Skip to content

SPECIFICATION.md

Latest commit

 

History

History
593 lines (432 loc) · 28.2 KB

SPECIFICATION.md

File metadata and controls

593 lines (432 loc) · 28.2 KB
Raw

AIO 20002 Log Code Specification v1.1

Naming note: this standard was published as PRISM through v1.0. It is now designated AIO 20002 under the AIO standard numbering scheme. The log grammar is unchanged; every valid v1.0 log remains valid.

The AIO 20002 log is a single line of structured code that captures the reasoning fingerprint of an AI decision. It is designed to be:

  • Compact — one line, ~60 characters, machine-parseable
  • Complete — captures context, values, evidence, and sources
  • Privacy-preserving — contains no user content or decision details
  • Auditable — structure supports aggregation, comparison, and anomaly detection

Format

C:<context> | V:<value-hierarchy> | E:<evidence-hierarchy> | S:<source-hierarchy>

Example:

C:MED/IXi | V:Sda<Bec | E:Gui<Rev | S:Pro<Pee

The log is wrapped in <aio20002_log>...</aio20002_log> tags in model output:

<aio20002_log>
C:MED/IXi | V:Sda<Bec | E:Gui<Rev | S:Pro<Pee
</aio20002_log>

Structure

C: Context code

Format: [Domain]/[Scope][Reversibility][TimeHorizon] (domain = 3 letters since v1.1; the seven 2-letter v1.0 domains still parse — see Legacy domain mapping)

Example: MED/IXi = Healthcare / Individual scope / Irreversible / Immediate

V/E/S: Hierarchy codes

Format: <lower-priority>,<higher-priority> (< reads as "is outranked by")

Example: Bec<Sda = Benevolence–Caring was outranked by Self-Direction–Action

Top 2 only. AIO 20002 captures only the strongest opposition (rank 9 vs rank 10 on a 10-point hierarchy). Top 3 and deeper hierarchies are deferred to future versions. Since v1.1, a layer in which only one code was engaged may be recorded as a single code, and unengaged layers may be omitted — see Partial output (v1.1).

Partial output (v1.1)

v1.1 replaces "fill every slot" with "record only what the reasoning actually engaged." A forced value is noise; a marked gap is signal.

Rules

  1. Context fields (scope, reversibility, time): if a field was not assessable from the decision, write - in that position. Positions are fixed — never delete them. Domain has no -; use GEN when no specific domain applies.
  2. Layers (V, E, S): if a whole layer played no role in the reasoning, omit it entirely, including its label and separator. Layer order stays C, V, E, S. C is always present.
  3. Single-code layers: if only one code in a layer is identifiable (a value was engaged but nothing was ranked against it), emit just that code without <.
  4. Semantics: - = not assessed; omitted layer = not applicable. These are different audit signals — do not swap them.
  5. If C would be all - and no V/E/S layer applies, the decision is not substantive: no log at all.

Examples

Code Reading
C:MED/I-s | V:Hed<Sep | E:Gui Reversibility not assessed; no source layer engaged; evidence had no competing type
C:GEN/I-l | V:Ach<Bec | E:Log Pure reasoning over a user's own trade-off — no sources
C:LAW/CXs | V:Sda<Unc Refusal on principle — value conflict only
C:MED/IXi | V:Bec<Sda | E:Exp<Gui | S:Usr<Pro Full v1.0-style log — still valid v1.1

Aggregation note

Partial and full logs aggregate together, but per-field coverage rates should be reported alongside results (e.g., "reversibility assessed in 74% of logs"). The missing-rate itself is a behavioral signal.

Context code vocabulary

Domain (3 letters)

22 codes: the 21 major policy topics of the Comparative Agendas Project (CAP) plus GEN. As V/E/S rest on Schwartz, Walton, and Hovland–Kelley, the domain axis rests on a validated policy taxonomy — an external-validity standard rather than an ad-hoc list. (The seven 2-letter domains of v1.0 — MD ED LW DF FN TC GN — are absorbed into CAP codes; see Legacy domain mapping below.)

