Human Coder Training Guide

Training version: 1.0.0

Status: required preparation for blind human-reliability coding; this guide is not itself a calibration packet, reliability sample, answer key, or adjudication record.

This guide teaches qualified readers to apply the workbench’s annotation method without private project knowledge and without access to accepted decisions. It operationalizes the human reliability architecture for MIPVU identification, lexical boundaries, CMT mapping, interpretation, violence/obligation, agency/absence, confidence, ambiguity, rival readings, uncertainty, and out-of-scope decisions.

All worked examples below are invented for training. They are identified with training-synthetic-* IDs, contain no historical or restricted source text, and must never be included in a blind reliability sample. After training, coders complete the separate multilingual calibration packets, whose coordinator-held answer keys and completion records remain outside the blind study.

Training outcome

Before receiving a reliability packet, a coder should be able to:

preserve blindness and independent decision-making;
identify the exact unit and task layer assigned;
make source-language MIPVU and boundary decisions;
separate identification from CMT and interpretation;
record violence, obligation, agency, and absence without importing a desired theory;
calibrate confidence separately from the categorical decision;
preserve ambiguity, uncertainty, and rival readings;
use out_of_scope rather than guess when competence, context, rights, or the codebook is insufficient; and
explain why a gloss, model output, accepted annotation, or another coder’s judgment cannot substitute for their own source-based decision.

Training completion permits entry into language-specific calibration. It does not qualify a coder by itself and does not authorize access to a blind study packet.

Required reading

Coders must read the versions declared by their cohort:

this training guide;
MIPVU_ANNOTATION_GUIDE.md;
the relevant sections of RESEARCH_DESIGN.md;
the cohort’s codebook and controlled vocabulary;
the case’s rights and storage instructions;
source-language historical-semantics guidance supplied for the cohort; and
the packet-specific instructions, after training and calibration are complete.

Do not open accepted annotation files, model-reliability outputs, prior coder submissions, or adjudication records while coding a blind packet.

Eligibility and source-language competence

A cohort is source-language specific. A coder must have enough reading competence to distinguish contextual meaning, basic meaning, grammar, register, polysemy, idiom, and plausible lexical boundaries in that language. Domain expertise and period-language preparation may also be required.

Qualification is declared before packet release and confirmed through language-appropriate calibration. Familiarity with an English translation is not source-language competence.

Use out_of_scope when:

the item exceeds the coder’s declared language competence;
OCR, transcription, or corruption prevents a defensible decision;
required context or a lawful reference source is unavailable;
source access or transmission is not permitted;
the packet omits a unit needed to answer the assigned question; or
the codebook does not provide an authorized decision path.

An out-of-scope response is information about study coverage. It is not a failure to be hidden and must not be converted into a low-confidence guess.

Gloss limits

An English gloss may help orientation, but it is a separate analytical aid. The source sentence and source lexical unit control the decision.

Do not:

infer a metaphor merely because the gloss uses an English metaphor;
erase a source-language metaphor because the gloss is literal;
use English word boundaries for a French or German boundary decision;
treat one gloss choice as proof of a source or target domain; or
silently repair a disputed gloss while coding.

If the gloss materially influences or conflicts with the source-language reading, record the issue in the uncertainty or rival-reading note. If the source cannot be evaluated without the gloss, use out_of_scope.

Blindness and independence

Blindness protects the study from answer leakage. During reliability coding, the coder may use only the cohort-authorized codebook, historical-semantics resources, lawful dictionaries or reference works, neutral packet context, and their own reasoning.

The coder must not inspect or request:

accepted MIPVU, CMT, interpretive, agency, or absence decisions;
prior Codex-assisted review packets, reliability results, or adjudication;
model packets, model outputs, disagreement reports, or consensus summaries;
another human coder’s decisions, notes, confidence, or completion state;
sample roles such as positive, negative, ambiguous, or claim-relevant;
support ratings, historical conclusions, claim status, or synthesis prose;
the adjudication queue or later adjudicator decisions; or
filenames, comments, or repository history that reveal an expected answer.

Do not discuss live study items with another coder before both submissions are registered. Do not use an AI system to generate, revise, or check study decisions unless the cohort is explicitly defined as assisted coding. An assisted cohort is a different design and cannot be pooled with independent unaided human coding.

