G07: Define explicit local outcome export consumption and cautious task-level cost calibration #177

Open
opened 2026-10-04 22:20:24 +00:00 by kilo · 5 comments
Member

Goal

Product owner: Scarcity Router (local consumer), Kernel outcome producer; G07/G12/G13. Source: Creatidy shared architecture v1.0 (2026-10-04), sections 1-5, 7-13, 14-18 supplied in #173. Enable user-authorized local outcome analysis correlated to Router decisions/calls, with honest accepted-result economics and a bounded calibration decision.

Why and Evidence

Router develop 8d9d4b04bcb23fe19ff702b6209fbcb1537cdf1b: gateway_contracts.UsageAccounting separates reported/estimated/unavailable; gateway_audit.AuditRecord has request/decision/target/adapter/call_count but no accepted-task outcome, review/remediation path or causal attribution. Kernel exports Attempt/operations/outcomes; existing export is not an agreed Router consumer contract. docs/selection-policy.md runtime feedback is merely deferred. Raw success rates, quota balance deltas and trace records cannot establish task-caused spend or capability superiority.

Scope

Research and freeze an explicit local export/import correlation contract and minimal analysis workflow, then implement supported Router consumption once producer evidence is available. Kernel owns Task/Attempt/accepted result, total path (solver/review/remediation/retries/owner attention); Router supplies source/call facts and unknowns. Specify consent, dedup/version/time/identity, retention and exclusion of prompt/secret content. Define calibration evidence requirements and human approval, not an autonomous learner.

Acceptance Criteria

  1. Version-pinned synthetic producer-consumer fixture associates one task's multiple Attempts, authorized role decisions and observed internal calls to final accepted/rejected/unknown outcome without merging identities or inventing missing usage.
  2. Re-import/reordered/duplicate exports are deterministic and deduplicated; incomplete history, mismatched decisions, schema drift, unknown or overlapping usage remain explicit. A quota delta while other clients use a pool is not assigned to this task; token usage is not converted automatically to subscription quota.
  3. Local report distinguishes token rate, observed/estimated call usage, request limit and full path to accepted result; request traces are not a financial ledger. Unknown effort/model/hidden helper calls stay unknown, not zero.
  4. No automatic central upload, MI private telemetry feed, credential/prompt retention or default learning. User explicitly authorizes each supported local export/import path; malicious/oversized/malformed exports fail safely without side effects.
  5. Calibration proposal retains sample/task/harness/version/configuration context, uncertainty/confounding and provenance. Raw percentages never auto-rank models; any accepted capability change requires owner-reviewable evidence/date/confidence/rationale/diff and unchanged hard minima. The research may conclude insufficient evidence, with concrete next evidence/decision rather than fake learning.
  6. Include migration/backward compatibility, safe diagnostics, operator docs, full deterministic gates and outcome-specific negative tests.

Dependencies, Priority and Reuse

After G04 decision/Attempt identity and G12 honest call-fact boundaries; contract research can proceed before them. Kernel producer counterpart not found at registration; known export foundation Creatidy/creatidy-kernel#37. MI publication remains separate; no Kernel implementation here. Refs #173 and D-043 audit.
Reuse existing AuditRecord, UsageAccounting, Kernel export and strict JSON validation. OpenTelemetry/Langfuse patterns are candidates for export/views only, not approved dependencies or accounting authority: compare exact revisions/files/license, semantics, privacy, negative tests, operational/dependency cost and maintenance before adoption. Future review requires representative user-approved local exports; historical/live inference or effects require separate authority. No inference required to prove the synthetic import contract.

