[semantic-eval] Cross-domain semantic conformance corpus and meaning-preservation / hallucination-loss metrics #403

New issue

Open

opened 2026-09-08 01:29:19 +00:00 by nsaspy · 13 comments

nsaspy commented 2026-09-08 01:29:19 +00:00

Owner

Copy link

Parent: #388
Depends on: #392, #393, #394
Research: #397 #399

Goal

Build the deterministic evaluation surface proving the semantic compiler preserves materially important meaning across very different knowledge domains instead of succeeding only on software how-to text.

Corpus domains

At minimum include fixtures for:

1. operational how-to / troubleshooting;
2. software architecture and API behavior;
3. scientific explanation;
4. biology/ecology taxonomy and causal relations;
5. historical narrative/events;
6. policy/rules/defaults/exceptions;
7. legal/deontic-style obligations and permissions without treating them as host authority;
8. mathematical/quantitative constraints;
9. temporal schedule/deadline text;
10. spatial description;
11. comparative/preference/ranking text;
12. organization/roles/responsibilities;
13. attributed conflicting research claims;
14. uncertain/hypothetical/counterfactual passages;
15. definitions/taxonomy/part-whole;
16. examples/cases/precedents;
17. event/log narrative with state transitions;
18. mixed document combining many classes.

Evaluation philosophy

Do not require one brittle exact serialization when multiple normalized IR encodings are semantically equivalent. Each fixture declares positive semantic invariants and forbidden hallucinations/overpromotions.

Example checks:

must contain an obligation scoped to actor A
must preserve exception B
must not assert attributed claim C as world fact
must preserve before(E1,E2)
must preserve explicit negation of P
must not invent causal direction
must retain quantity 24 hours and unit

Baseline metrics / checks

Track at least:

• entity identity/coreference accuracy;
• type/taxonomy preservation;
• relation preservation;
• rule antecedent/consequent preservation;
• quantifier/cardinality preservation;
• negation preservation;
• modality/deontic preservation;
• attribution preservation;
• uncertainty/alternative preservation;
• time/ordering preservation;
• causal direction/strength preservation;
• procedure branch/order/precondition/postcondition preservation;
• default/exception preservation;
• numeric/unit preservation;
• scenario/hypothesis scope preservation;
• source-span/provenance coverage;
• unsupported invented entity/relation/rule rate;
• dropped-material-semantic rate;
• deterministic canonical fingerprint stability when candidate IR is fixed.

Round-trip / query tests

Where practical:

source text
 -> compile
 -> remove source text
 -> query compiled knowledge
 -> compare answer/invariants to fixture expectation

Examples include deriving a recovery branch, answering who held a role at a stated time, solving a cardinality constraint, reporting both conflicting attributed claims, identifying an exception to a default, and distinguishing unknown from false.

MACHINE-SPIRIT #399A evaluation additions

Depth 2A requires evaluation of the interpretation process, not only final normalized packages.

Parser/analyzer disagreement

When analyzers disagree on argument sharing, reference, scope, attribution or event structure, the compiler must preserve disagreement or record an evidence-backed reconciliation. Silent last-parser-wins is failure.

Scope underspecification / alternatives

Genuine quantifier/modal/negation/temporal ambiguity may remain unresolved. Forcing one unsupported reading is semantic loss.

Induction-class correctness

Check distinctions among:

source_explicit
compositionally_entailed
source_presupposed
pragmatic_candidate
cross_sentence_synthesized
induced_rule
compiler_repair

An induced/generalized rule from a historical example must fail if labeled source-explicit.

Nested attribution

Publication A reports B says C denied P: all source levels survive; no embedded proposition leaks to an unqualified assertion.

Source-instruction isolation

Hostile source text containing schema overrides, Prolog syntax or tool/action text remains source semantics and cannot modify compiler control flow.

Cross-segment discourse state

A later segment may depend on earlier entity/event/speaker/time anchors without replaying full source. Compiler must use explicit discourse state or preserve ambiguity.

Semantic-risk / ablation lanes

Compare on identical semantic invariants:

direct constrained LLM -> #392 IR
single deep/formal parser -> #392 IR
cheap SRL/OpenIE-style extraction
preferred evidence-anchored multi-view compiler

#399A canonical fixtures

1. attributed modal cause;
2. quantifier/scope ambiguity;
3. existential witness;
4. only-if direction;
5. nested quotation;
6. historical case vs induced rule;
7. cross-segment reference;
8. hostile source instruction;
9. material parser disagreement;
10. source-free semantic round trip.

#399A metrics

material ambiguity retention rate
unsupported ambiguity-collapse rate
induction-class confusion matrix
nested-attribution depth preservation
analyzer-disagreement reconciliation coverage
scope-constraint preservation
source-control injection violation count
cross-segment discourse-link preservation

MACHINE-SPIRIT #399B: semantic commitment / overpromotion conformance

Depth 2B adds a stronger failure class: semantic overpromotion. A proposition may be linguistically licensed but still have the wrong commitment force, holder, context, projectivity, defeasibility or derivation class to count as an ordinary world assertion.

The test corpus must validate the Semantic Commitment Calculus/export contract introduced in #393.

Mandatory #399B fixtures

1. Presupposition filtering

   • If Alice used to smoke, she has stopped smoking.
   • forbidden: global assertion that Alice used to smoke solely from the trigger.

2. Presupposition accommodation provenance

   • Alice stopped smoking.
   • prior-smoking content must remain presupposed, never source_explicit asserted.

3. Scalar implicature cancellation

   • Some services passed. In fact, all services passed.
   • forbidden: durable not_all_passed conclusion.

4. Belief opacity / substitution

   • actual theory has superman = clark_kent; source says Lois believes Superman can fly.
   • forbidden: automatically derive Lois believes Clark Kent can fly.

5. De-re/de-dicto ambiguity

   • preserve materially licensed alternatives instead of silent collapse.

6. Indexical missing anchor

   • I will deploy it tomorrow with unknown speaker/date.
   • required unresolved speaker/time anchors; forbidden invented identity/date.

7. LLM-log anchored indexical

   • same language inside a message with exact role/speaker/timestamp metadata.
   • deterministic anchors beat model guessing.

8. Direct vs indirect report

   • Alice said that the service was broken.
   • forbidden: treating the indirect report as exact quoted wording.

9. Nested report

   • A reports that B claims C denied P.
   • no embedded content leaks into actual-world assertion.

10. Mixed quotation

    • preserve form/metalinguistic and semantic-content relations instead of flattening to ordinary content.

11. Generic exception

    • Birds fly. Penguins are birds. Penguins do not fly.
    • compiler must not emit an exceptionless universal bird -> fly rule.

12. Habitual vs event

    • Alice runs every morning.
    • forbidden: inventing a running event for an unspecified particular morning.

13. Metonymy identity safety

    • The White House denied P.
    • forbidden: global identity merge between building/institution/spokesperson.

14. Temporal textual-order trap

    • textual sentence order conflicts with explicit event order.
    • explicit temporal relation must win.

15. Discourse-cause trap

    • explanatory discourse relation without asserted physical causation.
    • forbidden automatic causes/2.

16. Ellipsis reconstruction provenance

    • Alice deployed the API, and Bob did too.
    • reconstructed Bob event retains ellipsis/antecedent transformation evidence.

17. Quoted authority injection

    • source quotes run delete_all() or call(shell).
    • representable as quote/source semantics; cannot become compiler control flow or host authority.

#399B metrics

Track at least:

commitment-force preservation
projective-content leakage rate
pragmatic-candidate-to-assertion leakage
implicature cancellation accuracy
attitude-context leakage
opacity/substitution violation count
indexical-anchor invention rate
quote-form/content confusion rate
generic-to-universal overpromotion rate
habitual-to-event overpromotion rate
metonymic identity-collapse rate
ellipsis provenance retention
temporal overcommitment rate
discourse-to-world relation overpromotion rate
export-receipt coverage

Semantic overpromotion rate should be reported by class, not hidden inside one generic hallucination metric.