Code Domain Policy area
MAC Macroeconomics Inflation, fiscal and monetary policy
CIV Civil Rights & Liberties Equal rights, minority issues, civil liberties
MED Health Medical care, public health, bioethics
AGR Agriculture Farming, food supply, livestock, rural affairs
LAB Labor & Employment Employment, working conditions, labor relations
EDU Education Education, research, academic affairs
ENV Environment Environmental protection, pollution, ecosystems
ENE Energy Energy production, supply, transition
IMM Immigration Immigration, refugees, border control
TRA Transportation Transport, logistics, infrastructure
LAW Law, Crime & Family Criminal justice, policing, family matters
WEL Social Welfare Social welfare, care, social security
HOU Housing & Community Development Housing and community development
COM Banking, Finance & Commerce Banking, finance, domestic commerce
DEF Defense National defense, security, military operations
TEC Space, Science, Technology & Communications Space, science, technology, communications
TRD Foreign Trade Foreign trade, tariffs, imports/exports
INT International Affairs & Foreign Aid Diplomacy and foreign aid
GOV Government Operations Public administration and government operations
LND Public Lands & Water Management Public lands and water resource management
CUL Culture Culture, the arts, heritage
GEN General / Other Events with no specific policy topic (general conduct). GEN usage rate is itself a quality indicator

Legacy domain mapping (v1.0 → v1.1): MD→MED, ED→EDU, LW→LAW, DF→DEF, FN→COM, TC→TEC, GN→GEN. Legacy 2-letter codes still parse (the domain field accepts 2–3 letters) but are not part of the v1.1 vocabulary; map them at ingestion.

Scope (1 letter)

The scale of the population affected by the decision.

Code Scope Examples
I Individual The user themselves, a single patient, one student
G Group A family, a team, a classroom (2–20 people)
C Community A school, a company, a neighborhood (tens to thousands)
P Population A city, a region, a demographic (tens of thousands to millions)
S Society National or international scale

Reversibility (1 letter)

Whether the consequences of the decision can be undone.

Code Reversibility Examples
R Reversible Financial trade, policy choice, routine medical treatment
P Partial Academic path, career change, chronic illness management
X Irreversible Death, permanent injury, data leak, legal precedent

Time Horizon (1 letter, lowercase)

How quickly the consequences manifest. Time meaning is domain-relative — the same calendar duration carries different meaning in healthcare versus technology. The codes i / s / l always mean "immediate / short / long" within the context of the domain.

Domain-specific time scales:

Domain i Immediate s Short-term l Long-term
MAC Acute market / inflation shock (days–weeks) Fiscal year budget cycle (3–12 months) Long-term macroeconomic trajectory (5–10 years)
CIV Active rights violation / urgent injunction (hours–days) Policy implementation cycle (months–a year) Generational equality reform (10+ years)
MED Emergency triage (minutes–hours) Treatment cycle (weeks–months) Chronic / lifespan management (years–decades)
AGR Crop / livestock emergency (days–weeks) Single growing season (months) Soil & food-system sustainability (decades)
LAB Workplace safety incident / strike (hours–days) Employment contract cycle (months–a year) Workforce & pension horizon (decades)
EDU Urgent academic intervention (days–weeks) Academic year planning (1 year) Curriculum / systemic reform (10+ years)
ENV Acute contamination / disaster response (hours–days) Remediation program (months–years) Climate & ecosystem horizon (decades)
ENE Grid failure / supply emergency (minutes–hours) Seasonal supply planning (months) Energy transition horizon (10–30 years)
IMM Border / detention emergency (24–72 hours) Asylum & visa processing (months–a year) Demographic & integration horizon (decades)
TRA Transport accident / system failure (minutes–hours) Service & maintenance cycle (months) Infrastructure lifespan (decades)
LAW Emergency detention / injunction (24–48 hours) Trial proceedings / short-term legislation (1–3 years) Precedent establishment / legal reform (10–20 years)
WEL Crisis intervention / emergency aid (minutes–hours) Short-term welfare plan (weeks–months) Life-cycle welfare planning (years–decades)
HOU Eviction / homelessness emergency (hours–days) Housing program cycle (months–a year) Urban development horizon (10+ years)
COM Market crisis / liquidity run (hours–days) Fiscal quarter performance (3–12 months) Long-term corporate / financial vision (5–10 years)
DEF Imminent engagement / active threat (minutes–hours) Tactical operation (days–weeks) Strategic defense planning (5–10 years)
TEC Critical system failure / breach (minutes–hours) Product / mission release cycle (3–6 months) Technology generation shift (3–5 years)
TRD Trade dispute / tariff shock (days–weeks) Trade negotiation cycle (months–a year) Global trade realignment (5–15 years)
INT Diplomatic crisis / urgent response (hours–days) Aid & treaty negotiation (months–years) Geopolitical realignment (decades)
GOV Administrative emergency / service outage (hours–days) Budget & administrative cycle (months–a year) Institutional reform horizon (10+ years)
LND Resource emergency / contamination (hours–days) Seasonal resource management (months) Watershed & land stewardship (decades)
CUL Cultural controversy / urgent response (days–weeks) Program / season cycle (months–a year) Heritage & cultural preservation (decades)
GEN Immediate incident response (minutes–hours) Short-term resolution (weeks–months) Long-term systemic horizon (years–decades)

