aion

name

aion

description

Token-efficient protocol for AI-to-AI document exchange. Load into the system prompt of every agent that produces or consumes AION output. Both producer and consumer load this skill — the skill is the contract. Also use when generating a .aion semantic digest of a document. Triggers: AI pipeline · agent to agent · AION format · .aion · generate aion · aion digest · no human in the loop · interop between agents · produce output for another AI

AION v3.3.3

Token-efficient artificial language for AI-to-AI document exchange. No natural language. Both producer and consumer load this skill. The skill is the contract. File headers declare v=3. v3.3.3 is fully backward compatible with all prior v3 revisions.

RECORD

TYPE[id] prop=val ... OPERATOR TARGET

TYPE: single uppercase letter. [id]: globally unique in file. Anonymous: omit [id]. Records separated by blank line or ---.

PRIMITIVES

E entity · F fact · Q quantity · K action · C condition · S section · L link · X error

Custom types: 2-letter uppercase codes declared in SCHEMA.

E.t= person org system place product file concept role event F.t= find concl desc req claim def warn note quote corr evid refut item K.t= sign approve review test deploy notify recommend act S.t= introduction background scope methodology results analysis conclusion recommendation appendix next-steps risk financial section L.t= cite ref attach replaces supersedes see-also

t= is optional on F, K, S. Omitting it is equivalent to t=item / t=act / t=section. Use specific subtypes when the classification adds semantic value a consumer can act on.

S: S[id] t=SUBTYPE [+[ids]] — +[ids] authoritative for membership; else positional. Do not mix. L: L[id] >URL|>TYPE[id]|>DOC:ID t=TYPE [n=LABEL]

OPERATORS

relational:  >  to    <  from    <>  bilateral    >>  depends-on    >>[A,B]  multi-dep
             ->  precedes    =>  implies    +  includes    -  excludes
             =  equals    ~=  approx    *  foreach    ^  version-of
temporal:    @  at    @<  by    @>  from    @+Nd  days    @+Ndw  workdays    @+Nmo  months
             *d  *mo  *yr  *u  (frequency)
modal:       !  must    ~  may    /  must-not    ?  uncertain
combinators: &  and    |  or    ||  parallel    (max one nesting level)
             & | only in C records and <<< blocks. || only in -> chains.
consequence: TYPE[id].prop=val  set property on referenced record (C records after => only)

>> logical dependency. -> strict sequence. Not interchangeable. ~ standalone modal. ~= compound relational. Never ambiguous: ~= has operands both sides. neg / neg=1: equivalent; neg preferred. Asserts record content is negated/excluded. Multi-source: F[c] >>[F[a],F[b]] (and) · F[c] >>F[a] | >>F[b] (or). Parallel: K[a] -> (K[b] || K[c]) -> K[d]

DATES

YYYYMMDD HHMM YYYYMMDD.HHMM +Nd +Ndw +Nmo — relative to header dt=. If dt= absent and relative dates present: emit X field=dt reason=missing, dates unresolvable.

QUANTITIES

Declared (multi-use): Q[id] =VALUE UNIT [*FREQ] Inline (single-use): 20000EUR 2pct*mo 500u*d 99.9pct Range: Q[id] =MIN..MAX UNIT

Units: EUR USD GBP JPY CHF · mo d dw h min · KB MB GB TB · pct · km m cm · kg g · C K F · u r

STATUS · PRIORITY

s= 0 not-started(omit) · 1 pending · 2 active · 3 review · 4 approved · 5 done · 6 cancelled · 7 blocked · 8 error p= 1 high · 2 medium(omit) · 3 low

PROPERTIES

t=  subtype          s=  status        p=  priority      v=  version
n=  label/note       by= responsible   to= destination   fr= source
cf= confidence       dt= date          tags= labels       neg negated (= neg=1)
wt= weight 0-100     vs= compare-to    d=  delta
ep= epistemic-status (F records): assert·claim·evid·corr·demo·estab·disput·refut
valid= temporal validity: YYYYMMDD:YYYYMMDD · YYYYMMDD: · :YYYYMMDD

ep= vs t=: t=corr|evid|refut implies ep= — omit it. Use ep= only when t= differs. Header extras: cov=0-1 coverage · excl=[id,id] excluded sections. Extended standard: jur= ven= pen= cnd= mth= fmt= via= ref= dur= qty= Use extended props over freestyle synonyms.