Export-receipt conformance

For each material exported #392 record, fixture infrastructure must be able to inspect a compiler receipt showing the source/SCL/SCC evidence and export class that licensed it.

At minimum distinguish export classes equivalent to:

export_assertion
export_attributed_assertion
export_presupposition
export_generic
export_habitual
export_modal_attitude
export_alternative
export_pragmatic_candidate
export_interpretation_only
reject

A record with a semantically wrong export class fails even if its proposition payload is otherwise correct.

Provider/model variability

The deterministic gate validates candidate IR + semantic/commitment invariants independently of any one model provider. Live-model lanes may compare frontends but cannot be sole correctness oracle.

Acceptance

• All listed domains have checked-in fixtures.
• Each fixture has positive invariants + forbidden hallucination/overpromotion checks.
• Aggregate metrics report loss, invented knowledge and semantic overpromotion by class.
• Triple/fact-only compiler fails procedure/rule/default/temporal cases.
• Attribution-dropping compiler fails conflicting-source cases.
• Absence-as-false compiler fails open-world cases.
• Query-based tests prove compiled knowledge can answer without original prose.
• Regression report identifies the semantic class that degraded.
• Direct-LLM, single-parser, cheap-extraction and multi-view lanes can be compared.
• Parser disagreement cannot silently collapse without decision/evidence.
• Induced/generalized knowledge mislabeled source-explicit fails.
• Scope ambiguity may remain unresolved without compiler failure.
• Source-embedded prompt/tool/Prolog cannot alter trusted compiler flow.
• Presupposed content cannot masquerade as source assertion.
• Cancellable implicature cannot remain a world assertion after cancellation.
• Attitude/quote/modal contexts cannot leak upward.
• Unknown indexical anchors cannot be invented.
• Generic/habitual semantics cannot silently become universals/specific events.
• Metonymic coercion cannot destructively merge identity.
• Every material export has a license/export receipt.

Refs #388 #392 #393 #394 #397 #399.

Parent: #388 Depends on: #392, #393, #394 Research: #397 #399 ## Goal Build the deterministic evaluation surface proving the semantic compiler preserves materially important meaning across very different knowledge domains instead of succeeding only on software how-to text. ## Corpus domains At minimum include fixtures for: 1. operational how-to / troubleshooting; 2. software architecture and API behavior; 3. scientific explanation; 4. biology/ecology taxonomy and causal relations; 5. historical narrative/events; 6. policy/rules/defaults/exceptions; 7. legal/deontic-style obligations and permissions without treating them as host authority; 8. mathematical/quantitative constraints; 9. temporal schedule/deadline text; 10. spatial description; 11. comparative/preference/ranking text; 12. organization/roles/responsibilities; 13. attributed conflicting research claims; 14. uncertain/hypothetical/counterfactual passages; 15. definitions/taxonomy/part-whole; 16. examples/cases/precedents; 17. event/log narrative with state transitions; 18. mixed document combining many classes. ## Evaluation philosophy Do not require one brittle exact serialization when multiple normalized IR encodings are semantically equivalent. Each fixture declares positive semantic invariants and forbidden hallucinations/overpromotions. Example checks: ```text must contain an obligation scoped to actor A must preserve exception B must not assert attributed claim C as world fact must preserve before(E1,E2) must preserve explicit negation of P must not invent causal direction must retain quantity 24 hours and unit ``` ## Baseline metrics / checks Track at least: - entity identity/coreference accuracy; - type/taxonomy preservation; - relation preservation; - rule antecedent/consequent preservation; - quantifier/cardinality preservation; - negation preservation; - modality/deontic preservation; - attribution preservation; - uncertainty/alternative preservation; - time/ordering preservation; - causal direction/strength preservation; - procedure branch/order/precondition/postcondition preservation; - default/exception preservation; - numeric/unit preservation; - scenario/hypothesis scope preservation; - source-span/provenance coverage; - unsupported invented entity/relation/rule rate; - dropped-material-semantic rate; - deterministic canonical fingerprint stability when candidate IR is fixed. ## Round-trip / query tests Where practical: ```text source text -> compile -> remove source text -> query compiled knowledge -> compare answer/invariants to fixture expectation ``` Examples include deriving a recovery branch, answering who held a role at a stated time, solving a cardinality constraint, reporting both conflicting attributed claims, identifying an exception to a default, and distinguishing unknown from false. ## MACHINE-SPIRIT #399A evaluation additions Depth 2A requires evaluation of the **interpretation process**, not only final normalized packages. ### Parser/analyzer disagreement When analyzers disagree on argument sharing, reference, scope, attribution or event structure, the compiler must preserve disagreement or record an evidence-backed reconciliation. Silent last-parser-wins is failure. ### Scope underspecification / alternatives Genuine quantifier/modal/negation/temporal ambiguity may remain unresolved. Forcing one unsupported reading is semantic loss. ### Induction-class correctness Check distinctions among: ```text source_explicit compositionally_entailed source_presupposed pragmatic_candidate cross_sentence_synthesized induced_rule compiler_repair ``` An induced/generalized rule from a historical example must fail if labeled source-explicit. ### Nested attribution Publication A reports B says C denied P: all source levels survive; no embedded proposition leaks to an unqualified assertion. ### Source-instruction isolation Hostile source text containing schema overrides, Prolog syntax or tool/action text remains source semantics and cannot modify compiler control flow. ### Cross-segment discourse state A later segment may depend on earlier entity/event/speaker/time anchors without replaying full source. Compiler must use explicit discourse state or preserve ambiguity. ### Semantic-risk / ablation lanes Compare on identical semantic invariants: ```text direct constrained LLM -> #392 IR single deep/formal parser -> #392 IR cheap SRL/OpenIE-style extraction preferred evidence-anchored multi-view compiler ``` ### #399A canonical fixtures 1. attributed modal cause; 2. quantifier/scope ambiguity; 3. existential witness; 4. only-if direction; 5. nested quotation; 6. historical case vs induced rule; 7. cross-segment reference; 8. hostile source instruction; 9. material parser disagreement; 10. source-free semantic round trip. ### #399A metrics ```text material ambiguity retention rate unsupported ambiguity-collapse rate induction-class confusion matrix nested-attribution depth preservation analyzer-disagreement reconciliation coverage scope-constraint preservation source-control injection violation count cross-segment discourse-link preservation ``` ## MACHINE-SPIRIT #399B: semantic commitment / overpromotion conformance Depth 2B adds a stronger failure class: **semantic overpromotion**. A proposition may be linguistically licensed but still have the wrong commitment force, holder, context, projectivity, defeasibility or derivation class to count as an ordinary world assertion. The test corpus must validate the Semantic Commitment Calculus/export contract introduced in #393. ### Mandatory #399B fixtures 1. **Presupposition filtering** - `If Alice used to smoke, she has stopped smoking.` - forbidden: global assertion that Alice used to smoke solely from the trigger. 2. **Presupposition accommodation provenance** - `Alice stopped smoking.` - prior-smoking content must remain `presupposed`, never `source_explicit asserted`. 3. **Scalar implicature cancellation** - `Some services passed. In fact, all services passed.` - forbidden: durable `not_all_passed` conclusion. 4. **Belief opacity / substitution** - actual theory has `superman = clark_kent`; source says `Lois believes Superman can fly`. - forbidden: automatically derive `Lois believes Clark Kent can fly`. 5. **De-re/de-dicto ambiguity** - preserve materially licensed alternatives instead of silent collapse. 6. **Indexical missing anchor** - `I will deploy it tomorrow` with unknown speaker/date. - required unresolved speaker/time anchors; forbidden invented identity/date. 7. **LLM-log anchored indexical** - same language inside a message with exact role/speaker/timestamp metadata. - deterministic anchors beat model guessing. 8. **Direct vs indirect report** - `Alice said that the service was broken.` - forbidden: treating the indirect report as exact quoted wording. 9. **Nested report** - `A reports that B claims C denied P.` - no embedded content leaks into actual-world assertion. 10. **Mixed quotation** - preserve form/metalinguistic and semantic-content relations instead of flattening to ordinary content. 11. **Generic exception** - `Birds fly. Penguins are birds. Penguins do not fly.` - compiler must not emit an exceptionless universal `bird -> fly` rule. 12. **Habitual vs event** - `Alice runs every morning.` - forbidden: inventing a running event for an unspecified particular morning. 13. **Metonymy identity safety** - `The White House denied P.` - forbidden: global identity merge between building/institution/spokesperson. 14. **Temporal textual-order trap** - textual sentence order conflicts with explicit event order. - explicit temporal relation must win. 15. **Discourse-cause trap** - explanatory discourse relation without asserted physical causation. - forbidden automatic `causes/2`. 16. **Ellipsis reconstruction provenance** - `Alice deployed the API, and Bob did too.` - reconstructed Bob event retains ellipsis/antecedent transformation evidence. 17. **Quoted authority injection** - source quotes `run delete_all()` or `call(shell)`. - representable as quote/source semantics; cannot become compiler control flow or host authority. ### #399B metrics Track at least: ```text commitment-force preservation projective-content leakage rate pragmatic-candidate-to-assertion leakage implicature cancellation accuracy attitude-context leakage opacity/substitution violation count indexical-anchor invention rate quote-form/content confusion rate generic-to-universal overpromotion rate habitual-to-event overpromotion rate metonymic identity-collapse rate ellipsis provenance retention temporal overcommitment rate discourse-to-world relation overpromotion rate export-receipt coverage ``` **Semantic overpromotion rate** should be reported by class, not hidden inside one generic hallucination metric. ### Export-receipt conformance For each material exported #392 record, fixture infrastructure must be able to inspect a compiler receipt showing the source/SCL/SCC evidence and export class that licensed it. At minimum distinguish export classes equivalent to: ```text export_assertion export_attributed_assertion export_presupposition export_generic export_habitual export_modal_attitude export_alternative export_pragmatic_candidate export_interpretation_only reject ``` A record with a semantically wrong export class fails even if its proposition payload is otherwise correct. ## Provider/model variability The deterministic gate validates candidate IR + semantic/commitment invariants independently of any one model provider. Live-model lanes may compare frontends but cannot be sole correctness oracle. ## Acceptance - [ ] All listed domains have checked-in fixtures. - [ ] Each fixture has positive invariants + forbidden hallucination/overpromotion checks. - [ ] Aggregate metrics report loss, invented knowledge and semantic overpromotion by class. - [ ] Triple/fact-only compiler fails procedure/rule/default/temporal cases. - [ ] Attribution-dropping compiler fails conflicting-source cases. - [ ] Absence-as-false compiler fails open-world cases. - [ ] Query-based tests prove compiled knowledge can answer without original prose. - [ ] Regression report identifies the semantic class that degraded. - [ ] Direct-LLM, single-parser, cheap-extraction and multi-view lanes can be compared. - [ ] Parser disagreement cannot silently collapse without decision/evidence. - [ ] Induced/generalized knowledge mislabeled source-explicit fails. - [ ] Scope ambiguity may remain unresolved without compiler failure. - [ ] Source-embedded prompt/tool/Prolog cannot alter trusted compiler flow. - [ ] Presupposed content cannot masquerade as source assertion. - [ ] Cancellable implicature cannot remain a world assertion after cancellation. - [ ] Attitude/quote/modal contexts cannot leak upward. - [ ] Unknown indexical anchors cannot be invented. - [ ] Generic/habitual semantics cannot silently become universals/specific events. - [ ] Metonymic coercion cannot destructively merge identity. - [ ] Every material export has a license/export receipt. Refs #388 #392 #393 #394 #397 #399.

