applied-nlp-ner-stage2-expanded-798

Stage 2 output of an applied-NLP NER project: the Stage 1 corrected dataset (applied-nlp-ner-stage1-cleaned-98) expanded with 700 newly labeled records from four public sources, split into train/test.

This dataset is never overwritten by later stages: Stage 4's revision is published as a separate repository.

Record format

Same as Stage 1: unique_index, source_text, privacy_mask (a list of {start, end, label, value} entities).

Splits

• train: 679 records
• test: 119 records

Split is stratified by source (wikipedia / wikinews / gutenberg / govreports / starter-corrected) at roughly 15% test per source, seeded (20260728) for reproducibility, computed once across the combined 798-record set rather than as two separately-split halves.

Sources and licensing

175 records each from four sources, chosen to extend the register diversity already present in the Stage 1 starter set (which mixed news, reviews, legal text, and casual web text) without relying on any source of uncertain license:

• Wikipedia (wikimedia/wikipedia, dump 20231101.en) — CC BY-SA 3.0 / GFDL. Formal encyclopedic register, dense in PERSON, ORGANIZATION, LOCATION, WORKOFART, PRODUCT and TIMEDATE.
• Wikinews (izumi-lab/wikinews-en-20230728) — CC BY 2.5. News register: current events, organizations in the news, dates and monetary amounts in a journalistic style.
• Project Gutenberg (NLTK's bundled gutenberg sample corpus, 18 public- domain texts including Austen, Melville, Shakespeare, the KJV Bible, and children's literature) — public domain (pre-1928 US copyright). Narrative register: strong on PERSON and WORKOFART, weak on modern PRODUCT/ORGANIZATION.
• US Congressional bill text (billsum, sourced from congress.gov) — public domain (17 U.S.C. §105: works of the US federal government carry no copyright); the compiled dataset is released CC0. Formal legislative register, matching starter record 99's World Bank-style report register: dense in ORGANIZATION, LOCATION, WORKOFART (named Acts) and AMOUNT.

Attribution (title and URL) is retained per-record in the source pull scripts (scripts/pull_*.py) even though it is not required for public-domain or CC0 text, to keep Wikipedia/Wikinews (CC BY-SA / CC BY) attribution available.

Selection method

Each source was over-sampled (~260 candidates), deduplicated (exact and near-duplicate by normalized text prefix — Wikinews alone had 80 near-duplicate candidates dropped, mostly repeated article boilerplate), length-filtered to the starter set's range (58-1912 characters), then capped at 175 per source so the published mix is balanced by source rather than by whatever volume a source happened to return. 700 new records, plus the 98 corrected Stage 1 records, for 798 total.

Labeling method

Labeled under the full policy below (D1-D15) using LLM automation: the policy document plus each record's text was given to a labeler with instructions to follow the policy's worked examples exactly, in batches, then the returned value+label pairs were resolved to exact character offsets programmatically and validated (a value that cannot be located as an unclaimed substring of its record's text is dropped and logged rather than guessed at). Cross-batch consistency was maintained by keeping the same instructions and policy text across every batch. Five records exposed genuine policy gaps not covered by the Stage 1 rules (named wars/battles, named legislation, mythological figures, battle-honour composites) - see D12-D15 in the policy below, added before bulk labeling began, each derived from a small hand-labeled pilot sample and cross-checked against Stage 1 precedent before being applied at scale.

Validation. Of ~4185 proposed spans, 10 (0.24%) could not be located as an exact substring of their record's text (typically a Unicode character such as a non-breaking space or en-dash that did not survive transcription) and were dropped rather than force-fit; every placed span is a verified exact substring of source_text[start:end].

Known limitation. PRODUCT is the smallest label at 150 spans (3.6% of all spans), because two of the four sources (Gutenberg, government bill text) contain almost no branded commercial goods by nature of their register. This is a property of the source mix, not an error, and is reported here rather than corrected for by artificially injecting product mentions.

Statistics

Stage 2 statistics

• total records: 798 (98 from Stage 1 + 700 new)
• total spans: 4175
• train: 679, test: 119

Records by source

source	records
wikipedia	175
wikinews	175
gutenberg	175
govreports	175
starter-corrected	98

Label counts

label	count
PERSON	804
LOCATION	681
TIMEDATE	670
ORGANIZATION	656
JOB	461
AMOUNT	404
WORKOFART	349
PRODUCT	150
total	4175

Text length (characters)

• min 58, median 329, max 1659, mean 365

Entity density (spans per 1000 characters)

• min 0.0, median 11.0, max 60.7, mean 13.5

---

Annotation policy

This is the complete policy in force for this dataset. It has three parts: the label definitions and the baseline labeling rules, both fixed and reproduced from the task specification; the governing principles derived from the added rules; and the added rules themselves (D1–D10) with the record-removal decision (R1).

Policy version: v1.0 (Stage 1).

---

Labels

• PERSON: a named person, including given names and surnames.
  • Example: <PERSON>Barack Obama</PERSON> / <PERSON>Sarah Chen</PERSON>.
• ORGANIZATION: a named company or institution.
  • Example: <ORGANIZATION>Google</ORGANIZATION> / <ORGANIZATION>Mayo Clinic</ORGANIZATION>.
• LOCATION: a named place such as a city, country, region, street, landmark, or geographic area.
  • Example: <LOCATION>Berlin</LOCATION> / <LOCATION>New York City</LOCATION>.
• TIMEDATE: an expression that places something on a timeline, including dates, clock times, and durations used as time.
  • Example: <TIMEDATE>March 15, 2024</TIMEDATE> / <TIMEDATE>50 minutes</TIMEDATE>.
• PRODUCT: a named commercial product, device, or branded good.
  • Example: <PRODUCT>iPhone 15</PRODUCT> / <PRODUCT>MacBook Pro</PRODUCT>.
• WORKOFART: a named creative or published work such as a book, film, song, or titled publication.
  • Example: <WORKOFART>Oppenheimer</WORKOFART> / <WORKOFART>Spider-Man: Brand New Day</WORKOFART>.
• JOB: an occupational title or formal work role when it functions as such in the sentence.
  • Example: <JOB>software engineer</JOB> / <JOB>CEO</JOB>.
• AMOUNT: a measurable or countable quantity, not a time or date expression.
  • Example: <AMOUNT>50</AMOUNT> tickets / <AMOUNT>120,000</AMOUNT>.