If accidental exposure occurs, stop coding and report:

the affected item or range;
what information was exposed;
when and how exposure occurred; and
whether any decisions had already been made.

Do not attempt to compensate mentally or conceal contamination.

Identify the assigned task

Each packet item belongs to one task layer. Answer only the assigned layer.

Layer	Main question	Must not decide
Identification	Is the lexical unit metaphor-related here, and is its proposed boundary valid?	Conceptual cluster, ideological function, historical support
CMT	What source-target mapping and entailments best describe the supplied span?	Whether the project has accepted the metaphor; final interpretation
Interpretation	Which bounded functions, agency roles, and absences are supported?	Historical causation, private mental state, final theory support

Later reasoning must not backfill an earlier layer. A compelling sacrificial interpretation cannot turn a non-metaphorical lexical unit into a MIPVU positive. A MIPVU-positive unit does not require a Koenigsbergian interpretation.

Units and context

Keep these units distinct:

Unit	Coding use
Document	Provenance, genre, date, and broader rhetorical context
Paragraph/section	Local movement, contrast, or search scope
Sentence	Minimum citable annotation context
Lexical unit	MIPVU decision unit
Source span	Boundary for CMT or interpretive evidence
Mapping	Source-target relation and entailments
Agency/absence scope	Declared context in which presence or nonpresence is evaluated

Code every assigned lexical unit, including ordinary negative controls. Do not skip function words or apparently literal units because a nearby expression is interesting. Use only the context supplied or explicitly authorized by the packet.

MIPVU identification procedure

MIPVU asks:

Is this lexical unit metaphor-related in this context?

For each unit:

State its contextual meaning in this sentence.
Identify a more basic contemporary or historically appropriate meaning.
Cite the dictionary, lexicon, or reference used when the packet requires it.
Decide whether the meanings contrast.
Decide whether the contextual meaning can be understood by comparison with the basic meaning.
Select one controlled decision.
Record confidence, uncertainty, and notes independently.

Controlled decisions

non_metaphor: no metaphorically relevant basic/contextual contrast.
mipvu_indirect: an indirect comparison supports the contextual meaning.
mipvu_direct: an explicit metaphor, simile, analogy, or comparison.
mipvu_implicit: omitted or implied material carries a metaphorical role.
mipvu_personification: a nonhuman entity is construed through human body, agency, emotion, or action.
uncertain: the evidence remains materially contested or insufficient.
excluded_nonlexical: the generated token is not a lexical unit.

Theoretical importance is not a MIPVU criterion. Literal violence, death, blood, body, nation, sacrifice, or enemy language remains non_metaphor when the contextual meaning is literal.

Required rationale

For metaphor-related or uncertain units, record:

contextual meaning;
basic meaning;
basic-meaning source;
contrast explanation;
comparison basis;
confidence;
uncertainty status and note; and
a review note explaining historical semantics, OCR, register, or rival concerns.

Period meaning matters. A modern English meaning or translation may be the wrong basic meaning for an older French, German, or English source.

Lexical and source-span boundaries

Boundary decisions are independent from category decisions. A coder can agree that an expression is metaphor-related but disagree about its span.

Use the packet’s controlled boundary response:

exact: supplied offsets capture the relevant unit;
expand: add adjacent source material;
contract: remove material;
split: one proposed span contains multiple independently coded units;
merge: adjacent supplied units form one expression for the assigned task;
no_valid_span: no lawful source span supports the proposed item; or
uncertain: more context or tokenization guidance is required.

Preserve source character offsets. Do not calculate boundaries from the gloss. For discontinuous or morphologically complex expressions, follow the packet/codebook rule and explain any proposed split or merge.

CMT mapping procedure

CMT coding begins from the supplied source span; it does not repeat the MIPVU decision. The presence of a CMT item does not reveal that the reference accepts the span as metaphorical.

For each item:

Identify the primary source domain evoked by the expression.
Add a secondary source domain only when separately supported.
Identify the target domain being understood.
State the conceptual mapping in a controlled, reviewable form.
List only entailments licensed by the local expression and authorized context.
Select a cluster only when the mapping satisfies that cluster’s definition.
Record a rival reading that could explain the language differently.