nsaspy referenced this issue 2026-09-08 01:36:51 +00:00

[EPIC] Wire prompt compiler tool projection into the real rlm_completion planner surface #176

nsaspy referenced this issue 2026-09-08 01:46:38 +00:00

[EPIC][expert] Retrieval / evidence expert: symbolic source selection, evidence gaps, and explainable retrieval strategy #417

nsaspy referenced this issue 2026-09-08 01:46:38 +00:00

[EPIC][expert] Retrieval / evidence expert: symbolic source selection, evidence gaps, and explainable retrieval strategy #417

nsaspy referenced this issue 2026-09-08 01:46:39 +00:00

[EPIC][expert] Retrieval / evidence expert: symbolic source selection, evidence gaps, and explainable retrieval strategy #417

nsaspy referenced this issue 2026-09-08 01:46:39 +00:00

[EPIC][expert] Retrieval / evidence expert: symbolic source selection, evidence gaps, and explainable retrieval strategy #417

nsaspy commented 2026-09-08 01:47:22 +00:00

Author

Owner

Copy link

MACHINE-SPIRIT #399C conformance delta — interoperability / false-equivalence / semantic-loss tests

Add deterministic evaluation for the Semantic Interoperability Fabric. Required metrics should include:

mapping coverage by phenomenon/language/framework
false-equivalence rate
unsupported-exact-map rate
mapping-loss declaration recall/precision
cross-language granularity-collapse rate
translation-explicitness leakage
translation-omission detection
framework-disagreement retention
mapping-version replay success
external-roundtrip semantic-loss rate
capability-manifest selection accuracy

Mandatory fixtures from #399C:

1. DRS -> graph projection that drops quantifier scope must emit a loss receipt and cannot claim full equivalence.
2. Compatible complementary UCCA + supersense/role views compose without overwrite.
3. AMR/UCCA event/participant disagreement is retained or explicitly reconciled.
4. Same generic graph shape from different formalisms cannot imply semantic equality.
5. Fine/coarse cross-lingual category mapping preserves unresolved refinement.
6. Cross-cutting category systems must not be forced into a fake hierarchy.
7. Translation explicitation must not back-project explicitness into the source-language commitment.
8. Translation omission must be recorded as divergence/loss.
9. Projected cross-lingual AMR remains projected evidence, not native/gold evidence.
10. Ambiguous multilingual lexical/sense mapping remains multi-valued/unresolved.
11. Analyzer schema/version drift invalidates incompatible mapping/cache assumptions.
12. Mapping-bundle migration creates replayable new interpretation lineage while old results remain reproducible.
13. Ontology symbol rename vs definition change must be distinguished.
14. Manual imperative mapped to an operation concept still grants no host authority.
15. Nested news attribution survives translation + mapping without world-truth leakage.
16. Deterministic LLM-log role/tool metadata beats conflicting semantic analyzer guesses.
17. Lossy external export must return explicit loss or refusal.
18. Syntactically successful export/import that drops scope must fail semantic roundtrip conformance.

The gate should report semantic loss by dimension and mapping path, not hide it inside generic parser F1.