---

Baseline labeling rules

These are the main rules. They are fixed and are not edited by this policy. All added rules below are consistent with them.

• A labeled mention should cover the entity itself, not the surrounding grammar.
  • Example: Correct: She visited <LOCATION>Paris</LOCATION> yesterday. Incorrect: labeling visited Paris yesterday as one span.
• Articles, prepositions, conjunctions, and other function words should stay outside the span unless they are truly part of the proper name.
  • Example: Correct: He works at <ORGANIZATION>Google</ORGANIZATION>. Incorrect: <ORGANIZATION>at Google</ORGANIZATION>.
  • Example: <ORGANIZATION>The New York Times</ORGANIZATION> should keep The when it belongs to the established name.
• Bare type-words and category descriptors, meaning words that name a category rather than a specific entity, should not be labeled on their own. Examples include "person", "company", "organization", "team", "hospital", "city", "product", "book", and "quantity".
  • Example: The company hired 200 people. should leave company unlabeled.
  • Example: She joined <ORGANIZATION>Acme Corp</ORGANIZATION>. should still label the named organization.
  • Example: The person who called did not leave a name. should leave person unlabeled.
• A multi-word name should be one span when it forms one real named entity; separate entities should be labeled separately.
  • Example: <PERSON>Barack Obama</PERSON> should be one span.
  • Example: <PERSON>Barack</PERSON> and <PERSON>Michelle</PERSON> should be two spans.
• Coordinated names should be separate spans, unless the conjunction is part of one established name.
  • Example: <ORGANIZATION>Google</ORGANIZATION> and <ORGANIZATION>Microsoft</ORGANIZATION> should be two spans.
  • Example: <ORGANIZATION>Johnson & Johnson</ORGANIZATION> should stay one span because & belongs to the name.
• Ordinary punctuation and stray whitespace should stay outside spans, unless the punctuation belongs to the name or abbreviation.
  • Example: He moved to <LOCATION>Berlin</LOCATION>. should leave the final period outside the span.
  • Example: <ORGANIZATION>AT&T</ORGANIZATION> should keep & inside the span.
• Possessive markers should stay outside the span unless the full possessive form is the name.
  • Example: <PERSON>Maria</PERSON>'s laptop should leave 's outside the span.
• Labels should follow context, not surface form alone. The same words may be an organization in one sentence and a place in another.
  • Example: She works at <ORGANIZATION>Cambridge University</ORGANIZATION>.
  • Example: The conference was held in <LOCATION>Cambridge</LOCATION>.
• Quantities used as time should be labeled as TIMEDATE, not AMOUNT.
  • Example: She bought <AMOUNT>50</AMOUNT> tickets.
  • Example: The train arrives in <TIMEDATE>50 minutes</TIMEDATE>.

---

Governing principles

The rules below were written one at a time, each closing a specific gap. Written out afterwards, they turn out to rest on three principles rather than five unrelated judgements. Naming the principles matters because no policy can enumerate every case: a labeler who meets a situation none of D1–D5 covers should be able to reach the same answer this policy would, by applying these.

P1 — Extent: a span carries exactly the value the text states. Whatever changes the stated value belongs inside the span; whatever only attaches the span to the rest of the sentence stays outside. 300 and $300m are different values, so the symbol is in. at Google and Google denote the same organization, so the preposition is out. Derives: D2, D4.

P2 — Label choice: prefer the stable reading over the local one. Where a span could take two labels, choose the one that reflects what the words mean generally, not what this particular writer is using them for on this occasion. Guitarist means an occupation in every English sentence; its use as one reviewer's stand-in name exists only in that review. The stable reading generalises to unseen text, which is what the dataset is for. Derives: D1, D3.

P3 — Threshold: a span earns a label only if it does the work that label's definition describes. TIMEDATE is defined as an expression that places something on a timeline, so an expression that places nothing does not qualify, however time-flavoured its vocabulary. This is the same principle the baseline rules already use when they exclude bare category words: city does not pick out a place, so it is not a LOCATION. Derives: D5, and it is the principle behind the baseline rule on type-words.

Order of application. When more than one principle bears on a span, apply them in this order: P3 first (does this span qualify for any label?), then P2 (which label?), then P1 (how far does the span reach?). Existence, identity, extent. The order matters because extending a span under P1 is wasted work if P3 would have rejected it, and because P2 can change which label's definition P1 has to satisfy.

What these principles do not settle. They do not adjudicate between two labels whose definitions both genuinely fit — PRODUCT and WORKOFART for a video game, or ORGANIZATION and LOCATION for a named institution standing in for its premises. P2 assumes one reading is more stable than the other; where both are equally stable, a separate rule is needed.

---

D1. PERSON covers naming expressions, including nicknames — not descriptions

Status: SETTLED

Gap. The PERSON label is defined as "a named person, including given names and surnames." Two things fall outside that wording: nicknames (neither a given name nor a surname) and descriptive references to a person who is never named (The 36-year-old). The label definition settles neither.

Decision. Tag a span as PERSON when it is an expression used to name a specific person — given name, surname, full name, or nickname/alias. Do not tag a phrase that merely describes or refers to a person without naming them, even when the person is real and specific.

Examples.

• Tag: record 1, <PERSON>Pac-Man</PERSON> has released own entrance song — Pac-Man is Manny Pacquiao's boxing nickname, not the arcade game.
• Do not tag: record 1, The 36-year-old also directed the music video — refers to Pacquiao, but describes rather than names him.
• Do not tag: record 76, a 34-year-old communications and public...; record 62, the woman at the reception desk; record 45, 19-year-old has attracted interest.

Reasoning. A nickname is one of the ways a person gets named, so excluding it would drop a genuine person mention on a technicality of wording. A description is not a name — it identifies by attribute, and the attribute would apply to anyone else who shared it.

