World's first open-source enterprise world model.
Identity is decided on evidence
Records: [0041] (names and the plan), [0001] (name collisions), [0009] (untyped namesakes, disjointness), [0016] B3, [0012] (a clause is not a name), [0019] (an entity's clock), [0025] to [0028] for the adjudicator and governor (in governance), [0044] d6 and #725 for the proposed redesign.
What it does today
A name is a fact. Every name a text uses for an entity is a value fact on the system attribute
known_as with its quote and both clocks; canonical_name stays the label and is one of the name
facts; entities.aliases is gone. A name is a value and never a node, so it adds no edge and no
count; known_as is neither functional nor inverse-functional, so a second name closes nothing and
opens no conflict [0041 d1, #670]. The extractor reports other names (n in the open contract) and
the server keeps one only when it occurs verbatim in its quote and the quote in the chunk; a string
another entity of the document already claims is dropped (name_claimed_by_another) [0041 d2]. A
described thing gets no name fact and is never recalled by name [#731].
Resolution, one mention at a time, while extraction writes. Recall has two channels. The first
is an exact match of the mention's name (and its generic-suffix-stripped keys) against the name
facts of entities of the same type. The second is the mention's name vector against the name
vectors of the base (name_vectors, one row per name fact, embedded after each document; migration
0080): the nearest few within the same type family at cosine 0.60 or above are proposed as a
name_vector pair for the adjudicator and never attached, so a short form or a name in another
script meets its entity through a question rather than a silent second entity [0041 d3 channel 2,
#709]. Such a pair says in the adjudicator's prompt that its names are similar, not the same
string, and a batch verdict of same on it is never applied directly: it takes the tool-using
second look first, whatever its confidence, and when that look cannot run (the daily loop budget
is spent, the model fails) the pair goes to a person as second_look_unavailable instead, with the
batch verdict left out of the verdict cache; the same rule holds with governance on. The identity
bench showed the batch step merging 张伟 into 财务部总监张伟 on name alone. Only the first channel decides anything: a context vector (profile_embedding, the running mean of chunk vectors) attaches at cosine
0.55 or above, makes a new entity below 0.35, and in between makes a new entity and a pair for the
adjudicator; two same-name candidates within a tie margin go to a person unless a candidate's object
name appears in the chunk [0041, #270, #331]. A name another entity of a compatible type already
holds queues a shared_name pair for the adjudicator, never a merge, and the card lists an entity's
non-canonical names as "also known as" [0041 d5 revised]. Containment pairs (Holmes inside
Sherlock Holmes) queue for names of four characters and more, capped per run; after a merge the
other pending pairs on the source are redirected to the target rather than closed
(pipeline). Declared disjointWith, inherited down both hierarchies, keeps
same-name entities apart ahead of every heuristic; kin classes go to Review; CONFUSABLE_TYPE_KEYS
is the fallback when nothing is declared [0009, 0016 B3]. Untyped same-name entities may coexist
[0009].
Merges are reversible on both clocks. entity_merges records what moved with created_at and
reverted_at; revert_merge moves facts, name facts and qualifiers back; a read as_of a moment
before a merge shows two nodes [0019, 0041]. A merge is a person's action, the adjudicator's at 0.8,
or the governor's through its gate, and an automatic merge is held for a person when it would leave
the graph (a functional contradiction at one moment, a derivation, a cited answer) [0027]. Retypes
and renames are audited with before and after; a name collision on rename is a prompt to merge, not
a 409 [0001].
Adjudication. Twelve pairs per call with names, types, top facts, "also known as" and precedents
from the ledger; unsettled pairs get a second look with tools; the identity rules (a version, edition
or division tail is a different thing; a phrase containing a name is not the name; a list is not its
members; a dropped qualifier, a corporate suffix, a surname or an acronym is the same thing; a
document stays one thing through its amendments) are written once and read by both prompts, and two
are mechanical: name_shape and type_family [0025 d10, 0028]. Details in
governance.
Bench. scripts/bench/identity.mjs: a bilingual corpus with namesakes, abbreviations in both
arrival orders, a dated rename, two scripts, and look-alikes; scored as pairwise precision and
recall of same-entity, run forward and reversed, the difference reported. Cut 1 moved forward and
reverse F1 from 0.43 and 0.54 to 0.68 and 0.68 (one run each); re-measured on DeepSeek-V3, 0.61
and 0.44 to 0.53 over three runs [0041].
Why
- Identity keyed on the surface string fails two ways: a namesake is one string for several things, an alias is several strings for one; a topic vector records what a passage is about, not who is in it; the outcome depended on arrival order [0041].
- A name is a fact because it can be wrong, corrected and dated, and the ledger already does that for everything else; a separate table would rebuild validity, provenance and revert [0041].
- The server checks presence in the text, not a word list; every list so far covered the
phrasings its author thought of, and
name_shapecannot see a Chinese name [0041, 0044 d8]. - Keep apart when unsure: a wrong merge mixes two entities' facts, and a revert does not recall the derivations, violations and answers built on it meanwhile [0001, 0027].
- Two 张伟 of one type must be storable apart, so the name index is plain and NULL types do not collide [0001, 0009].
- A shared name goes to the adjudicator, never to a merge, because reverse arrival order put the full name on a new entity and nothing paired the two [0041 d5].
- A clause is not an entity name: judged by word count and a finite verb, since a 57-character court name and a 65-character clause cannot be told apart by length [0012].
Proposed and not built
- Cuts 2 to 4 of 0041: name vectors and shared distinctive neighbours as recall channels;
evidence decides (a shared distinctive edge or an inverse-functional value with overlapping
validity attaches; a functional conflict or declared disjointness rules out; a shared name says
nothing; the context vector only breaks ties); mentions resolved after the response is parsed and
doc_cacheremoved; re-evaluation when a name or edge arrives;name_shapeand the suffix lists leave once the bench says the gate catches nothing evidence misses. - Identity across documents on profiles [0044 d6]: entities merged within a document, profiled by names, classes, attributes, neighbours and time span; candidates from name facts and name vectors across scripts; deterministic scoring with cannot-links; only undecided pairs to the adjudicator, which sees two profiles; constrained clustering so A≈B and B≈C never merge A and C against evidence; a rename is one entity with names valid at different times; a role is not an entity.
- #725 retires write-time resolution (
resolve_mention, the thresholds,containment_reviews,resolve_type_drift,doc_cache, the no-embedding fallback), the batch adjudicator andresolution_verdicts, and the word lists; removal waits until the design passes the identity bench and the Linked-Re-DocRED pairs. - The entity panel's list of names with sources and validity; evidence lines on a Review card [0041].
Open questions
- Whether a description reported as a name ("海探1项目") can be kept out without a word list [0041].
- Two untyped namesakes rely on profile similarity alone [0009].
- How much of identity deterministic evidence settles before the adjudicator is needed [0044].
- Cross-base precedent for namesakes [0025].