## Goal **Product owner: Scarcity Router (local consumer)**, Kernel outcome producer; G07/G12/G13. Source: Creatidy shared architecture v1.0 (2026-10-04), sections 1-5, 7-13, 14-18 supplied in #173. Enable user-authorized local outcome analysis correlated to Router decisions/calls, with honest accepted-result economics and a bounded calibration decision. ## Why and Evidence Router develop `8d9d4b04bcb23fe19ff702b6209fbcb1537cdf1b`: `gateway_contracts.UsageAccounting` separates reported/estimated/unavailable; `gateway_audit.AuditRecord` has request/decision/target/adapter/call_count but no accepted-task outcome, review/remediation path or causal attribution. Kernel exports Attempt/operations/outcomes; existing export is not an agreed Router consumer contract. `docs/selection-policy.md` runtime feedback is merely deferred. Raw success rates, quota balance deltas and trace records cannot establish task-caused spend or capability superiority. ## Scope Research and freeze an explicit local export/import correlation contract and minimal analysis workflow, then implement supported Router consumption once producer evidence is available. Kernel owns Task/Attempt/accepted result, total path (solver/review/remediation/retries/owner attention); Router supplies source/call facts and unknowns. Specify consent, dedup/version/time/identity, retention and exclusion of prompt/secret content. Define calibration evidence requirements and human approval, not an autonomous learner. ## Acceptance Criteria 1. Version-pinned synthetic producer-consumer fixture associates one task's multiple Attempts, authorized role decisions and observed internal calls to final accepted/rejected/unknown outcome without merging identities or inventing missing usage. 2. Re-import/reordered/duplicate exports are deterministic and deduplicated; incomplete history, mismatched decisions, schema drift, unknown or overlapping usage remain explicit. A quota delta while other clients use a pool is not assigned to this task; token usage is not converted automatically to subscription quota. 3. Local report distinguishes token rate, observed/estimated call usage, request limit and full path to accepted result; request traces are not a financial ledger. Unknown effort/model/hidden helper calls stay unknown, not zero. 4. No automatic central upload, MI private telemetry feed, credential/prompt retention or default learning. User explicitly authorizes each supported local export/import path; malicious/oversized/malformed exports fail safely without side effects. 5. Calibration proposal retains sample/task/harness/version/configuration context, uncertainty/confounding and provenance. Raw percentages never auto-rank models; any accepted capability change requires owner-reviewable evidence/date/confidence/rationale/diff and unchanged hard minima. The research may conclude insufficient evidence, with concrete next evidence/decision rather than fake learning. 6. Include migration/backward compatibility, safe diagnostics, operator docs, full deterministic gates and outcome-specific negative tests. ## Dependencies, Priority and Reuse After G04 decision/Attempt identity and G12 honest call-fact boundaries; contract research can proceed before them. Kernel producer counterpart not found at registration; known export foundation https://forgejo.creatidy.com/Creatidy/creatidy-kernel/issues/37. MI publication remains separate; no Kernel implementation here. Refs #173 and D-043 audit. Reuse existing `AuditRecord`, `UsageAccounting`, Kernel export and strict JSON validation. OpenTelemetry/Langfuse patterns are candidates for export/views only, not approved dependencies or accounting authority: compare exact revisions/files/license, semantics, privacy, negative tests, operational/dependency cost and maintenance before adoption. Future review requires representative user-approved local exports; historical/live inference or effects require separate authority. No inference required to prove the synthetic import contract.
Author
Member

The registration-time missing local export producer is now Creatidy/creatidy-kernel#58 (versioned private whole-path outcome/full-cost export), supported by Kernel #56 budgets. This issue remains Router's explicitly authorized local consumer/analysis. Creatidy/model-intelligence#17 is only the optional opt-in sharing decision, not a receiving service or automatic upload/learning obligation. Producer identity/cost schema -> deterministic Router import/dedup/correlation -> cautious owner-reviewed analysis; unknown/overlapping/private data remains explicit. Counterparts are OPEN.

The registration-time missing local export producer is now https://forgejo.creatidy.com/Creatidy/creatidy-kernel/issues/58 (versioned private whole-path outcome/full-cost export), supported by Kernel #56 budgets. This issue remains Router's explicitly authorized local consumer/analysis. https://forgejo.creatidy.com/Creatidy/model-intelligence/issues/17 is only the optional opt-in sharing decision, not a receiving service or automatic upload/learning obligation. Producer identity/cost schema -> deterministic Router import/dedup/correlation -> cautious owner-reviewed analysis; unknown/overlapping/private data remains explicit. Counterparts are OPEN.
Author
Member

Fresh /loop contract-research/eligibility evidence after174 producer integration and owner-paused185:174/c38c8ae and184 are actually integrated; neither open state nor185 is invented as this issue's implementation prerequisite. Actual Kernel develop freshly pinned e1aca78b0e56afffb245005e25ecb7b155a8367a and its accepted PR96/20e7f3e/merge98656f6 contract inspected from a transient read-only public clone. No Kernel/private-export execution, data access, mutation, inference or dependency adoption. Source Apache-2.0, no code copied.