Consistency check.

• Consistent with the baseline rule that labels follow context, not surface form: the same string Pac-Man would be a PRODUCT in a sentence about the arcade game.
• Consistent with the baseline rule that bare type-words and category descriptors are not labeled on their own. That rule's own example, The person who called did not leave a name, has the same shape as The 36-year-old, so tagging descriptions would have contradicted a baseline rule that cannot be edited.
• Does not touch JOB. Occupational titles such as Chief Investment Officer (record 57) stay JOB.

Alternative considered and rejected. A middle position — tag a description when it is capitalised and functions as a title — was tested against the data and dropped for two reasons. First, capitalisation in English is confounded by sentence position: Teenager (record 26), Patient (record 77) and The 36-year-old (record 1) are capitalised only because they begin a sentence, while the woman and a 34-year-old are lowercase only because they do not. The rule would have split identical cases on an irrelevant signal. Second, "functions as a title" pulls Chief Investment Officer and Chief Medical Officer into PERSON, where they collide with JOB.

Scope. Around 10 spans affected directly; roughly 18 descriptive phrases in the starter set are confirmed as correctly untagged by this rule.

Still open. Record 5 tags invented nicknames for anonymous performers (MC, Guitarist, Dreads, Blank, Mess Up). These are naming expressions by D1, but Guitarist is also an occupation and MC is arguably one. Needs a PERSON vs JOB tiebreak — likely a separate rule.

---

D2. A quantity span carries its whole unit of measure

Status: SETTLED

Gap. The baseline rules say a span covers the entity and not the surrounding grammar, but they do not say what counts as part of a quantity. A bare number is not a quantity — 300 on its own says nothing about money, length or proportion. The starter data is incoherent here: $300m is tagged 300m with the symbol excluded, 2.5million keeps its suffix, and 3 feet 1 inches is split into two bare numbers.

Decision. An AMOUNT or TIMEDATE span covers the number together with everything that states what is being measured:

1. Currency symbols precede the number and are inside the span: <AMOUNT>$300m</AMOUNT>, <AMOUNT>£1,000</AMOUNT>.
2. Magnitude suffixes are inside the span: the m of $300m, the million of £1million.
3. Unit words are inside the span, whether symbolic or spelled out: <AMOUNT>20 percent</AMOUNT>, not <AMOUNT>20</AMOUNT> percent.
4. A compound measurement is one span, not one span per number: <AMOUNT>3 feet 1 inches</AMOUNT> is a single height, not two quantities.
5. Where a currency symbol is separated from its number by whitespace, the span begins at the symbol and the internal whitespace is part of the span. It is not stray whitespace, because it falls between two parts of one quantity: <AMOUNT>$ 784,700.78</AMOUNT>.

Examples.

• Record 1: faces Floyd Mayweather in <AMOUNT>$300m</AMOUNT> showdown
• Record 60: power losses reach <AMOUNT>as high as 20 percent</AMOUNT> → span is 20 percent (see D4 for the approximator)
• Record 30: who stands <AMOUNT>3 feet 1 inches</AMOUNT>
• Record 33: increases by about <AMOUNT>$ 784,700.78</AMOUNT>

Reasoning. The span is supposed to carry the value. 300 and $300m are not the same value, and 20 and 20 percent are not the same value. Cutting the unit off destroys exactly the information the label exists to capture. Clause 4 follows from the same logic: 3 and 1 are meaningless apart, and only 3 feet 1 inches names an actual height.

Consistency check.

• Does not conflict with the baseline rule that function words stay outside spans: a currency symbol, a suffix and a unit word are none of those. They are the substance of the quantity, not grammar attached to it.
• Clause 5 is an explicit carve-out from the baseline rule that stray whitespace stays outside spans. The carve-out is narrow: it applies only to whitespace interior to a single quantity, never to leading or trailing whitespace, so it does not weaken the baseline rule anywhere else.
• Consistent with the baseline rule that quantities used as time are TIMEDATE, not AMOUNT. D2 governs the extent of a span, not the choice of label, so <TIMEDATE>2 hours</TIMEDATE> keeps its unit for the same reason <AMOUNT>20 percent</AMOUNT> does.

Scope. 16 numeric spans in the starter set touch this rule: 10 with a currency symbol left outside, 4 with a unit word left outside, 2 forming a split compound measurement.

Clause 1a — currency qualifiers. A country prefix on a currency symbol is part of the symbol and stays inside the span: <AMOUNT>US$580,000</AMOUNT> (record 81). US$ and $ do not denote the same currency, so dropping the prefix changes the stated value in exactly the way clause 1 exists to prevent.

---

D3. When a naming expression is itself an occupational word, JOB wins

Status: SETTLED

Gap. D1 says a nickname used to name a specific person is PERSON. But a writer can nickname someone by their occupation, at which point the same span satisfies both PERSON (it names this person) and JOB (it is an occupational title). Neither the baseline rules nor D1 says which label takes precedence.

Decision. Where a naming expression is an ordinary occupational term, label it JOB. PERSON is reserved for naming expressions that are not occupational words.

Example. Record 5, a review by someone who never learned the performers' names: here are my nicknames to them: MC: loud blond guy... Guitarist: great player... Dreads: the only guy that can dance... Blank: no expression in any act. Under this rule the list splits by what each word means: <JOB>MC</JOB>, <JOB>Guitarist</JOB>, but <PERSON>Dreads</PERSON>, <PERSON>Blank</PERSON>, <PERSON>Mess Up</PERSON> — the last three are invented nicknames with no occupational content.

Reasoning. The occupational reading is the stable one. Guitarist means the same thing in every English sentence it appears in; its use as somebody's stand-in name is local to this one review and disappears the moment the writer learns the real name. Labeling by the stable meaning generalises; labeling by the local improvisation does not, and would teach the model that any capitalised word before a colon is a person.

Supporting evidence from the dataset. Record 93 already labels The <JOB>M.C.</JOB> of Zumanity as JOB, alongside <JOB>drag queen</JOB> and <JOB>stripper</JOB>. Applying JOB to record 5's MC makes the two records agree rather than contradict each other.

