR Article 22 (decisions based solely on automated processing)	Turns on: A decision with legal or similarly significant effect on a person made solely by automated means, including profiling. Our facts: No decision is made here about anyone; no profiling or scoring occurs. Customer decision records are typed and signed by the customer's named officer; we countersign the hash and time only. We do not claim: Anything about the customer's downstream processing. official page
NIST AI Risk Management Framework 1.0 (voluntary)	Turns on: Govern / Map / Measure / Manage functions for organizations that design, develop, deploy or use AI. Our facts: Map: eight named surfaces, listed below with who reaches them. Measure: every call in one ledger with provider and model. Manage: the receipt-build guard and AI-off mode. Govern: RULES.md and the AI & Automation Policy. We do not claim: Certification against the framework; there is none to claim. official page

What this page does not claim

This page is a statement of facts about this platform, not legal advice and not a compliance certification of any kind.
It says nothing about whether a customer's use of a receipt, an explanation or a draft meets the customer's own obligations under any of these instruments.
Laws and effective dates change; each row links the official page. Where this page and the official page differ, the official page is right and this is a corrections matter (/corrections).

Policy: https://alphacovenantholdings.com/legal/ai-usage · JSON: /api/public/ai-position · doctrine: RULES.md in the repository.