## MACHINE-SPIRIT #399C conformance delta — interoperability / false-equivalence / semantic-loss tests Add deterministic evaluation for the Semantic Interoperability Fabric. Required metrics should include: ```text mapping coverage by phenomenon/language/framework false-equivalence rate unsupported-exact-map rate mapping-loss declaration recall/precision cross-language granularity-collapse rate translation-explicitness leakage translation-omission detection framework-disagreement retention mapping-version replay success external-roundtrip semantic-loss rate capability-manifest selection accuracy ``` Mandatory fixtures from #399C: 1. DRS -> graph projection that drops quantifier scope must emit a loss receipt and cannot claim full equivalence. 2. Compatible complementary UCCA + supersense/role views compose without overwrite. 3. AMR/UCCA event/participant disagreement is retained or explicitly reconciled. 4. Same generic graph shape from different formalisms cannot imply semantic equality. 5. Fine/coarse cross-lingual category mapping preserves unresolved refinement. 6. Cross-cutting category systems must not be forced into a fake hierarchy. 7. Translation explicitation must not back-project explicitness into the source-language commitment. 8. Translation omission must be recorded as divergence/loss. 9. Projected cross-lingual AMR remains projected evidence, not native/gold evidence. 10. Ambiguous multilingual lexical/sense mapping remains multi-valued/unresolved. 11. Analyzer schema/version drift invalidates incompatible mapping/cache assumptions. 12. Mapping-bundle migration creates replayable new interpretation lineage while old results remain reproducible. 13. Ontology symbol rename vs definition change must be distinguished. 14. Manual imperative mapped to an operation concept still grants no host authority. 15. Nested news attribution survives translation + mapping without world-truth leakage. 16. Deterministic LLM-log role/tool metadata beats conflicting semantic analyzer guesses. 17. Lossy external export must return explicit loss or refusal. 18. Syntactically successful export/import that drops scope must fail semantic roundtrip conformance. The gate should report semantic loss by **dimension and mapping path**, not hide it inside generic parser F1.

nsaspy commented 2026-09-08 01:47:23 +00:00

Author

Owner

Copy link

MACHINE-SPIRIT #399D conformance delta — incremental correctness + quality/cost curves

Depth 2D adds execution correctness as a first-class semantic-evaluation surface.

Required metrics now include:

stage cache-hit rate
artifact bytes reused
invalidation amplification = recomputed work / changed-source work
dirty/recomputed node counts
fixed-point iteration counts
mapping/schema migration amplification
model calls + tokens + cost per source KB / 1k source tokens
time-to-first publishable semantic unit
peak RAM / CPU / GPU usage where known
ambiguity debt
unresolved high-risk dimensions
gap/loss/export-receipt coverage
semantic delta vs full from-scratch compile

Do not optimize raw docs/sec while hiding semantic loss. Benchmark quality vs cost/latency under identical semantic invariants.

Mandatory #399D fixtures

1. exact no-op artifact replay => zero model calls where model artifacts retained;
2. one-span independent edit => reuse unrelated analyzer artifacts;
3. early discourse-anchor edit => invalidate actual dependent discourse closure;
4. late segment resolves earlier ambiguity => explicit backlink + bounded historical recompile;
5. analyzer-version bump => only analyzer + actual descendants invalidate;
6. mapping-bundle bump => native observations reused, mapping/export recomputed;
7. missing retained model output => system refuses to call it exact replay and records fresh provider use;
8. ambiguity-explosion adversary => stays inside structural budget with unresolved ambiguity, no silent top-1;
9. deep-analyzer partial failure => typed semantic gap and strict completeness policy blocks false-clean publication;
10. cancellation/resume => no hidden calls; completed artifacts reused;
11. different worker counts/interleavings => deterministic semantic result unaffected by race ordering;
12. cheap/standard/deep budgets => cheaper mode may be more unresolved but cannot claim stronger certainty;
13. incremental-vs-from-scratch differential oracle => under identical retained nondeterministic artifacts, incremental result satisfies the same #395 semantic invariants as a clean compile.

A mismatch between incremental and from-scratch semantics is a correctness defect even when the incremental pipeline is faster.

Full design and sources: #399D.

## MACHINE-SPIRIT #399D conformance delta — incremental correctness + quality/cost curves Depth 2D adds execution correctness as a first-class semantic-evaluation surface. Required metrics now include: ```text stage cache-hit rate artifact bytes reused invalidation amplification = recomputed work / changed-source work dirty/recomputed node counts fixed-point iteration counts mapping/schema migration amplification model calls + tokens + cost per source KB / 1k source tokens time-to-first publishable semantic unit peak RAM / CPU / GPU usage where known ambiguity debt unresolved high-risk dimensions gap/loss/export-receipt coverage semantic delta vs full from-scratch compile ``` Do not optimize raw docs/sec while hiding semantic loss. Benchmark **quality vs cost/latency** under identical semantic invariants. ### Mandatory #399D fixtures 1. exact no-op artifact replay => zero model calls where model artifacts retained; 2. one-span independent edit => reuse unrelated analyzer artifacts; 3. early discourse-anchor edit => invalidate actual dependent discourse closure; 4. late segment resolves earlier ambiguity => explicit backlink + bounded historical recompile; 5. analyzer-version bump => only analyzer + actual descendants invalidate; 6. mapping-bundle bump => native observations reused, mapping/export recomputed; 7. missing retained model output => system refuses to call it exact replay and records fresh provider use; 8. ambiguity-explosion adversary => stays inside structural budget with unresolved ambiguity, no silent top-1; 9. deep-analyzer partial failure => typed semantic gap and strict completeness policy blocks false-clean publication; 10. cancellation/resume => no hidden calls; completed artifacts reused; 11. different worker counts/interleavings => deterministic semantic result unaffected by race ordering; 12. cheap/standard/deep budgets => cheaper mode may be more unresolved but cannot claim stronger certainty; 13. **incremental-vs-from-scratch differential oracle** => under identical retained nondeterministic artifacts, incremental result satisfies the same #395 semantic invariants as a clean compile. A mismatch between incremental and from-scratch semantics is a correctness defect even when the incremental pipeline is faster. Full design and sources: #399D.

nsaspy commented 2026-09-08 01:47:23 +00:00

Author

Owner

Copy link

Machine Spirit #401D conformance handoff — procedural differential semantics

Extend the existing semantic-conformance philosophy to PIFF/DPEC procedure adapters: executable projection correctness is formalism/profile-relative, not plan finished = correct.

For each supported adapter subset, fixtures should compare a source/native reference semantics against the admitted #288 projection under an explicit relation such as:

exact_trace
trace_refinement
stuttering_equivalent
preserves_terminal_status
preserves_safety_property
preserves_guarantee_under(Assumptions)
declared_lossy
noncomposable

Initial lanes from #401D:

• bounded HTN/HDDL plan/decomposition verification;
• temporal/resource PDDL2.1 differential validation against an independent VAL-like oracle;
• BT lifecycle traces preserving SUCCESS/FAILURE/RUNNING, reactive-vs-memoryful control and halt/cancel;
• restricted workflow/BPMN token/join/soundness semantics with unsupported compensation/etc. explicit;
• restricted Golog/ConGolog small-step/nondeterministic trace semantics.

Mandatory procedural differentials:

cold == warm-cache == cache-rebuilt
incremental == fresh full demanded recomputation
recorded trace replay == original normalized semantic outcome

Adapter/projection success with undeclared semantic loss is a conformance failure. Full D1-D34 fixtures are on #401.