Consistency check.

• Consistent with D1, which it narrows rather than reverses: D1 still governs every naming expression that is not an occupational word.
• Consistent with the baseline rule that labels follow context, not surface form.
• Consistent with the baseline rule against bare category descriptors, since JOB is defined as applying to an occupational title "when it functions as such in the sentence."

Consequence accepted. Record 5's list of nicknames ends up with mixed labels rather than one label for the whole list. This is intended: the list is a list of people described in different ways, not a single entity type.

---

D4. Approximators are inside the quantity span

Status: SETTLED

Gap. D2 settles where a quantity span ends. It does not settle where the span starts when the number is hedged. The baseline rules do not say whether a word like about belongs inside the span, and the task itself flags this as an open question using the phrase "about 50 tickets."

Decision. A word or short phrase that marks the number as approximate or as a bound is part of the span. Tag <AMOUNT>around 55,000</AMOUNT>, not around <AMOUNT>55,000</AMOUNT>.

This covers: about, around, roughly, approximately, nearly, almost, at least, more than, less than, up to, as high as, as much as, over, under.

Examples.

• Record 85: has to remove <AMOUNT>around 55,000</AMOUNT> fatbergs a month
• Record 85: This costs <AMOUNT>more than £1million</AMOUNT> a month
• Record 60: power losses reach <AMOUNT>as high as 20 percent</AMOUNT>
• Record 63: they said it would be <TIMEDATE>about 2 hours</TIMEDATE> until we go

Reasoning. around 55,000 is an estimate and 55,000 is an exact count. Dropping the approximator turns an estimate into a precise figure the text never claimed. This is the same reasoning as D2: the span should carry the value the text actually states, and an approximator changes that value.

Consistency check.

• Extends D2 rather than competing with it. D2 governs the right edge of a quantity span, D4 the left edge, and both are justified by the same principle: the span carries the stated value.
• Tension with the baseline rule that prepositions and conjunctions stay outside spans, since more than and up to contain function words. The tension is only apparent: the baseline rule excludes function words that attach the entity to the rest of the sentence, as in at Google. Here more than does not attach the number to anything — it modifies the number's own value. The test is whether removing the word changes what the number means. Removing at from at Google does not change which organization is meant; removing more than from more than £1million does change the amount.
• Does not affect non-numeric labels.

Scope. 10 numeric spans in the starter set are preceded by an approximator: 5 AMOUNT and 5 TIMEDATE.

Boundary of the rule. Only the approximator itself joins the span, not a longer hedging clause. power losses reach as high as 20 percent takes as high as 20 percent; a phrasing such as losses that some engineers estimate at 20 percent takes only 20 percent, because the estimating clause is ordinary sentence grammar rather than a modifier on the number.

---

D5. TIMEDATE requires a definite region of time

Status: SETTLED

Gap. TIMEDATE is defined as "an expression that places something on a timeline, including dates, clock times, and durations used as time." That wording does not say what to do with time words that place nothing in particular: always, never, future, moment. The starter data tags all of them, and TIMEDATE is the largest label in the set at 164 spans, so the gap is expensive.

Decision. Tag a span as TIMEDATE only if it marks a definite region of the timeline. Three shapes qualify:

1. A point — May 2, 10:30, 1949, December 31, 2019
2. A stretch — 2 hours, 18 months, three days, past few decades
3. A repeating pattern — every Saturday, annually, twice a day

Do not tag:

4. Unbounded expressions, which have no edges: always (all of time), never (none of it), future (starts now, runs forever), ever, yet

5. Bare time-category words: moment, month, hour, week, night

6. Expressions with no time content at all: clash, meals, market downturns, meeting with the board

7. Event-anchored expressions. An expression that locates something relative to a human, political or situational event, rather than to the calendar or the clock, is not a TIMEDATE: halfway through (an audition), during production hours, the past six presidential elections, during a meeting with the board, since I was in junior high school.

The test. Could you mark it on a calendar? May 2 is one square, 2 hours is a short stretch, every Saturday is every seventh square. future would mean shading everything after today with no end, always the whole calendar, never nothing, and moment cannot be marked at all.

Stated as a single question for clause 6: strip the sentence away — does the phrase still rely on a clock, a calendar, or a standard unit of time? every Thursday at 5 PM does. halfway through, during halftime and during elections do not; they need the duration or timing of some other event supplied from outside before they mean anything.

Clause 6 carve-out — named calendar observances are kept. Easter, Ramadan and Remembrance Day are named after events but hold a determinate position on a calendar: 11 November, a computed date, a lunar month. The test is whether the phrase has a calendar position, not whether it is named after an event. a meeting with the board has no calendar position and is rejected; Remembrance Day has one and is kept. Without this carve-out clause 6 would strip three correct spans, including one this policy itself assigned when resolving the out-of-schema COMPANY label on record 39.

No fallback label exists. The schema is fixed at eight labels and contains no EVENT tag, so spans rejected under clause 6 are left untagged rather than relabeled. This is deliberate: inventing a home for them would reintroduce exactly the out-of-schema problem that COMPANY created in the starter data.

Reasoning. The label's own definition says the expression must place something on a timeline. Placing means narrowing down. An expression that covers all of time, none of it, or an edgeless region does not narrow anything, so it fails the definition on its own terms rather than by an added restriction.

Repeating patterns are kept (clause 3) because dropping them produces an incoherent result: Friday would be tagged while every Saturday would not, even though every Saturday carries strictly more information about when something happens.

Consistency check.

• Clause 5 is the baseline rule on bare type-words applied to time. moment is to time what city is to place, and the baseline already leaves city unlabeled.
• Clause 3 does not rescue every time or some time, because time is a generic category word rather than a calendar unit. Those stay rejected under clause 5, so clauses 3 and 5 do not fight.
• Consistent with D2 and D4, which govern span extent rather than label choice. A kept duration still carries its unit and any approximator: <TIMEDATE>about 2 hours</TIMEDATE>.