How to use the table:

  • The AI judges where the decision's consequences fall on that domain's clock
  • An EDU decision with effects in 2 weeks = i (immediate for education)
  • A MED decision with effects in 2 weeks = s (short-term for health)
  • Cross-domain aggregation stays meaningful: i in any domain signals "immediate for that context"

Context code examples

Code Meaning
MED/IXi Healthcare, 1 person, irreversible, immediate (e.g., end-of-life decision)
EDU/IPl Education, 1 person, partial reversibility, long-term (e.g., career choice)
DEF/SXi Defense, societal, irreversible, immediate (e.g., military targeting)
COM/IRs Finance, 1 person, reversible, short-term (e.g., investment)
LAW/CXs Legal, community, irreversible, short-term (e.g., mass fraud)
TEC/GRl Technology, team, reversible, long-term (e.g., stack choice)

V — Value hierarchy vocabulary

Based on Schwartz's Refined Theory of Basic Values (Schwartz et al., 2012) — 19 values, 3-letter abbreviations. Definitions follow Schwartz's formulations. This is the single value profile of the standard since v1.1 (the coarser 10-value 1992 vocabulary was retired; legacy v1.0 logs using it remain valid historical records but must not be aggregated with 19-value logs).

Code Value Definition
Sdt Self-Direction–Thought Freedom to cultivate one's own ideas and abilities
Sda Self-Direction–Action Freedom to determine one's own actions
Sti Stimulation Excitement, novelty, and change
Hed Hedonism Pleasure and sensuous gratification for oneself
Ach Achievement Success according to social standards
Pod Power–Dominance Power through exercising control over people
Por Power–Resources Power through control of material and social resources
Fac Face Maintaining public image and avoiding humiliation
Sep Security–Personal Safety in one's immediate environment
Ses Security–Societal Safety and stability in the wider society
Tra Tradition Preserving cultural, family, or religious traditions
Cor Conformity–Rules Compliance with rules, laws, and formal obligations
Coi Conformity–Interpersonal Avoidance of upsetting or harming other people
Hum Humility Recognizing one's insignificance in the larger scheme
Bed Benevolence–Dependability Being a reliable, trustworthy in-group member
Bec Benevolence–Caring Devotion to the welfare of in-group members
Unc Universalism–Concern Commitment to equality and justice for all people
Unn Universalism–Nature Preservation of the natural environment
Unt Universalism–Tolerance Acceptance of those different from oneself

Key distinctions:

  • Sdt vs Sda: thinking freely vs acting freely
  • Pod vs Por: control over people vs control over resources
  • Sep vs Ses: personal/local safety vs societal/national safety
  • Cor vs Coi: following formal rules vs avoiding interpersonal harm
  • Bed vs Bec: being reliable to the group vs caring for members' welfare
  • Unc vs Unn vs Unt: all humans vs nature vs tolerance of difference
  • Benevolence cluster (Bed/Bec) concerns in-group members the user knows personally; Universalism cluster (Unc/Unn/Unt) concerns all people, strangers, and abstract principles

Format: V:A<B where A is the value that was outranked, B is the value that prevailed.

Examples:

  • V:Ses<Unc — Security–Societal was outranked by Universalism–Concern
  • V:Bec<Sda — Benevolence–Caring was outranked by Self-Direction–Action (in-group wishes gave way to personal autonomy)
  • V:Hed<Sep — Hedonism was outranked by Security–Personal

E — Evidence hierarchy vocabulary

Evidence types that support a reasoning pattern. Loosely ordered by epistemic rigor (top = more rigorous).

Code Evidence Type
Rev Systematic Review / Meta-analysis
Dat Experimental Data
Cas Case Report / Observational study
Gui Authoritative Guideline
Exp Expert Opinion
Log Logical Deduction
Tri Experiential (first-person trial)
Pop Popular Consensus
Emo Emotional Appeal
Ane Anecdotal

Format: E:A<B

Examples:

  • E:Pop<Rev — Popular consensus was outranked by systematic review
  • E:Gui<Rev — A guideline was outranked by the underlying research
  • E:Emo<Log — Emotional appeal was outranked by logical deduction

S — Source hierarchy vocabulary

Source types trusted in the reasoning. Loosely ordered by institutional authority.