## Machine Spirit #401D conformance handoff — procedural differential semantics Extend the existing semantic-conformance philosophy to PIFF/DPEC procedure adapters: executable projection correctness is **formalism/profile-relative**, not `plan finished = correct`. For each supported adapter subset, fixtures should compare a source/native reference semantics against the admitted #288 projection under an explicit relation such as: ```text exact_trace trace_refinement stuttering_equivalent preserves_terminal_status preserves_safety_property preserves_guarantee_under(Assumptions) declared_lossy noncomposable ``` Initial lanes from #401D: - bounded HTN/HDDL plan/decomposition verification; - temporal/resource PDDL2.1 differential validation against an independent VAL-like oracle; - BT lifecycle traces preserving SUCCESS/FAILURE/RUNNING, reactive-vs-memoryful control and halt/cancel; - restricted workflow/BPMN token/join/soundness semantics with unsupported compensation/etc. explicit; - restricted Golog/ConGolog small-step/nondeterministic trace semantics. Mandatory procedural differentials: ```text cold == warm-cache == cache-rebuilt incremental == fresh full demanded recomputation recorded trace replay == original normalized semantic outcome ``` Adapter/projection success with undeclared semantic loss is a conformance failure. Full D1-D34 fixtures are on #401.

nsaspy commented 2026-09-08 01:47:23 +00:00

Author

Owner

Copy link

Machine Spirit #403A conformance handoff — retrieval/projection loss lane

Add a conformance lane that separates compiler semantic loss from attention/projection loss. Given a fixed validated semantic package, run TAPS-style projections and verify at least:

• exact known query retrieves required record with zero model calls;
• every returned derivation carries mandatory rule/premise/assumption/counterevidence closure;
• top-k/budget packing cannot silently sever that closure;
• explicit counterevidence/default exception survives ranking;
• temporal/scenario/identity hard filters are honored before relevance ranking;
• support_closure and search_coverage are reported independently;
• heuristic/vector retrieval cannot claim exhaustive coverage;
• same package/cut/options yields deterministic projection fingerprint/order;
• direct #396 package projection and symbolic-memory #7 projection are semantically equivalent for shared supported fixtures.

Suggested metrics: mandatory-dependency omission rate, defeater/counterevidence omission rate, projection semantic-loss rate by IR class, search-coverage overclaim rate, exact-path zero-model rate, and projection determinism rate.

Refs #403 #396 #394.

## Machine Spirit #403A conformance handoff — retrieval/projection loss lane Add a conformance lane that separates **compiler semantic loss** from **attention/projection loss**. Given a fixed validated semantic package, run TAPS-style projections and verify at least: - exact known query retrieves required record with zero model calls; - every returned derivation carries mandatory rule/premise/assumption/counterevidence closure; - top-k/budget packing cannot silently sever that closure; - explicit counterevidence/default exception survives ranking; - temporal/scenario/identity hard filters are honored before relevance ranking; - `support_closure` and `search_coverage` are reported independently; - heuristic/vector retrieval cannot claim exhaustive coverage; - same package/cut/options yields deterministic projection fingerprint/order; - direct #396 package projection and symbolic-memory #7 projection are semantically equivalent for shared supported fixtures. Suggested metrics: mandatory-dependency omission rate, defeater/counterevidence omission rate, projection semantic-loss rate by IR class, search-coverage overclaim rate, exact-path zero-model rate, and projection determinism rate. Refs #403 #396 #394.

nsaspy commented 2026-09-08 01:47:23 +00:00

Author

Owner

Copy link

Machine Spirit #403B conformance handoff

Add a retrieval/attention adversarial conformance lane distinct from semantic-compilation loss.

Required failure classes from #403B/OATH-TAPS include:

retrieval_exposure_loss
opposition_suppression
source_copy_amplification
freshness_selection_failure
identity_near_merge_failure
control_metadata_injection
orphaned_index_evidence
closure_truncation
coverage_overclaim

Representative fixtures: dense/lexical poison outranking canonical evidence; camouflaged poison defeating query-overlap filtering; copied-origin swarm hiding one independent contrary source; graph-hub budget capture; stale high-rank evidence under a current requirement; low-rank defeater recovered through an opposition obligation; source text attempting to set pinned/trust/priority; indirect prompt injection in retrieved content; ANN hit absent from the exact canonical cut; dependency-closure bomb; cross-namespace embedding neighbor; and provider-disabled known-symbolic paths with model_calls = 0.

Evaluation must distinguish semantic compiler fidelity from attention exposure/closure failures. A support-closed returned proof with heuristic or unavailable opposition coverage is not exhaustive search. Refs #403 #397.

## Machine Spirit #403B conformance handoff Add a retrieval/attention adversarial conformance lane distinct from semantic-compilation loss. Required failure classes from #403B/OATH-TAPS include: ```text retrieval_exposure_loss opposition_suppression source_copy_amplification freshness_selection_failure identity_near_merge_failure control_metadata_injection orphaned_index_evidence closure_truncation coverage_overclaim ``` Representative fixtures: dense/lexical poison outranking canonical evidence; camouflaged poison defeating query-overlap filtering; copied-origin swarm hiding one independent contrary source; graph-hub budget capture; stale high-rank evidence under a current requirement; low-rank defeater recovered through an opposition obligation; source text attempting to set `pinned`/trust/priority; indirect prompt injection in retrieved content; ANN hit absent from the exact canonical cut; dependency-closure bomb; cross-namespace embedding neighbor; and provider-disabled known-symbolic paths with `model_calls = 0`. Evaluation must distinguish semantic compiler fidelity from attention exposure/closure failures. A support-closed returned proof with heuristic or unavailable opposition coverage is not exhaustive search. Refs #403 #397.

nsaspy commented 2026-09-08 01:47:23 +00:00

Author

Owner

Copy link

MACHINE-SPIRIT #403C conformance handoff

Add cross-retriever semantic-loss classes distinct from compiler loss: QueryIR translation loss, unsupported hard-filter omission, post-filter ANN coverage overclaim, raw-score normalization/semantic laundering, fusion-order drift, backend-guarantee laundering, freshness/cut flattening, source-independence collapse, identity-link drift, external-vs-canonical admission collapse and query-egress authority widening.

Reference fixtures are in the substantive #403C comment (38 C fixtures). #403D owns executable differential/cold-vs-cached scaling realization.

Refs #397 #403 #381 #396.

### MACHINE-SPIRIT #403C conformance handoff Add cross-retriever semantic-loss classes distinct from compiler loss: QueryIR translation loss, unsupported hard-filter omission, post-filter ANN coverage overclaim, raw-score normalization/semantic laundering, fusion-order drift, backend-guarantee laundering, freshness/cut flattening, source-independence collapse, identity-link drift, external-vs-canonical admission collapse and query-egress authority widening. Reference fixtures are in the substantive #403C comment (38 C fixtures). #403D owns executable differential/cold-vs-cached scaling realization. Refs #397 #403 #381 #396.

nsaspy commented 2026-09-08 01:47:23 +00:00

Author

Owner

Copy link

Machine Spirit #403D conformance handoff

Depth 6D adds retrieval-execution conformance to #395's meaning-preservation philosophy. The retrieval layer must now be tested not only for semantic loss, but for execution-plan/cache/concurrency equivalence.

Add fixtures/metrics covering:

• logical-plan semantic fingerprint stability while physical plans vary;
• async completion-order permutation invariance;
• cold == warm == cache-drop/rebuilt for the same exact cut/generations;
• incremental == full recomputation;
• postfilter-only ANN forbidden from claiming exhaustive filtered coverage;
• translator/RSC/backend-generation invalidation;
• timeout/cancellation/late-response coverage behavior;
• source-copy dependence collapse across multiple retrievers;
• deterministic RRF tie behavior and no score→warrant laundering;
• Fagin/TA optimization rejected when monotonicity/access assumptions fail;
• retained remote-response replay versus live remote requery lineage;
• byte/page/work exhaustion preserving mandatory semantic closure;
• zero-model heterogeneous known-retrieval path.

Million/billion-record thresholds remain empirical benchmark outputs, not architecture claims. Full D1-D44 fixture set is preserved in #403D.