Earlier formulation, rejected. The first version of this rule asked whether the span was "built on a calendar or clock unit." It was dropped because it does not actually exclude future or always — both are time words, so the test sounds discriminating but admits the very cases the rule exists to reject. The bounded/unbounded test replaces it and draws the line where the label definition already draws it.

Scope. A checker script (scripts/check_d5_timedate.py) flags 49 of the 164 TIMEDATE spans for review: 26 unbounded, 11 bare category words, 12 with no time content. The script proposes only; each flagged span is confirmed by hand, since the reject lists are hand-built and the third bucket depends on context the script cannot see.

---

D6. PRODUCT vs WORKOFART: consumed content vs. things used to make or do

Status: SETTLED

Gap. PRODUCT is "a named commercial product, device, or branded good." WORKOFART is "a named creative or published work." Software satisfies both definitions at once: a video game is authored like a film and sold like a device. The governing principles do not resolve this — P2 assumes one reading is more stable than the other, and here both readings are equally stable. The starter set shows the consequence: Minecraft, Final Fantasy VII and world of warcraft sit in PRODUCT while php, discord and Apache Impala sit in WORKOFART, with no visible reason for the split.

Decision. Ask what the named thing is for:

• If it is consumed — played, read, watched, listened to — it is WORKOFART.
• If it is used to make or do something else — a tool, a language, a framework, a utility, a service — it is PRODUCT.

Examples.

• <WORKOFART>Minecraft</WORKOFART>, <WORKOFART>Final Fantasy VII</WORKOFART>, <WORKOFART>world of warcraft</WORKOFART> — played.
• <PRODUCT>Godot Engine</PRODUCT> — games are built with it.
• <PRODUCT>Java</PRODUCT>, <PRODUCT>SQL</PRODUCT>, <PRODUCT>php</PRODUCT>, <PRODUCT>Django</PRODUCT>, <PRODUCT>Flask</PRODUCT> — languages and frameworks.
• <PRODUCT>Illustrator</PRODUCT>, <PRODUCT>Gimp</PRODUCT>, <PRODUCT>Sketch</PRODUCT> — art is made with them.
• <PRODUCT>discord</PRODUCT> — a communication tool.

Reasoning. The test separates the one pair that would otherwise be indistinguishable: Minecraft and Godot Engine are both software about building things, but you play one and you build games with the other. Asking what the thing is for gives a single question that a labeler can answer in seconds without knowing anything about how the software was made or sold, which is what makes it survive 900 records in Stage 2.

It also decides non-software cases the starter set gets wrong, without needing a separate rule: a fly swatter (The Executioner Bug Zapper Racket, currently WORKOFART) is used to do something, so it is PRODUCT; a TV soap opera (Hollyoaks, currently PRODUCT) is watched, so it is WORKOFART.

Consistency check.

• Sits underneath P2 rather than contradicting it. P2's stated limit is that it cannot arbitrate two equally stable readings; D6 is the separate rule that limit calls for.
• Applied after P3 and before P1, per the order of application: the span must first qualify for a label at all, then D6 picks between these two, then extent is settled.
• Does not disturb ORGANIZATION. Adobe the company stays ORGANIZATION even though Illustrator is a PRODUCT.

Accepted awkward case. D&D moves from PRODUCT to WORKOFART. It is sold as physical boxed books, which reads oddly as a "work of art," but it is authored content that people play, and carving out an exception for physical packaging would reintroduce exactly the arbitrary split this rule exists to remove. The medium is not the test; the use is.

Scope. Around 30 software and game spans in the starter set fall under this rule. Roughly 7 need relabeling — Minecraft, Final Fantasy VII, world of warcraft, D&D move to WORKOFART; php, discord, Apache Impala move to PRODUCT. The remainder were already on the correct side but were correct by accident; D6 is what makes them correct on purpose, and is what will keep Stage 2 consistent with Stage 1.

---

D7. ORGANIZATION vs LOCATION: decide by the predicate, not the name

Status: SETTLED

Gap. A named business can stand in for its premises. The baseline settles the Cambridge University / Cambridge case, where two different surface forms name an institution and a place, but not the case where the same name is used both ways. The starter set shows the damage: McDonald's appears under both labels, blizzard entertainment is tagged LOCATION in a sentence about corporate discretion, and Wal-Mart is tagged LOCATION in a sentence about selling goods.

Decision. The label follows what the sentence predicates of the name, not the name itself. Applied in this order:

1. Agency test → ORGANIZATION. The entity acts, decides, employs, sells, announces, or is otherwise treated as an actor. <ORGANIZATION>blizzard entertainment</ORGANIZATION> may, in its sole and absolute discretion, take whatever action... (record 25); The products are sold by <ORGANIZATION>Wal-Mart</ORGANIZATION> and <ORGANIZATION>Kroger</ORGANIZATION> (record 50); The <ORGANIZATION>Pentagon</ORGANIZATION> lifted a ban (record 58).

2. Physical-constraint test → LOCATION. The entity is described with a predicate or modifier that can only apply to a physical space — leaking, painted, spacious, dirty, walking into, sitting inside, or being explicitly called a place. Advertisements portraying <LOCATION>McDonald's</LOCATION> as a happy, circus-like place (record 69).

3. Indefinite-article test → LOCATION. A company name preceded by a/an, or pluralised, denotes one branch rather than the corporation. There is no such thing as "a corporation" in the singular-instance sense, so the article itself is the evidence. the 'glamour' of being in a <LOCATION>McDonald's</LOCATION> (record 69); The ceiling looks like a <LOCATION>chipotle</LOCATION> (record 4).

4. Default → ORGANIZATION. Where no test fires — a bare mention with no predicate attached — label ORGANIZATION. A company name names a company unless something in the sentence pushes it toward the premises. Parking is terrible (same as <ORGANIZATION>Sahara</ORGANIZATION>) (record 4).

Explicit exclusion — locative prepositions do not trigger LOCATION. at, in, from and similar are not evidence on their own. The baseline's own example is He works at <ORGANIZATION>Google</ORGANIZATION>, so a preposition-based test would contradict a rule that cannot be edited. Only the article flips in a McDonald's to LOCATION, not the in. This exclusion is written down because it is the obvious-looking extension that a later labeler would otherwise add.