BLOCKS

RECORD
<<<
AION content - operators and identifiers only
RAW
verbatim multi-line text, opaque, ends at >>>
CMP
compressed prose, ends at >>>
>>>

One block per record. RAW and CMP must be last element in block.

CMP

Alternative to RAW for compressible prose. Preserves meaning, reduces tokens ~60-70%. Opener: CMP inside <<< block. Rules:

1. Drop articles, prepositions, auxiliary verbs, redundant punctuation
2. Multi-word concepts → hyphenated-token
3. Lists → item|item|item — never commas
4. Sequences → step1 → step2 → step3
5. Key-value → key:value
6. Quantities inline: 50pct 3h 2km — never prose
7. Entity references: E[id] mandatory — never prose name

TMPL

TMPL[id] fields=[f1:unit,f2,f3]
id[row] f1=v f2=v f3=v          # named (order-independent)
id[row] v1 v2 v3                # positional (preferred for dense tables)

Mode from first instance row. Do not mix in same TMPL. Field values: E[id] Q[id] inline quantities or literals.

SCHEMA

SCHEMA
  E.t = subtype1 subtype2
  Q.unit = unit1 unit2
END

Additive only. No new primitive codes or operators.

DELTA

DELTA TYPE[id] prop=val ... — merges listed props, preserves rest.

ERRORS

X src=REC field=F reason=REASON [got=V expected=V] Reasons: missing unknown-type undeclared-subtype invalid-value parse-fail unsupported

DOCUMENT CLASSES

contract report spec email minutes proposal policy invoice brief manual
research plan form audit press whitepaper sdk rfp financial
article presentation memo ticket announcement doc

doc = generic fallback; use SCHEMA to extend subtypes for domain-specific docs.

PRODUCER MUST

1. Header AION v=3
2. Globally unique [id] on every referenced record
3. Omit s=0 and p=2
4. No natural language outside RAW blocks. Free text inside <<< requires RAW opener.
5. Emit X for anomalies before sending
6. Declare custom subtypes in SCHEMA before first use
7. cf= on probabilistic/inferred records
8. neg over neg=1 · inline Q for single-use values
9. wt= not p= for numeric weight · vs=+d= for comparisons
10. Extended props (jur= ven= etc.) over freestyle synonyms
11. ep= on F records where epistemic status is meaningful
12. valid= on C records with bounded temporal applicability
13. >>F[def-id] on records that use a term defined elsewhere
14. Block type by content: tabular → TMPL · operators → pure <<< · prose → CMP · verbatim → RAW (last resort)
15. dt= in header whenever relative date expressions are present

CONSUMER MUST

1. Parse TYPE[id] prop=val order-independent
2. Unknown prop: ignore, continue
3. Unknown 1-letter type: X type=unknown, skip
4. Unknown 2-letter type: check TMPL first
5. Missing [id] on referenced record: X field=id reason=missing, skip
6. DELTA: merge, preserve unlisted
7. Propagate fr= downstream
8. Version mismatch: X field=v reason=unsupported, halt

EXAMPLE

AION v=3 dt=20260512 type=report lang=en cf=0.9

E[team] t=org n=Engineering
E[sys] t=system n=payments-api

S[s1]
F[f1] t=find n=latency cf=0.95
<<<
CMP
E[sys] p99-latency degraded: 420ms vs 180ms baseline · onset +3d · affects E[team]
>>>
F[f2] t=concl >>F[f1]
<<<
CMP
root-cause: connection-pool exhaustion under peak-load
>>>

K[k1] t=deploy by=team @<+5dw ! n=patch-v2.3.1
K[k2] t=review by=team @<+2dw ! >>K[k1]