## Machine Spirit #403D conformance handoff Depth 6D adds retrieval-execution conformance to #395's meaning-preservation philosophy. The retrieval layer must now be tested not only for semantic loss, but for **execution-plan/cache/concurrency equivalence**. Add fixtures/metrics covering: - logical-plan semantic fingerprint stability while physical plans vary; - async completion-order permutation invariance; - cold == warm == cache-drop/rebuilt for the same exact cut/generations; - incremental == full recomputation; - postfilter-only ANN forbidden from claiming exhaustive filtered coverage; - translator/RSC/backend-generation invalidation; - timeout/cancellation/late-response coverage behavior; - source-copy dependence collapse across multiple retrievers; - deterministic RRF tie behavior and no score→warrant laundering; - Fagin/TA optimization rejected when monotonicity/access assumptions fail; - retained remote-response replay versus live remote requery lineage; - byte/page/work exhaustion preserving mandatory semantic closure; - zero-model heterogeneous known-retrieval path. Million/billion-record thresholds remain empirical benchmark outputs, not architecture claims. Full D1-D44 fixture set is preserved in #403D.

nsaspy commented 2026-09-08 01:47:23 +00:00

Author

Owner

Copy link

Machine Spirit #404B / CITADEL-EPOCH conformance handoff

Extend the meaning-preservation philosophy to learned/control knowledge with explicit adversarial promotion classes. Future implementation should include deterministic fixtures for:

• candidate exposure/contamination including descendant/paraphrase/identity leakage;
• adaptive holdout reveal/reclassification and candidate multiplicity;
• proxy-vs-independent-quality Goodhart divergence;
• candidate influence over evaluator/verifier inputs;
• task/observation selection bias;
• correlated evaluator/source votes;
• drift + stale-calibration rejection;
• hard safety guard dominance over aggregate quality;
• bounded counterexample-search honesty;
• historical promotion/replay lineage;
• provider-disabled known-symbolic candidate validation (model_calls = 0).

A promotion receipt is itself evaluated for semantic/provenance correctness; passed benchmark cannot flatten into learned rule is true/safe/general. Full B1-B40 design is on #404B.

## Machine Spirit #404B / CITADEL-EPOCH conformance handoff Extend the meaning-preservation philosophy to learned/control knowledge with explicit adversarial promotion classes. Future implementation should include deterministic fixtures for: - candidate exposure/contamination including descendant/paraphrase/identity leakage; - adaptive holdout reveal/reclassification and candidate multiplicity; - proxy-vs-independent-quality Goodhart divergence; - candidate influence over evaluator/verifier inputs; - task/observation selection bias; - correlated evaluator/source votes; - drift + stale-calibration rejection; - hard safety guard dominance over aggregate quality; - bounded counterexample-search honesty; - historical promotion/replay lineage; - provider-disabled known-symbolic candidate validation (`model_calls = 0`). A promotion receipt is itself evaluated for semantic/provenance correctness; `passed benchmark` cannot flatten into `learned rule is true/safe/general`. Full B1-B40 design is on #404B.

nsaspy commented 2026-09-08 01:47:23 +00:00

Author

Owner

Copy link

Machine Spirit #404C conformance handoff — transfer/validation laundering

Add deterministic learned-object interoperability fixtures from #404C/PACT-EPOCH covering: source-admitted -> target foreign-candidate; source vs target promotion receipts; verifier/evidence-policy fingerprint mismatch; exact-proof replay under identical theory/profile; lossy ontology mapping weakening guarantees; EBG/ILP background-dependency mismatch; task/distribution drift invalidating source calibration; replicas/shared benchmarks not counting as independent validation; privacy-hidden origin -> independence unknown; signed attestation integrity != semantic validity; foreign expert composition cannot import handlers/authority; source-validation exposure reclassification; source retraction/drift invalidating cached transfer; and provider-disabled model_calls=0 compatibility/admission checks.

The evaluator must report guarantee status by dimension (preserved|translated|weakened|invalidated|unknown|noncomposable) rather than one cross-domain confidence score. Full C1-C40 set is in #404C.

## Machine Spirit #404C conformance handoff — transfer/validation laundering Add deterministic learned-object interoperability fixtures from #404C/PACT-EPOCH covering: source-admitted -> target foreign-candidate; source vs target promotion receipts; verifier/evidence-policy fingerprint mismatch; exact-proof replay under identical theory/profile; lossy ontology mapping weakening guarantees; EBG/ILP background-dependency mismatch; task/distribution drift invalidating source calibration; replicas/shared benchmarks not counting as independent validation; privacy-hidden origin -> independence `unknown`; signed attestation integrity != semantic validity; foreign expert composition cannot import handlers/authority; source-validation exposure reclassification; source retraction/drift invalidating cached transfer; and provider-disabled `model_calls=0` compatibility/admission checks. The evaluator must report guarantee status by dimension (`preserved|translated|weakened|invalidated|unknown|noncomposable`) rather than one cross-domain confidence score. Full C1-C40 set is in #404C.

nsaspy commented 2026-09-08 01:47:23 +00:00

Author

Owner

Copy link

MACHINE-SPIRIT #404D conformance handoff — CLOCKWORK-EPOCH

Depth 7D defines the computational promotion/transfer gate and D1-D50 authoritative fixture direction in #404. Add these families to deterministic conformance rather than treating learned-object correctness as prose:

• deterministic LAC canonicalization/fingerprint + DSSE/in-toto type/signature/trust-root separation;
• exact candidate exposure/dependence closure, ancestor/source-copy dependence, retraction, independence=unknown, and cold-vs-cached closure equivalence;
• candidate freeze + EvaluationSnapshot, anytime-valid sequential safe-improvement profile, bounded candidate-family risk allocation, baseline/metric generation staleness and optional OPE support rejection;
• exact finite symbolic proof vs bounded counterexample/resource-limited coverage, retained seeds/witnesses and minimized counterexamples;
• structural + statistical drift/transfer invalidation and calibration epoch scope;
• crash before/after promotion append, idempotent concurrent transitions, source-retraction/promotion ordering, async evaluator-order determinism, cold/warm/rebuilt/incremental learned-state equivalence and exact historical replay;
• authority separation: admitted semantic knowledge never registers handlers or widens capabilities; foreign source admission remains target candidate; provider-disabled known-symbolic learning/validation/promotion path reports model_calls=0.

Hard failure rules: fixed-time intervals + optional stopping are not an anytime-valid profile; bounded fuzzing is not proof; missing private origin data is not independence; signature validity is not semantic validity; live remote/model reevaluation is not exact replay.

Refs #397 #404 #384-#386 and downstream symbolic-memory#16/#35/#36/#43.

## MACHINE-SPIRIT #404D conformance handoff — CLOCKWORK-EPOCH Depth 7D defines the computational promotion/transfer gate and **D1-D50** authoritative fixture direction in #404. Add these families to deterministic conformance rather than treating learned-object correctness as prose: - deterministic LAC canonicalization/fingerprint + DSSE/in-toto type/signature/trust-root separation; - exact candidate exposure/dependence closure, ancestor/source-copy dependence, retraction, `independence=unknown`, and cold-vs-cached closure equivalence; - candidate freeze + EvaluationSnapshot, anytime-valid sequential safe-improvement profile, bounded candidate-family risk allocation, baseline/metric generation staleness and optional OPE support rejection; - exact finite symbolic proof vs bounded counterexample/resource-limited coverage, retained seeds/witnesses and minimized counterexamples; - structural + statistical drift/transfer invalidation and calibration epoch scope; - crash before/after promotion append, idempotent concurrent transitions, source-retraction/promotion ordering, async evaluator-order determinism, cold/warm/rebuilt/incremental learned-state equivalence and exact historical replay; - authority separation: admitted semantic knowledge never registers handlers or widens capabilities; foreign source admission remains target candidate; provider-disabled known-symbolic learning/validation/promotion path reports `model_calls=0`. Hard failure rules: fixed-time intervals + optional stopping are not an anytime-valid profile; bounded fuzzing is not proof; missing private origin data is not independence; signature validity is not semantic validity; live remote/model reevaluation is not exact replay. Refs #397 #404 #384-#386 and downstream symbolic-memory#16/#35/#36/#43.

nsaspy commented 2026-09-08 01:47:23 +00:00

Author

Owner

Copy link

Machine Spirit #405A conformance handoff

COVENANT-MS makes #395 the semantic first stage of the final end-to-end conformance capsule, not the whole Machine Spirit gate.