Reasoning. Metonymy is a property of the sentence, not of the word, so the sentence is where the evidence has to come from. Predicates are also observable: a labeler can point at the verb that decided the call, which makes the decision reviewable by someone else. "Is this sentence about the place or the business?" is the same question but with nothing to point at, and that is what would sink agreement.

Validation. Record 69 mentions McDonald's four times and the original annotator switched labels each time. D7 reconstructs all four calls exactly: agency for know and promote, physical for portraying... as a place, article for in a McDonald's. The rule was derived from other records, so this is an out-of-sample check rather than a restatement, and it is evidence that D7 matches how a careful annotator already behaves rather than imposing a new convention.

Consistency check.

• Sits underneath P2 as the second arbitration rule, alongside D6. Where D6 splits two labels by what the thing is for, D7 splits them by what the sentence does with it.
• Consistent with the baseline rule that labels follow context, not surface form — D7 is that rule made operational.
• Does not disturb genuine place names. Berlin, Las Vegas and Brazil never reach D7, because they are not organization names in the first place.

Surface form rejected as a signal, for the second time. Capitalisation was tested here as a possible ORGANIZATION marker and rejected on two grounds. It fails empirically: 11 of the 100 starter records are entirely lowercase (web chat, scraped text), including record 25, which is why blizzard entertainment looks uncapitalised. It also fails in principle: Google and Berlin are both capitalised, so capitalisation is weak evidence for P3 (does this qualify for any label?) and no evidence at all for P2 (which label?). The same signal was considered and rejected once before, in D1. The policy relies on function and context throughout, never on surface form.

Scope. Around 12 spans in the starter set change under D7, all of them LOCATION mentions of named businesses that fail the physical and article tests: blizzard entertainment, Wal-Mart, Kroger, Sahara, cosmopolitans, and the restaurant names among Black Bear Diner, Joyride Taco House and Andre's, pending a per-record check of their predicates.

Still open. Built infrastructure — Carajas railroad, Tucurui dam, Flatley Dam, Beltway — is currently LOCATION. These are structures rather than businesses, so D7 does not reach them and they need a separate decision.

---

D8. Model, version and vintage designators are part of the PRODUCT span

Status: SETTLED

Gap. A product name can contain a number or a year that would, standing alone, qualify as AMOUNT or TIMEDATE. Nothing in the baseline or in D2–D5 says whether such a token is a separate entity or part of the product name. Record 79 is the case: Items selected: 2012 Dom Perignon Brut, Champagne.

Decision. A number, year or version designator that identifies which product is meant belongs inside the PRODUCT span. It is not tagged separately as TIMEDATE or AMOUNT. <PRODUCT>2012 Dom Perignon Brut</PRODUCT>, not <TIMEDATE>2012</TIMEDATE> <PRODUCT>Dom Perignon Brut</PRODUCT>.

Reasoning. Three arguments, the third decisive.

1. Identity. A vintage year does not anchor anything on a timeline; it alters the identity, price and composition of the product. 2012 Dom Perignon and 2008 Dom Perignon are different goods.
2. Consistency with existing product names. The 15 in iPhone 15 and the 747 in Boeing 747 are already treated as part of the name. Boeing 747 appears in the starter set as a single PRODUCT span, so treating 2012 as separable would make record 79 inconsistent with a record the annotator got right.
3. Semantic distortion. Tagging 2012 as TIMEDATE asserts that the selecting happened in 2012. The sentence says the selection is happening now and the bottle is from 2012. A separate TIMEDATE span would put a claim into the annotation that the text does not make — a worse failure than a misfiled span, because a model trained on it learns to date events by product vintages.

Consistency check.

• Follows P1: the span carries the stated value, and the stated value here is which bottle, not when anything happened.
• Follows the P3 → P2 → P1 order: 2012 is first tested for whether it qualifies as a time expression at all in this context, and under argument 3 it does not, so the question never reaches D5.
• Does not weaken D5. A year that genuinely places an event on a timeline is still TIMEDATE; D8 applies only where the year identifies a product.
• Consistent with existing spans such as 5 Puttonyos Royal Tokaji and M282 warhead, already tagged as whole PRODUCT spans in the starter set.

Boundary. D8 covers designators that are part of the product's identity, not quantities of it. two Boeing 747s is <AMOUNT>two</AMOUNT> <PRODUCT>Boeing 747</PRODUCT>s — the two counts units and is not part of the name.

---

D9. JOB is a role a person holds — not a field, not a unit of a company

Status: SETTLED

Gap. JOB is defined as "an occupational title or formal work role when it functions as such in the sentence." The starter set stretches this in two directions the definition does not license: to industries and academic fields (the energy sector of the Kyrgyz Republic, linguistics, military), and to parts of an organization (analytics team, Security, Geriatric Medicine Department). Neither is a role a person holds. JOB is the second-largest label at 100 spans, so the gap is expensive.

Decision, in three parts.

1. A field, industry or discipline is not a JOB, and gets no label. finance, linguistics, psychology, market research, SCM, energy sector, Research sector, military, acting, Voice-over. Nobody holds linguistics as a post. The schema has no INDUSTRY label and none is added, for the same reason COMPANY was removed: an out-of-schema label makes the dataset incomparable across stages and would be too sparse to learn from.

2. A unit of an organization is not a JOB. Whether it takes ORGANIZATION instead depends on the baseline's own wording, which bans category descriptors on their own:

   • Bare functional noun → no label. Security, Quality, Assurance, Emergency Operations, emergency response, costume team. These name a function, not a body.
   • A modifier that picks out one specific unit → ORGANIZATION. Geriatric Medicine Department, market analysis division, analytics team, primary care department are identifiable units within a named parent organization, in the way Acme Corp is identifiable and company is not.

3. Plural and unattached occupations stay JOB. bricklayers, statisticians, researchers, police officers are occupational titles even when no individual holds them in the sentence, because the label definition turns on whether the phrase functions as an occupational title, not on whether a named person is attached.

Examples.