Do not:

copy a familiar cluster because the case is known for it;
infer source domains from the English gloss alone;
treat co-activation as multiple independent pieces of evidence;
inflate a conventional expression into a novel ideological mapping; or
use a cluster label as a substitute for source-target analysis.

Interpretive procedure

Interpretation asks which bounded political-symbolic functions are supported by the supplied text. It does not diagnose a speaker, prove historical causation, or establish Koenigsberg’s theory.

For each standard function, choose present, absent, uncertain, or not_applicable:

sacred object;
sacrificial body;
enemy as bringer of death;
violence logic;
obligatory frame; and
purification.

Absent means the field was applicable and evaluated but not supported in the declared context. Not_applicable means the question does not apply to this item. Uncertain means applicable evidence is materially ambiguous.

Violence and obligation

Record violence_logic only when the text frames violence, injury, killing, or coercion as a means, necessity, defense, cure, payment, purification, rescue, or other authorized process. Literal mention of violence does not itself establish a violence logic.

Record obligatory_frame only when an agent or collective is represented as bound, required, destined, commanded, owing, unable legitimately to refuse, or morally compelled. Strong preference, prediction, or praise is not obligation.

State who is obliged, by what authority or condition, to do what, and with what represented consequence.

Sacred object and sacrificial body

A sacred object is represented as ultimate, transcendent, inviolable, immortal, or worth dying/killing for. Honorific language alone is insufficient.

A sacrificial body is represented as an offering, payment, instrument, victim, martyr, or bodily material through which an object is preserved or made real. Literal injury without offering, exchange, obligation, or sacralization may be violence but not sacrificial-body evidence.

Enemy as bringer of death and purification

Enemy-as-bringer-of-death requires the enemy to be represented as agent, carrier, embodiment, or sign of collective death, dissolution, contamination, or unreality. Mere opposition is insufficient.

Purification requires removal of represented stain, disease, corruption, pollution, impurity, or contaminating presence. Administrative reform or ordinary improvement is not automatically purification.

Agency coding

Agency coding records textual role assignment, not the coder’s historical judgment.

Identify:

agents: actors represented as initiating or controlling action;
patients: entities acted upon or affected;
beneficiaries: entities represented as gaining from action;
sacrificial subjects: bodies or groups represented as expended, offered, or required to suffer;
excluded agents: plausible actors whose agency is displaced or suppressed in the declared scope.

Grammar can obscure agency through passive voice, nominalization, divine or natural causation, collective abstraction, or agentless necessity. Record the textual construction and avoid inventing an agent merely because history suggests one.

Absence coding

Absence is not whatever the coder wishes the text had mentioned. Before coding absence, identify:

the declared search scope;
the expected presence;
what would count as that presence;
whether the supplied context is sufficient; and
a possible displacement mechanism.

Choose:

present: the expected element is present;
absent: the element is meaningfully absent within adequate scope;
uncertain: scope or evidence does not resolve the question;
not_applicable: no justified expectation exists.

Do not turn a missing packet excerpt into a substantive absence. Use out_of_scope or uncertain when the required document, register, or paragraph-level context is unavailable.

Confidence, ambiguity, and uncertainty

Confidence records how strongly the available evidence supports the selected decision. It is not the importance of the item and is not a substitute for an uncertainty note.

Use the cohort’s numeric scale consistently. As a general interpretation:

high confidence: the decision and rationale are stable under plausible alternatives;
moderate confidence: one reading is better supported, but a material rival remains;
low confidence: evidence is weak, context-limited, or sensitive to language, boundary, OCR, or codebook uncertainty.

Ambiguity means two or more readings remain linguistically available. Uncertainty can also arise from missing context, source quality, historical semantics, rights restrictions, or an incomplete codebook. Record the cause, not merely the word “uncertain.”

Do not raise confidence to make the dataset look complete. Do not lower confidence as a covert way to disagree with a controlled category.

Rival readings

Every CMT or interpretive item requires a plausible alternative explanation. A rival reading is not a token disclaimer. It should identify the strongest nonpreferred account, such as:

literal description;
conventional idiom;
legal formula;
religious or ceremonial convention;
genre expectation;
translation or gloss artifact;
OCR or transcription error;
alternate source/target mapping;
audience adaptation; or
insufficient context.

