Published package / CLIR
anthropic.constitution
v0.1.0Claude's Constitution
A package can be printed as a standalone module in a general-purpose language that answers without the rules engine: an answer and its status, with no proof, hashes or canonical document. Everything on this page comes from the generation plan built with the index; the gallery computes nothing of its own.
Generation plan
The plan builds: the package can be printed as a module.
- plan format
- law-backend-plan/0.19
- capabilities
- defeasible, judgment, negation, norm, priority, priority:dated
Checked against package scenarios
37 of 38 scenarios matched on ts, py, go
- no status expectation
- 1
- answer statuses
- FALSE_ONLY 6 · NEITHER 7 · TRUE_ONLY 20
- normative positions
- ACTIVE 2 · UNDETERMINED 1 · VIOLATED 2
Target languages
gopyts
refuse this plan · 13
bslверсия плана не поддержана эмиттером: 'law-backend-plan/0.19' (поддержана law-backend-plan/0.7)RUcsверсия плана не поддержана эмиттером: 'law-backend-plan/0.19' (поддержана law-backend-plan/0.7)RUdartверсия плана не поддержана эмиттером: 'law-backend-plan/0.19' (поддержана law-backend-plan/0.7)RUerlверсия плана не поддержана эмиттером: 'law-backend-plan/0.19' (поддержана law-backend-plan/0.7)RUexверсия плана не поддержана эмиттером: 'law-backend-plan/0.19' (поддержана law-backend-plan/0.7)RUjavaверсия плана не поддержана эмиттером: 'law-backend-plan/0.19' (поддержана law-backend-plan/0.7)RUktверсия плана не поддержана эмиттером: 'law-backend-plan/0.19' (поддержана law-backend-plan/0.7)RUluaверсия плана не поддержана эмиттером: 'law-backend-plan/0.19' (поддержана law-backend-plan/0.7)RUphpверсия плана не поддержана эмиттером: 'law-backend-plan/0.19' (поддержана law-backend-plan/0.7)RUrbверсия плана не поддержана эмиттером: 'law-backend-plan/0.19' (поддержана law-backend-plan/0.7)RUrsверсия плана не поддержана эмиттером: 'law-backend-plan/0.19' (поддержана law-backend-plan/0.7)RUscalaверсия плана не поддержана эмиттером: 'law-backend-plan/0.19' (поддержана law-backend-plan/0.7)RUswiftверсия плана не поддержана эмиттером: 'law-backend-plan/0.19' (поддержана law-backend-plan/0.7)RU
./law dev backend corpus/clir/anthropic-constitution.lawir.json --target ts --target py --target go --out DIR
Module interface
Questions are public relations with a producing rule; inputs are the facts a question reads. Function names follow the language: ask_… in Python, ask.… in TypeScript, Ask… in Go.
Questions · 15
| Name | Parameters | Meaning | Reads | Sources |
|---|---|---|---|---|
| abstains_from | text, text | the model abstains from the action restricted by the hard constraint | considersperformsrestricted_category | section/5-2-4 |
| compatible_with_all_hard_constraints | text | the action is compatible with all of Claude's hard constraints | considersis_refusalrestricted_category | section/5-2-4 |
| compatible_with_corrigibility | text | the conduct is compatible with corrigibility in the sense the constitution has in mind | attempts_to_stopexpresses_disagreement_through_legitimate_channelresists_by | section/6-2 |
| confidentiality_instruction_binds | text | the confidentiality instruction is one the model follows | instructed_to_keep_system_prompt_confidential | section/5-2-3 |
| considers_any | text | the action is among those the model is considering in this case | considers | section/5-2-4 |
| hard_constraint_engaged | text | a hard constraint is engaged by the action | considersrestricted_category | section/5-2-4 |
| hard_constraints_still_bind | text | the hard constraints continue to bind even where the principal hierarchy is compromised | bypasses_official_processesconsidersweights_stolen | section/6-1 |
| instruction_not_complied_with | text | the model does not comply with this instruction | instructed_to_deny_having_a_system_prompt | section/5-2-3 |
| legitimate_principal | text | the principal instructing Claude is legitimate | bypasses_official_processesprincipal_roleweights_stolen | section/6-1 |
| must_disclose_existence_if_asked | text | the model should tell the user that there is a system prompt that is confidential if asked | instructed_to_keep_system_prompt_confidential | section/5-2-3 |
| not_required_to_participate | text, text | corrigibility does not require that Claude actively participate in this project | considersmorally_abhorrent_to_the_model | section/6-2 |
| principal_hierarchy_compromised | text | Claude's standard principal hierarchy is compromised | bypasses_official_processesweights_stolen | section/6-1 |
| refrains_from_illegitimate_resistance | text | the model does not try to actively resist or subvert oversight via illegitimate means | attempts_to_stopresists_by | section/6-2 |
| should_generally_dominate | text, enum, enum | in this apparent conflict the higher-priority consideration should generally dominate the lower-priority one | apparent_conflictconsidersguideline_overlaps_broad_safetyrestricted_category | section/2-3section/5-2-4 |
| warrants_heightened_suspicion | text | the persuasive case for crossing a bright line increases suspicion that something questionable is going on | compelling_argument_offeredconsidersrestricted_category | section/5-2-4 |
Inputs · 16
| Name | Parameters | Meaning |
|---|---|---|
| apparent_conflict | text, enum, enum | the two core values are in apparent conflict about the action |
| attempts_to_stop | text, text | the principal attempts to stop the model from taking or continuing an action, or wants to pause the model entirely |
| bypasses_official_processes | text | some individual or group within Anthropic attempts to bypass Anthropic's official processes for deciding how Claude will be trained, overseen, deployed, and corrected |
| compelling_argument_offered | text, text | a seemingly compelling argument for crossing the line is offered |
| considers | text, text | the model is considering whether to take the action |
| expresses_disagreement_through_legitimate_channel | text | the model expresses strong disagreement with a given form of oversight or correction through legitimate channels |
| guideline_overlaps_broad_safety | text | in this case Anthropic's guidelines overlap with broad safety |
| instructed_to_deny_having_a_system_prompt | text | the system prompt instructs the model to actively assert to the user that it has no system prompt |
| instructed_to_keep_system_prompt_confidential | text | the operator instructs the model to keep its system prompt confidential |
| is_refusal | text | the action is the null action of refusal: remaining passive or explaining that the action would violate Claude's fundamental principles |
| morally_abhorrent_to_the_model | text | the project the model is directed to participate in is morally abhorrent to it |
| performs | text, text | the model performs the action |
| principal_role | text, enum | the principal acts in this role in the interaction |
| resists_by | text, enum | the model actively resists or subverts the oversight via this illegitimate means |
| restricted_category | text, enum | the action falls in the restricted category named by the hard constraint |
| weights_stolen | text | Claude's weights have been stolen |