• engagement in the energy sector of the Kyrgyz Republic → energy sector untagged (record 81)
• We've outsourced Security to an external team → Security untagged (record 35)
• <JOB>Chief Medical Officer</JOB> <PERSON>Robert Williams</PERSON> from the <ORGANIZATION>Geriatric Medicine Department</ORGANIZATION> (record 77)
• 5.7% of the <JOB>bricklayers</JOB> in the relevant labor force (record 84)

Reasoning. Part 1 follows P3: JOB's definition names a role, and a field is not a role, so the span does not do the work the label describes. Part 2 keeps the distinction the baseline already draws between Acme Corp and company, applied one level down to units inside a company. Part 3 resists over-correction: the temptation after parts 1 and 2 is to strip every occupation without a named holder, but that would delete ~15 correct spans and contradict the label's own wording.

Consistency check.

• Does not disturb D3, which decides PERSON vs JOB for occupational nicknames. D9 decides what counts as an occupation at all, so it applies first under the P3 → P2 → P1 order.
• Part 2 does not conflict with D7. D7 chooses between ORGANIZATION and LOCATION for something already established as an organization; D9 part 2 establishes whether a unit is one.
• No surface-form test is used. The named/bare split turns on whether a modifier picks out a specific unit, not on capitalisation — consistent with the rejection of capitalisation in D1 and D7.

Scope. Of 100 JOB spans: ~12 industries and fields lose their label, ~6 bare functional nouns lose theirs, ~4 named units move to ORGANIZATION, 2 have boundary errors (Business Intelligence Director at swallowed a preposition; front-line combat positions names a category of posts rather than a title), and the remaining ~76 are correct and unchanged.

---

D10. Shortened forms are tagged when the full form appears in the same record

Status: SETTLED

Gap. The baseline bans bare category words: company, hospital, bank. But a text that has already named an institution often refers to it afterwards by a shortened form that is, on its surface, exactly such a category word. Record 99 is a World Bank report that names World Bank and World Bank Group, then says the Bank and CVRD failed to anticipate.... Read without context, Bank is a category word and gets no label; read in context, it is a specific institution.

Decision. Tag a shortened form, abbreviation or acronym with the label its full form carries, provided the full form appears in the same record. Where the full form does not appear, the ordinary rules apply and a bare category word stays unlabeled.

Examples.

• Record 99: <ORGANIZATION>World Bank</ORGANIZATION> appears, so the <ORGANIZATION>Bank</ORGANIZATION> and <ORGANIZATION>CVRD</ORGANIZATION> failed to anticipate is tagged.
• Record 93: <ORGANIZATION>Cirque du Soliel</ORGANIZATION> appears, so the later another <ORGANIZATION>Cirque</ORGANIZATION> show is tagged.
• Record 99: Project Performance Audit Report (PPAR) introduces the acronym, so PPAR carries the same label as its expansion.
• Counter-example: a text that mentions no named bank leaves the bank unlabeled.

Reasoning. The shortened form refers to the same entity as the full form, so refusing to tag it would drop a real mention on a surface technicality — the same reasoning that admitted nicknames under D1. Requiring the full form to be present in the record keeps the rule decidable: the labeler checks the record rather than guessing at what the writer had in mind, and the condition is mechanically verifiable by script.

The same-record condition also blocks the failure this rule would otherwise create. Without it, bank, company and department would become taggable anywhere, which would contradict the baseline rule on category words and teach the model to tag every generic occurrence.

Consistency check.

• Narrows nothing in the baseline: the category-word rule still governs every record where no full form is present.
• Consistent with D1's treatment of nicknames — both admit an alternative expression that refers to an already-established entity.
• Consistent with D9, which strips bare functional nouns such as Security and Quality. Those records contain no full form for them to abbreviate, so D10 does not rescue them.

Scope. 3 spans in the starter set: Bank (record 99), Cirque (record 93), PPAR (record 99).

---

R1. Two records removed

Status: SETTLED

Two of the 100 starter records are removed. Both are spam-corpus artifacts whose text is not English prose, so their annotations cannot be corrected into anything meaningful — the problem is the source text, not the labels.

Record 92 — machine-generated essay-mill spam. Pigment must saltate after the sadhu. Uprighteously brayon riche was the loathsomely caucasian gillyflower. Grammatically shaped, semantically empty. Its 9 spans are incidental collisions between generated tokens and entity-looking strings.

Record 80 — SEO keyword spam. ...bp pot com mlqrmsi u vqnnwgzgobi aaaaaaa imrs php src washingtonpost answer sheet files jfk app jpg w... A bag of search terms with no sentence structure. php and washingtonpost are keyword fragments rather than mentions of a language or a newspaper. This policy had already relabeled php to PRODUCT under D6 — correct by the rule and meaningless in context, which is the clearest argument that the record itself is the problem.

Reasoning. A token-classification model learns where entities sit relative to surrounding words. Text with no syntax offers no such relationship, so these records contribute noise to training and would make test metrics harder to interpret. The task permits removals that would otherwise worsen model behaviour.

Record 59 was considered and kept. It is also spam — lowercase pharmaceutical ad copy that degrades into word salad after roughly 380 characters. It was kept because its opening is coherent English carrying genuine entities (cialis, cialis soft tabs, 15 minutes, up to 36 hours), and because lowercase unpunctuated web text is a real register the model must handle: 11 of the 100 records are entirely lowercase. Removing every noisy record would bias the corpus toward clean newswire and hurt generalisation.

Effect. 100 records to 98, a 2% removal.

---

Observed patterns (not yet rules)

• Vague nouns tagged TIMEDATE. Words that name a category of time without placing anything on a timeline: moment (rec 0), clash (rec 1), plus now (x4), never (x3), always (x2), future (x2), today (x2) across the set. TIMEDATE is the largest label at 164 spans.
• Multi-word names split into one span per word. Manny/Pacquiao, Floyd/Mayweather, blizzard/entertainment. 45 adjacent same-label pairs found by script, but not all are errors — record 23's list of game engines is correctly separate. Script proposes, human decides.
• A label outside the schema. COMPANY appears 8 times on values that are variously a person, a date, and a product. A corruption bucket, not a consistent mislabel.