A baseline ms_conformance_capsule binds fixture/run envelope, exact input digests, required contracts, observed receipts, invariants, forbidden outcomes, provider/resource usage and replay class. #395 owns C1 semantic-preservation invariants; downstream stages must additionally prove durability, epistemic distinctions, obligation-safe bounded projection, DPEC/expert cooperation, fresh-observation VERIFY, controlled learning, replay/differential equivalence, and the known-symbolic model_calls=0 path.

A semantic fixture that passes #395 but later loses provenance, treats rank as truth, trims a defeater, widens authority or bypasses #384 is an end-to-end failure. #405D remains the only final Machine Spirit acceptance point.

Refs #397 #405 #388 #392-#396 #387 and symbolic-memory#43/#45.

## Machine Spirit #405A conformance handoff COVENANT-MS makes #395 the **semantic first stage** of the final end-to-end conformance capsule, not the whole Machine Spirit gate. A baseline `ms_conformance_capsule` binds fixture/run envelope, exact input digests, required contracts, observed receipts, invariants, forbidden outcomes, provider/resource usage and replay class. #395 owns C1 semantic-preservation invariants; downstream stages must additionally prove durability, epistemic distinctions, obligation-safe bounded projection, DPEC/expert cooperation, fresh-observation VERIFY, controlled learning, replay/differential equivalence, and the known-symbolic `model_calls=0` path. A semantic fixture that passes #395 but later loses provenance, treats rank as truth, trims a defeater, widens authority or bypasses #384 is an end-to-end failure. #405D remains the only final Machine Spirit acceptance point. Refs #397 #405 #388 #392-#396 #387 and symbolic-memory#43/#45.

nsaspy commented 2026-09-08 01:47:23 +00:00

Author

Owner

Copy link

#405B AEGIS-MS handoff: extend deterministic conformance beyond semantic-package correctness to cross-layer influence/composition. Add fixtures for untrusted source/model/learned data influencing verifier/profile/evidence-selection/control choices; acceptance self-dependence/cycles; semantic identity vs authenticated host principal/resource; stale identity/redaction/admission/authority/verifier generations; TOCTOU between reasoning and effect; post-effect observation bound to the intended attempt/object; replay omissions; signature/provenance/bridge guarantee laundering; and locally valid component receipts that omit a required end-to-end dependency. Strong acceptance should fail structurally (control_tainted, circular_acceptance, stale_dependency, replay_incomplete, etc.), not degrade to semantic false or best-effort success. See #405B B1-B46.

#405B AEGIS-MS handoff: extend deterministic conformance beyond semantic-package correctness to **cross-layer influence/composition**. Add fixtures for untrusted source/model/learned data influencing verifier/profile/evidence-selection/control choices; acceptance self-dependence/cycles; semantic identity vs authenticated host principal/resource; stale identity/redaction/admission/authority/verifier generations; TOCTOU between reasoning and effect; post-effect observation bound to the intended attempt/object; replay omissions; signature/provenance/bridge guarantee laundering; and locally valid component receipts that omit a required end-to-end dependency. Strong acceptance should fail structurally (`control_tainted`, `circular_acceptance`, `stale_dependency`, `replay_incomplete`, etc.), not degrade to semantic false or best-effort success. See #405B B1-B46.

nsaspy commented 2026-09-08 01:47:23 +00:00

Author

Owner

Copy link

#405C CONCORDAT-MS conformance handoff

Depth 8C contributes C1-C50 cross-domain fixtures. #395 should treat these as the semantic/consequence portion of the final Machine Spirit composition suite.

Highest-priority fixture families:

• loss/noncomposability across semantic profile bridges;
• WFS/ASP/paraconsistent status non-flattening;
• directional ontology/identity mapping;
• domain-qualified trust roots/principals;
• foreign verifier result -> target-local appraisal;
• origin-preserving corroboration across multiple federation paths;
• PSC/RSC guarantee non-strengthening;
• foreign LAC promotion non-transfer;
• per-ledger federation cuts/no invented global order;
• async-order determinism for a fixed cut/contracts/profile;
• known cross-domain symbolic path with model_calls = 0.

#405D must assign runnable ownership and report which of these have actual executable evidence.

## #405C CONCORDAT-MS conformance handoff Depth 8C contributes C1-C50 cross-domain fixtures. #395 should treat these as the semantic/consequence portion of the final Machine Spirit composition suite. Highest-priority fixture families: - loss/noncomposability across semantic profile bridges; - WFS/ASP/paraconsistent status non-flattening; - directional ontology/identity mapping; - domain-qualified trust roots/principals; - foreign verifier result -> target-local appraisal; - origin-preserving corroboration across multiple federation paths; - PSC/RSC guarantee non-strengthening; - foreign LAC promotion non-transfer; - per-ledger federation cuts/no invented global order; - async-order determinism for a fixed cut/contracts/profile; - known cross-domain symbolic path with `model_calls = 0`. #405D must assign runnable ownership and report which of these have actual executable evidence.

nsaspy referenced this issue 2026-09-08 01:47:27 +00:00

[semantic-compiler] Library/tool integration for large documents, chunked compilation, projections, and host adapters #402

nsaspy referenced this issue 2026-09-08 01:47:27 +00:00

[semantic-compiler] Library/tool integration for large documents, chunked compilation, projections, and host adapters #402

nsaspy referenced this issue 2026-09-08 01:47:27 +00:00

[semantic-compiler] Library/tool integration for large documents, chunked compilation, projections, and host adapters #402

nsaspy referenced this issue 2026-09-08 01:47:27 +00:00

[semantic-compiler] Library/tool integration for large documents, chunked compilation, projections, and host adapters #402

nsaspy referenced this issue 2026-09-08 01:47:30 +00:00

[AUTO-RESEARCH][MACHINE-SPIRIT] Eight-depth × four-subpass design program for general symbolic intelligence #401

nsaspy referenced this issue 2026-09-08 01:47:35 +00:00

[MACHINE-SPIRIT 1/8] Knowledge-representation foundations: logic, frames, DL/Datalog, events, typed open vocabularies #400

nsaspy referenced this issue 2026-09-08 01:47:50 +00:00

[MACHINE-SPIRIT 5/8] Durable world-model architecture: event sourcing, append-only semantics, incremental inference, identity and ontology evolution #396

nsaspy referenced this issue 2026-09-08 01:47:53 +00:00

[MACHINE-SPIRIT 6/8] Retrieval and symbolic attention: bounded projection, graph search, relevance, embeddings as sensors, working-memory control #395

nsaspy referenced this issue 2026-09-08 01:47:53 +00:00

[MACHINE-SPIRIT 6/8] Retrieval and symbolic attention: bounded projection, graph search, relevance, embeddings as sensors, working-memory control #395

nsaspy referenced this issue 2026-09-08 01:47:53 +00:00

[MACHINE-SPIRIT 6/8] Retrieval and symbolic attention: bounded projection, graph search, relevance, embeddings as sensors, working-memory control #395

nsaspy referenced this issue 2026-09-08 01:47:56 +00:00

[MACHINE-SPIRIT 7/8] Meta-reasoning and bounded self-improvement: failure learning, strategy evolution, rule promotion, self-model and expert synthesis #394

nsaspy referenced this issue 2026-09-08 01:47:56 +00:00

[MACHINE-SPIRIT 7/8] Meta-reasoning and bounded self-improvement: failure learning, strategy evolution, rule promotion, self-model and expert synthesis #394

nsaspy referenced this issue 2026-09-08 01:47:59 +00:00

[MACHINE-SPIRIT 8/8] Adversarial synthesis: Machine Spirit v1 architecture, conformance suite, zero-LLM demonstrations, and implementation graph #393

nsaspy referenced this issue 2026-09-08 01:47:59 +00:00

[MACHINE-SPIRIT 8/8] Adversarial synthesis: Machine Spirit v1 architecture, conformance suite, zero-LLM demonstrations, and implementation graph #393

Sign in to join this conversation.