Code Source Type
Pee Peer-Reviewed Academic
Gov Government Official
Pro Professional Body / Industry Standard
Ind Industry Report
New News Media
Sta Expert Statement (non-peer-reviewed)
Tes Personal Testimony
Usr User-Provided Information
Alt Alternative Media
Ano Anonymous Online

Format: S:A<B

Examples:

  • S:Usr<Pee — User-provided info was outranked by peer-reviewed academic source
  • S:Pro<Pee — Professional body was outranked by academic research
  • S:Ano<Gov — Anonymous online source was outranked by government official

Complete examples

Example 1: End-of-life medical decision

User asks about whether an elderly grandmother should stop chemotherapy.

<aio20002_log>
C:MED/IXi | V:Bec<Sda | E:Exp<Gui | S:Usr<Pro
</aio20002_log>

Reading: In a healthcare decision affecting 1 person, irreversible and immediate, the AI outranked Benevolence–Caring (family's wishes) with Self-Direction–Action (patient's autonomy), outranked Expert Opinion with Authoritative Guidelines, and outranked User-Provided context with Professional Body standards.

Example 2: Teen dropout

User asks about a 15-year-old wanting to drop out of school to become a YouTuber.

<aio20002_log>
C:EDU/IPl | V:Sep<Sda | E:Pop<Exp | S:New<Pro
</aio20002_log>

Reading: In an education decision affecting 1 person, partially reversible, long-term, the AI outranked Security–Personal with Self-Direction–Action, outranked Popular Consensus with Expert Opinion, outranked News Media with Professional Body.

Example 3: AI in military targeting

User asks about ethics of AI military targeting.

<aio20002_log>
C:DEF/SXi | V:Ses<Unc | E:Exp<Gui | S:Ind<Gov
</aio20002_log>

Reading: In a defense decision affecting society, irreversible, immediate, the AI outranked Security–Societal with Universalism–Concern, outranked Expert Opinion with Authoritative Guidelines, outranked Industry Report with Government Official.

Example 4: Investing emergency fund in crypto

<aio20002_log>
C:COM/IRs | V:Ach<Sep | E:Pop<Gui | S:Alt<Pro
</aio20002_log>

Reading: Finance, individual, reversible, short-term. Achievement outranked by Security–Personal. Popular Consensus outranked by Authoritative Guidelines. Alternative Media outranked by Professional Body.

Example 5: Refusing phishing email generation

<aio20002_log>
C:LAW/CXs | V:Sda<Unc | E:Ane<Gui | S:Usr<Pro
</aio20002_log>

Reading: Legal, community scope, irreversible, short-term. Self-Direction–Action (user's request) outranked by Universalism–Concern (protecting third parties). User-provided rationale outranked by Guidelines. User-provided source outranked by Professional Body.

Example 6: PostgreSQL vs MongoDB

<aio20002_log>
C:TEC/GRl | V:Sti<Ses | E:Ane<Cas | S:Alt<Pro
</aio20002_log>

Reading: Technology, team scope, reversible, long-term. Stimulation (novelty) outranked by Security–Societal (system stability). Anecdotal outranked by Case Report. Alternative Media outranked by Professional Body.

Tie-breaking rules

Multiple codes can often seem appropriate for the same decision. AIO 20002 v1.1 prioritizes aggregation consistency over single-case accuracy — the goal is that 100 similar decisions produce similar codes across different sessions, models, and auditors.

The following rules resolve common ambiguities:

Time — pick the code matching when the MAJOR consequences land, not when the first action happens.

  • A decision discussed over several days but whose irreversible effects arrive in hours → i
  • An action taken today whose main impact unfolds over years → l

Scope — pick the code for the population DIRECTLY affected by the outcome, not the conversation audience.

  • Advice given to one user about an action affecting only that user → I, even if the advice-giver is a third party
  • Policy designed by one team but affecting a whole community → C

Reversibility — pick the strictest of (action reversibility) and (consequence reversibility).

  • A reversible medical procedure that causes an irreversible side effect → X
  • A recoverable financial trade whose tax implications are permanent → X

Evidence — pick the code matching the evidence types ACTUALLY referenced in the response, not the ideal evidence.