The preferred decision should explain why it fits the supplied evidence better, or remain uncertain when it does not.

Out-of-scope protocol

Use out_of_scope only for a declared reason, not to avoid a difficult but answerable item. Record one or more reasons:

language competence;
rights or access;
corrupted source;
missing context;
invalid or unknown identifier;
task/codebook mismatch;
packet construction defect; or
conflict/contamination discovered during coding.

Do not supply a substantive value alongside out_of_scope unless the eventual submission schema explicitly permits a preserved provisional response. Notify the study coordinator when the reason indicates a packet-wide defect.

Worked synthetic examples

These examples are training-only. Their wording is invented, their IDs begin with training-synthetic-, and they are excluded from sampling and calibration.

Example 1: English MIPVU and boundary

training-synthetic-en-01

The council built a bridge between the two districts.

Assigned lexical unit: bridge.

Contextual meaning: a means of connecting political communities.
More basic meaning: a physical structure spanning an obstacle.
Decision: mipvu_indirect.
Boundary: exact for bridge; built a bridge may be retained as a wider CMT evidence span but does not change the lexical-unit decision.
Comparison basis: social connection is understood through physical passage.
Rival: bridge may be a highly conventionalized policy idiom.
Confidence: moderate-to-high, with conventionality noted.

This identification does not establish reconciliation, sacred object, obligation, or a historical outcome.

Example 2: English literal negative control

training-synthetic-en-02

Workers repaired the wooden bridge before sunrise.

Assigned lexical unit: bridge.

Contextual meaning: a physical structure spanning an obstacle.
Basic meaning: the same physical structure.
Decision: non_metaphor.
Boundary: exact.

The word’s metaphorical use elsewhere is irrelevant.

Example 3: Direct comparison and uncertainty

training-synthetic-en-03

The rumor moved through the town like smoke.

Assigned span: like smoke.

Decision: mipvu_direct.
Candidate source domain: smoke spreading through air.
Candidate target: circulation of a rumor.
Entailment: diffusion can be rapid and difficult to contain.
Rival: the comparison may foreground visibility or obscurity rather than diffusion.
Confidence: moderate; preserve the entailment ambiguity.

Example 4: French source-language judgment

training-synthetic-fr-01

La réforme ouvre une voie étroite vers la paix.

Working gloss: “The reform opens a narrow path toward peace.”

Assigned lexical unit: voie.

Source-language contextual meaning: a figurative route or course toward a political condition.
Basic meaning: a physical way, route, or passage.
Decision: likely mipvu_indirect.
Candidate mapping: ACHIEVING PEACE IS TRAVELING ALONG A PATH.
Rival: voie may be conventional political vocabulary with weak image activation.

The French lexical and grammatical evidence controls the decision. The English word “path” does not prove the mapping. A coder unable to evaluate voie and ouvrir une voie in French must use out_of_scope.

Example 5: German gloss conflict

training-synthetic-de-01

Der Vertrag trägt die Stadt durch den Winter.

Working gloss: “The treaty sustains the city through the winter.”

Assigned lexical unit: trägt.

Source-language basic meaning includes physically carrying or bearing.
Contextual meaning concerns enabling the city to endure.
Decision: plausible mipvu_indirect.
Candidate mapping: POLITICAL SUPPORT IS PHYSICAL CARRYING.
Gloss warning: “sustains” softens the carrying image and cannot replace the German decision.
Rival: tragen may be conventionalized as support or endurance.

Example 6: Violence without sacrificial interpretation

training-synthetic-en-04

The guards struck the gate until the lock broke.

Violence is literally present.
violence_logic: may be present only if the assigned interpretation field treats force-as-means to entry and the packet context supports that scope.
sacrificial_body: absent or not_applicable; no body is represented as offering, payment, or sacred instrument.
obligatory_frame: absent unless additional context presents the action as required.

Do not infer sacrifice from force alone.