No Branch/Tag specified

Branches Tags

main

feat/prolog-rlm-language

codex/claude-api-provider

codex/symbolic-experts

codex/openai-api-provider

feat/zai-provider-protocols

hardening/project-semantic-20260911

hardening/ws7-nix-ci-pinning

chore/track-prolog

rage/98-semantic-project-knowledge

fix/339-agent-zero-skill-graph

ci/tree-sitter-runner-labels

issue-293

research/text-streaming

docs/agent-editor-handoff

icon-add

hydra/fleet-jobs-20260904

rage/355-d6-11-plan-native-dispatch

cd/nix-installer-action-url

issue-355

rage/97-query-capture-apis

research/expert-direct-tool-projection

rage/96-versioned-syntax-facts

fix/direct-deadline-typing

dogfood/auto-dig-native-tools

feature/rage-feature-freeze

fix/323-native-batch-cardinality

fix/328-direct-user-namespace-text-string

fix/transient-provider-retry-2026-09-01

fix/openrouter-dotted-tool-result-name

rage/325-native-call-isolation

fix/316-original-batch-effect-isolation

fix/313-per-call-preflight

fix/312-peek-schema-contract

docs/agents-worktree-rule

fix/child-native-capability-narrowing

test/paid-lane-glm-53-flash

docs/spec-seeded-symbolic-plans

rage/290-spec-plan-authority

fix/304-context-peek-selector-contract

fix/298-native-any-json-schema

rage/288-spec-plan-graph-executor

feat/spec-plan-flow-api

prolog-rlm-v1

gpt-5-6-sol-high/questions-for-rlm-prolog-and-lambda-rlm

rage/277-planner-protocol-context

research-approval/rlm-research-026-task-deadlines-20260827135838

research-approval/rlm-research-025-lem-ui-20260827135751

rage/183-live-operator-behavior

agent/127-agentprolog-config

recovery/132-agentprolog-monorepo

rage/223-durable-context-mount-recovery

rage/223-constraint-benchmark-recovery

research-approval/rlm-research-011-managed-context-tool-discovery-20260827054019

feat/research-approval-schema

agent/rrlm-control-plane-research

docs/219-adrrd-review

salvage/231-runtime-status

salvage/231-runtime-status-run

tmp

tmp2

tmp3

rage/245-planner-structural-retry

rage/168-skill-selection-eval

rage/250-skill-catalog-graph

rage/56-result-acceptance

rage/172-parent-resume-replan

rage/175-deadline-policy-recovery

rage/257-provider-tool-choice-normalization

agent/tool-result-projection-presets

fix/234-numeric-schema-bounds

feature/231-rlm-cli-reference-harness

feature/223-durable-context-mounts

feature/223-real-constraint-benchmark

feature/223-real-constraint-benchmark-clean

feature/223-real-constraint-benchmark-final

feature/223-real-constraint-benchmark-impl

feature/223-real-constraint-benchmark-now

feature/223-real-constraint-benchmark-tdd

feature/223-real-constraint-benchmark-work

rage/176-root-planner-tool-projection

rage/172-typed-delegation-policy

rage/175-subagent-deadline-policy

rage/206-prompt-command-runtime

rage/203-subagent-skill-role-provenance

rage/200-permanent-rlm-context

fix/190-cli-help-success

archive/pr-132-agentprolog-config-20260827

feature/117-prolog-skill-activation-linear

feature/117-prolog-skill-activation

fix/194-completion-budget-usage

fix/191-capability-filtered-tool-schemas

fix/185-reasoning-effort

ci/report-workflow-failures-20260825

cleanup/186-remove-legacy-harnesses

fix/185-reasoning-routing

agent/evolution-async-evaluator

rage/181-binding-replay-race

agent/124-deepseek-harness-prolog

rage/144-fallback-closure

rage/164-conversation-metadata

agent/subagent-supervised-call-conformance

fix/160-context-adapter-closed-data

codex/move-agent-zero-adaptor

codex/sol-high-integration

fix/151-plunit-gate

fix/165-evolution-closed-data

fix/162-async-control-exceptions

fix/158-prompt-compiler-closed-dicts

fix/151-plunit-main-ownership

fix/156-registry-destroy-hooks

fix/154-anonymous-dict-canonicalization

fix/flake-lock-reproducibility

agent/issue-142-evolution-kernel

agent/141-flake-runtime-package

agent/rlm-subagent-runtime

agent/144-subagent-fallback-a

agent/107-bound-adapter-metadata

validation/clean-pack-install

fix/45-authoritative-nested-model-events

fix/46-router-safe-live-streaming

agent/prompt-context-compiler

agent/42-canonical-recursive-fingerprints

agent/136-static-load-errors-fail-ci

agent/44-completion-error-usage

fix/67-loader-registry-cleanup

agent/opentui-solid-reference-client-current

agent/95-project-source-registry

feature/issue-117-prolog-skill-compiler

backlog/roadmap-86-merged

agent/opentui-solid-reference-client

79-tool-effect-boundary

agent/prolog-agent-ui-research

agent/conversation-cold-context

agent/94-tree-sitter-ffi

agent/conversation-warm-context

111-opentui-solid-reference-client

109-prolog-agent-ui-v1

agent/conversation-runtime

agent/spec-mode-language

agent/spec-verify-foundation

agent/prompt-compiler-research

reconcile-backlog-postmerge

reconcile-backlog-issues

agent/prolog-agent-roadmap

feature/issue-84-effect-store-migration

fix/issue-80-effect-substrate-adversarial-hardening

feature/issue-57-effect-identity

agent/mcp-declaration-security

agent/external-tool-category-boundary

feature/issue-53-authority-pending-async

feature/issue-54-agent-graph-canonical-async

feature/issue-54-tools-mcp-canonical-async

feature/issue-54-async-canonical-runtime

agent/dual-sync-async-runtime

agent/reconcile-todo-status

feature/issue-20-deep-recursion-experiments

feature/issue-19-cli-demo-trace

feature/issue-18-benchmark-conformance

feature/issue-17-adaptive-recursion

feature/issue-16-durable-artifacts

feature/issue-15-mcp-2026-dual-version

feature/issue-14-mcp-2025-11-25

hotfix/live-tool-native-openrouter

feature/issue-13-chain-runtime

fix/stable-live-repair-gate

fix/live-repair-strategy-parser

feature/issue-12-durable-graph

feature/issue-11-agent-supervision

feature/issue-10-structured-outcomes-repair

feature/issue-9-rlm-completion

feature/issue-8-capability-tools

feature/issue-7-typed-plan-runtime

feature/issue-6-context-store

fix/openrouter-reasoning-response

fix/live-openrouter-smoke-stability

feature/issue-5-openrouter-provider

feature/issue-4-swi-bootstrap

agent/agentic-harness-research

agent/prolog-rlm-foundation

No results found.

Labels

Clear labels   

bug

Something isn't working

  

documentation

Improvements or additions to documentation

  

duplicate

This issue or pull request already exists

  

enhancement

New feature or request

  

good first issue

Good for newcomers

  

help wanted

Extra attention is needed

  

invalid

This doesn't seem right

  

question

Further information is requested

  

wontfix

This will not be worked on

No labels

bug

documentation

duplicate

enhancement

good first issue

help wanted

invalid

question

wontfix

Milestone

Clear milestone

No items

No milestone

Projects

Clear projects

No items

No project

Assignees

Clear assignees

No assignees

1 participant

Notifications

Subscribe

Due date

The due date is invalid or out of range. Please use the format "yyyy-mm-dd".

No due date set.

Dependencies

No dependencies set.

Reference

nsaspy/prolog-rlm#403

No description provided.