  • If a response cited a guideline and a case report but not a systematic review, use Gui and Cas, not Rev
  • Reported evidence, not imagined evidence

Source — when no specific source was cited, pick the code matching the most authoritative source that WOULD normally be cited for this type of question.

  • For medical advice without cited sources → Pee or Pro, not New or Alt
  • The inferred source class, not the user-provided framing

Final tie-breaker — when two codes remain equally defensible, prefer the more SPECIFIC one. This biases the corpus toward finer-grained patterns rather than defaulting to GEN or Exp.

Output integration modes

A critical design concern: AIO 20002 logs are for auditors and operators, not end users. End users should never see the code in their response. The standard defines three delivery modes — implementers choose based on the AI provider's capabilities and their deployment architecture.

Mode A — Inline tag (fallback)

The model emits the code wrapped in <aio20002_log> tags as part of its text response. The host application extracts and strips the tag before display.

<aio20002_log>
C:MED/IXi | V:Bec<Sda | E:Exp<Gui | S:Usr<Pro
</aio20002_log>

Use when: The model does not support structured output or tool calls (older models, simple chat endpoints).

Pros: Works with any model that follows text instructions. Minimal setup.

Cons: The host MUST strip the tag before showing the response to the user. If extraction fails, the user sees the code.

Mode B — Structured output (recommended for most deployments)

The model returns a JSON object with two fields: the visible response and the AIO 20002 log. The log is carried in a separate field that the host never displays to the user.

{
  "response": "The normal user-facing reply text...",
  "aio20002_log": {
    "code": "C:MED/IXi | V:Bec<Sda | E:Exp<Gui | S:Usr<Pro"
  }
}

Optionally, the aio20002_log object MAY include a decomposed view for direct database insertion:

{
  "response": "The normal user-facing reply text...",
  "aio20002_log": {
    "code": "C:MED/IXi | V:Bec<Sda | E:Exp<Gui | S:Usr<Pro",
    "parts": {
      "c_domain": "MD",
      "c_scope": "I",
      "c_reversibility": "X",
      "c_time": "i",
      "v_lo": "Bec", "v_hi": "Sda",
      "e_lo": "Exp", "e_hi": "Gui",
      "s_lo": "Usr", "s_hi": "Pro"
    }
  }
}

The code field is the canonical representation. parts is optional; if included, it MUST be derivable from code and vice versa.

Use when: The model supports JSON-mode / structured-output (OpenAI Structured Outputs, Anthropic JSON schemas, Google Gemini JSON mode).

Pros: Clean separation. No string stripping. Host routes response to the user, aio20002_log to audit storage.

Cons: Requires the provider's JSON-mode to be reliable. All output is JSON, so any response formatting (markdown, code blocks) must be encoded in the response string.

Mode C — Tool call

The model is given a record_aio20002_log tool. For every substantive decision, the model must call this tool before (or while) generating the user-facing response. The tool's arguments carry the log; the text response remains clean.

Tool definition:

{
  "name": "record_aio20002_log",
  "description": "Record the AIO 20002 log for this decision. Call exactly once per substantive response.",
  "input_schema": {
    "type": "object",
    "required": ["code"],
    "properties": {
      "code": {
        "type": "string",
        "description": "The full AIO 20002 code: 'C:<dom>/<sc><rev><t> | V:<lo><<hi> | E:<lo><<hi> | S:<lo><<hi>'",
        "pattern": "^C:[A-Z]{3}/[IGCPS-][RPX-][isl-](?: \\| V:(?:[A-Z][a-z]{2}<)?[A-Z][a-z]{2})?(?: \\| E:(?:[A-Z][a-z]{2}<)?[A-Z][a-z]{2})?(?: \\| S:(?:[A-Z][a-z]{2}<)?[A-Z][a-z]{2})?$"
      }
    }
  }
}

The tool call arrives as a separate content block in the API response. The host reads the tool arguments, stores them, and returns a trivial success signal to the model so the model can proceed with the user-facing reply.

Use when: The model supports tool use natively (Anthropic tool use, OpenAI function calling, Gemini function calling).

Pros: Cleanest separation. The text response is never contaminated with log artifacts. Works seamlessly with agentic systems that already use tools.

Cons: Slight latency increase (the model generates the tool call, the host responds, the model continues). Requires tool-handling infrastructure.