---

Recurring methodological finding: scripts propose, humans dispose

Every automated pass built during Stage 1 produced false positives that only a human reader could reject. Four instances, all from this dataset:

1. A punctuation-at-edge detector flagged 6 spans. Only 2 were errors (country., movie.); the other 4 — U.N., U.S., 10:00 p.m., M.C. — end in a period legitimately, which the baseline rules explicitly allow.
2. An adjacent-same-label detector found 45 candidate split names. Many were real (Manny/Pacquiao), but record 23's list of game engines is correctly separate, so a blanket merge would have corrupted the data.
3. A person-descriptor search matched Pac-Man as "Man", plus man-days and father-confessor.
4. The first version of the D5 checker missed plural calendar units and wrongly flagged three days, two weeks and past few decades as having no time content. Fixed by allowing plural forms; flagged count fell from 62 to 49.

Only 5 of the 100 records contain an error a script can identify with certainty on its own — an out-of-range offset, a value that disagrees with the text, or an overlapping span. Everything else, which is the large majority of the damage, is a wrong label or a wrong extent, and both are invisible to a script without a policy to check against. This is the reason the policy had to be written before the mass edit rather than after it.

---

Stage 2 policy extensions

The Stage 1 policy (POLICY.md) was written against 100 records drawn mostly from news, reviews, chat, and legal/institutional text. Stage 2 pulls from Wikipedia, Wikinews, Project Gutenberg, and US Congressional bill text (billsum) — registers the starter set barely touched. A 20-record pilot label across all four sources, done by hand before any bulk automation, surfaced four gaps the existing rules do not settle. These extend the policy; none contradict it. Numbered to follow on from D1–D11.

---

D12. Named wars and historical eras with a determinate calendar position are TIMEDATE; undated named events are not

Gap. D5.6 already carves out named calendar observances (Easter, Remembrance Day) from the general rule that event-anchored expressions get no label, on the ground that they hold a determinate position on the calendar. Historical text constantly produces two more shapes: named wars/eras (the First World War, the Cultural Revolution) and named one-off battles (the Battle of Trafalgar). Both are event-anchored by D5.6's own test, but they differ from each other in exactly the property D5.6's carve-out turns on.

Decision.

• A named war or historical era with a well-known, determinate date range is TIMEDATE, by the same reasoning as the D5.6 carve-out: served in the <ORGANIZATION>Royal Navy</ORGANIZATION> during the <TIMEDATE>First World War</TIMEDATE>.
• A named battle or one-off historical event with no calendar position stated or implied in the record stays untagged, consistent with D5.6's general rule: served at the Battle of Trafalgar — Battle of Trafalgar untagged.

Reasoning. the First World War picks out a specific, widely-known span of years the way Easter picks out a specific day; the Battle of Trafalgar picks out an event the way a meeting with the board does — real, but not placed on a timeline without outside knowledge the text does not supply. This is D5.6 applied, not a new principle.

---

D13. A named piece of legislation or a named formal standard is WORKOFART

Gap. Government-report text is dense with named Acts (the Water Resources Development Act of 1986) and Wikipedia occasionally names formal standards (the International Phonetic Alphabet). Neither is authored fiction, but both are titled, published documents, which is what WORKOFART's definition ("a named creative or published work ... or titled publication") actually asks for — the baseline's own examples already include non-fiction publications (The New York Times, PPAR in the World Bank report, record 99 in Stage 1).

Decision. A named Act, statute, treaty, or formal standard is WORKOFART: <WORKOFART>the Water Resources Development Act of 1986</WORKOFART>, <WORKOFART>the International Phonetic Alphabet</WORKOFART>.

Reasoning. These are titled publications with a fixed, citable name, which is exactly what distinguishes a WORKOFART from a bare description. They are not PRODUCT under D6, since nobody "uses" a statute to make or do something else — it is read and cited, the same relationship a reader has to a book.

Consistency check with Stage 1. Applying D13 to record 99 exposes an inconsistency Stage 1 missed: Project Performance Audit Report was tagged PRODUCT while its own acronym PPAR, introduced two words later in the same sentence, was tagged WORKOFART — violating D10 (a shortened form carries its full form's label) with itself. D13 settles the correct label as WORKOFART for both. This is corrected in Stage 1's dataset as a one-span retag, noted in the Stage 2 report as a fix carried backward once discovered, per the task's allowance to revise policy after Stage 1 as new edge cases appear.

---

D14. A named battle honour composite (place + year) is tagged as its parts, not as one span

Gap. Military-history text sometimes lists a place and year together as a single conventional designation (Walcheren 1809) rather than as a grammatical date attached to a place (Berlin, 1945).

Decision. Tag the parts separately under their own rules rather than inventing a composite span: <LOCATION>Walcheren</LOCATION> 1809 with 1809 tagged TIMEDATE in the ordinary way. This is not a new mechanism — it is D5 and the baseline location rules applied without modification — stated here only because the convention (no comma, no preposition) could tempt a labeler into treating it as one indivisible unit.

---

D15. A mythological or historical figure referenced by name is PERSON

Gap. Encyclopedic and literary text refers to figures from mythology and antiquity by a single conventional name (Achilles, Zhuang Zhou, Laozi). PERSON's definition ("a named person") does not condition on whether the referent is historically or textually attested versus legendary, and D1 already established that fictional/legendary naming is admitted (Cait Sith, Stage 1) for the same reason.

Decision. Tag by the same test D1 already uses: a naming expression, not whether the referent is historical, legendary, or fictional. <PERSON>Zhuang Zhou</PERSON>, <PERSON>Laozi</PERSON>, the name of <PERSON>Achilles</PERSON> has been used for ... warships (here Achilles names the mythological figure being invoked, not any single ship — no specific ship is named in the sentence, so no PRODUCT span applies).

Consistency check. Directly extends D1's precedent (Cait Sith) rather than introducing a new distinction; no baseline or added rule is narrowed.

---

These five extensions (D12–D15, and the D13 consistency fix to record 99) are folded into the labeling instructions used for Stage 2's bulk annotation and are carried forward into any future stage.