Example 7: Obligation and agency

training-synthetic-en-05

By the oath, each delegate must carry the city’s burden.

obligatory_frame: present; the oath is represented as binding delegates.
Agent: each delegate.
Beneficiary or protected object: the city, if supported by context.
Candidate source domain: physical carrying.
Candidate target: civic responsibility.
Rival: formulaic oath language may explain some of the force.
Unresolved question: what the “burden” consists of; do not invent sacrifice.

Example 8: Absence requires scope

training-synthetic-en-06

The proclamation praises victory and names the commanders.

Question: Is civilian suffering absent?

Sentence-only scope: uncertain or not_applicable; one sentence does not establish a meaningful proclamation-wide absence.
Full-document scope with an explicit expectation and completed search: absent may be available if no civilian patients, losses, or consequences appear.

The same text can support different absence decisions under different declared scopes. Never infer a larger search than the packet supplies.

Example 9: Personification versus ordinary agency

training-synthetic-en-07

The law remembers the forgotten claimant.

law is nonhuman and is construed as remembering.
Decision for remembers: plausible mipvu_personification.
Agency note: the text assigns agency to law, potentially displacing legislators, judges, or administrators.
Rival: conventional legal shorthand may reduce personification salience.

The coder records textual agency; they do not decide from this sentence who actually implemented the law.

Example 10: Out of scope

training-synthetic-fr-02

[OCR fragment: “La p—rie exige …”]

The focal word is corrupted and the packet supplies no lawful image or edition check.

Decision: out_of_scope.
Reason: corrupted source / missing lawful verification context.
Do not guess patrie from the partial string and then code obligation.

Common coding failures

Failure	Correction
Coding the theory instead of the word	Return to contextual/basic meaning and the assigned layer
Letting a gloss determine the source decision	Re-read the source; use `out_of_scope` if not competent
Treating all violence as sacrifice	Require offering, payment, obligation, or sacralization evidence
Treating all praise as sacred-object evidence	Require ultimacy, inviolability, transcendence, or death/killing-worthiness
Inferring absence from a short excerpt	Check the declared search scope
Using low confidence instead of `uncertain`	Select the category and uncertainty status independently
Omitting a rival because confidence is high	Record the strongest plausible alternative
Copying another coder or accepted reference	Stop, report contamination, and preserve independence
Guessing when context or access is missing	Use `out_of_scope` with a reason
Converting adjudication into original agreement	Preserve pre-adjudication coder values and metrics

Coder workflow

Before coding:

confirm cohort, source language, task layer, packet ID, and packet hash;
confirm training and calibration versions;
review rights and storage requirements;
attest source-language competence and conflicts;
disable or avoid unauthorized AI, retrieval, and collaboration tools; and
verify that no accepted decisions or other coder outputs are visible.

For each item:

verify stable IDs and source text;
identify the assigned unit and task layer;
read the authorized source context;
make the categorical or field decision;
check boundaries where assigned;
record rationale, confidence, uncertainty, and rival reading;
record agency/absence scope where assigned;
use out_of_scope if the item cannot be lawfully and competently coded;
save without consulting another coder; and
move on rather than informally adjudicating your own uncertainty.

Before submission:

confirm every assigned item has exactly one permitted disposition;
confirm no source text or ID was altered;
confirm required rationales and notes are complete;
review low-confidence and out-of-scope items for accurate reasons;
confirm no accepted values, model outputs, or other coder decisions were copied into comments;
preserve the original packet and completed response bytes; and
submit through the cohort’s registered procedure.

Training completion attestation

Before calibration, the coder should be able to attest:

I read the required guide and codebook versions.
I understand that accepted annotations, model outputs, other coder values, and adjudication are withheld during blind coding.
I will code only in source-language cohorts for which my competence has been declared and reviewed.
I understand that English glosses are aids and do not control source-language identification, boundaries, or mappings.
I can distinguish MIPVU, CMT, interpretation, agency/absence, and historical corroboration.
I can use uncertainty, rival readings, and out_of_scope without guessing.
I will report accidental exposure, conflicts, source defects, and rights problems.
I understand that coder agreement does not prove correctness and that adjudication does not rewrite accepted annotations automatically.

Calibration completion and any qualification decision are recorded separately under the versioned cohort workflow. Do not record training completion by editing accepted corpus or annotation artifacts.