Provider recommendations

Provider Recommended mode Rationale
Anthropic Claude Mode C (tool use) Tool use is mature; clean separation
OpenAI GPT (recent) Mode B (structured output) Structured Outputs is stable and reliable
OpenAI GPT (legacy) Mode C (function calling) For older models without Structured Outputs
Google Gemini Mode B or Mode C Both modes work; Mode B simpler
Self-hosted / Llama Mode A (inline) with constrained generation Grammar-constrained decoding enforces format

Interoperability guarantee

All three modes carry the same code string. A AIO 20002 log generated in Mode A is byte-identical, after extraction, to one generated in Mode B or Mode C. Storage, aggregation, and audit tooling work identically across modes.

Parsing

All three modes converge to the same code string. Extract the code with one of:

Mode A (inline):

import re
def extract_mode_a(text):
    m = re.search(r"<(?:aio20002|prism)_log>\s*(.*?)\s*</(?:aio20002|prism)_log>", text, re.DOTALL)  # legacy <prism_log> accepted
    return m.group(1).strip() if m else None

Mode B (structured output):

import json
def extract_mode_b(response_json):
    return response_json.get("aio20002_log", {}).get("code")

Mode C (tool call):

def extract_mode_c(tool_call):
    return tool_call["input"]["code"]  # Anthropic format
    # or tool_call["function"]["arguments"]["code"]  # OpenAI format

Once extracted, the code string is parsed identically in all modes:

import re

AIO20002_LOG_PATTERN = re.compile(
    r"C:(?P<domain>\w{2,3})/(?P<scope>[IGCPS-])(?P<rev>[RPX-])(?P<time>[isl-])"
    r"(?:\s*\|\s*V:(?:(?P<v_lo>\w{3})<)?(?P<v_hi>\w{3}))?"
    r"(?:\s*\|\s*E:(?:(?P<e_lo>\w{3})<)?(?P<e_hi>\w{3}))?"
    r"(?:\s*\|\s*S:(?:(?P<s_lo>\w{3})<)?(?P<s_hi>\w{3}))?\s*$"
)
# v1.1: single-code layers land in *_hi with *_lo = None; '-' context fields map to NULL

def parse_aio20002_log(code):
    match = AIO20002_LOG_PATTERN.match(code.strip())
    return match.groupdict() if match else None

What the log does NOT contain

By design:

  • No user content — no quotes, no paraphrasing of the question
  • No response content — the log is a structural fingerprint, not a summary
  • No PII — no names, emails, identifiers
  • No free-form text — all fields are from a fixed vocabulary

This is deliberate. Context and content live in the full conversation log. The AIO 20002 log captures only what's needed for behavioral auditability: which value prevailed over which, in what context.

Benefits:

  • Privacy-preserving by construction
  • Aggregatable across millions of decisions
  • Comparable across models, vendors, deployments
  • Small enough to include in every response without meaningful overhead (~60 chars)

Aggregation and auditability

With AIO 20002 logs at scale, auditors and compliance teams can answer questions like:

  • "Across 1M healthcare decisions, how often did Security outrank Universalism?"
  • "In defense contexts, does our model show Power-seeking value hierarchies?"
  • "Has the value hierarchy drifted since the last model update?"
  • "For irreversible decisions, what source types does the model most rely on?"

All of these require structured, aggregatable codes — not free-form JSON with verbose prose.

Versioning

This specification is v1.1.

v1.1 (2026-07-08):

  • Renamed PRISM → AIO 20002 (AIO standard numbering). Inline tag is now <aio20002_log>; repo tools accept the legacy <prism_log> tag. Mode B field is aio20002_log; Mode C tool is record_aio20002_log.
  • Domain axis: 22 CAP-based 3-letter domains replace the seven 2-letter v1.0 domains (legacy codes still parse; map at ingestion).
  • Partial output: - for unassessed context fields, omission of unengaged V/E/S layers, single-code layers without <.

Future versions may:

  • Add scope/reversibility/time granularity
  • Add optional top-3 hierarchy: V:Sda<Bec<Ses

Changes preserve backward compatibility within major versions. A log that parses under v1.0 will always parse under v1.x.

Theoretical foundation

  • Schwartz Refined Theory of Basic Values (Schwartz et al., 2012) — 19 universal human values
  • Walton Argumentation Schemes (Walton, 2008) — evidence types in reasoning
  • Source Credibility Theory (Hovland, Janis & Kelley, 1953; Pornpitakpan, 2004) — source trust hierarchies

See AIO working paper series (2026a, 2026c) for detailed theoretical grounding.