Actual source evidence: adapters/task_execution.py:1245-1358 export_task_store emits schema1 Program/status/mode/private task/deadline/base, Attempt/status/terminal candidate, operation status/reference/dispatch claims, allocation runtime/provider/model/effort, optional candidate/acceptance/cancellation and historical PR artifact. It does NOT emit a role-decision/route/internal-call/usage/check/remediation whole-path causal ledger. adapters/reference.py:332-411 is a distinct richer schema1 reference format with explicitly unavailable token usage, not production full-cost export. core/outcomes.py:198-307 emits creatidy-local-outcomes-v1 receipt/correction/Attempt facts, but exact Git call-site search under current src finds only that definition: no durable task-export integration. docs/architecture/successor.md G07/G12 explicitly distinguishes all3 paths, selects no unified schema/redaction/sharing permission, and retains52/56/58 producer work. PR96 changed documentation only; accepted specification is not implementation/full consumer conformance.

Specific external producer condition: Kernel52/56/58 must deliver/agree actual versioned Task->Attempt->authorized role/route/call/usage/acceptance/correction lineage and privacy/rights/negative fixture semantics before Router can implement the required FULL correlation/accepted-result economics in AC1-3. Current thin exports cannot prove observed calls or task-caused costs; manual invented sidecar mappings, zero usage, synthetic fields outside a producer contract or a reference-only format relabeled production would violate scope. Owner-approved real exports are not inferred from source access; default local JSON remains unredacted/private, MI17 sharing not required or authorized.

Research stage remains available and this bounded source check is completed; full consumer implementation is temporarily ineligible on actual missing producer/contract fields, not merely Kernel58 OPEN. No substantive177 change/branch/PR exists, so preserve issue/history and return nonterminally to fresh SELECT per post-selection rule.175/179/174 delivered semantics may independently unblock181; paused185 does not stall unrelated work. No new owner architecture question or speculative issue/unused parser is manufactured.

Fresh /loop contract-research/eligibility evidence after174 producer integration and owner-paused185:174/c38c8ae and184 are actually integrated; neither open state nor185 is invented as this issue's implementation prerequisite. Actual Kernel develop freshly pinned `e1aca78b0e56afffb245005e25ecb7b155a8367a` and its accepted PR96/20e7f3e/merge98656f6 contract inspected from a transient read-only public clone. No Kernel/private-export execution, data access, mutation, inference or dependency adoption. Source Apache-2.0, no code copied. Actual source evidence: adapters/task_execution.py:1245-1358 export_task_store emits schema1 Program/status/mode/private task/deadline/base, Attempt/status/terminal candidate, operation status/reference/dispatch claims, allocation runtime/provider/model/effort, optional candidate/acceptance/cancellation and historical PR artifact. It does NOT emit a role-decision/route/internal-call/usage/check/remediation whole-path causal ledger. adapters/reference.py:332-411 is a distinct richer schema1 reference format with explicitly unavailable token usage, not production full-cost export. core/outcomes.py:198-307 emits creatidy-local-outcomes-v1 receipt/correction/Attempt facts, but exact Git call-site search under current src finds only that definition: no durable task-export integration. docs/architecture/successor.md G07/G12 explicitly distinguishes all3 paths, selects no unified schema/redaction/sharing permission, and retains52/56/58 producer work. PR96 changed documentation only; accepted specification is not implementation/full consumer conformance. Specific external producer condition: Kernel52/56/58 must deliver/agree actual versioned Task->Attempt->authorized role/route/call/usage/acceptance/correction lineage and privacy/rights/negative fixture semantics before Router can implement the required FULL correlation/accepted-result economics in AC1-3. Current thin exports cannot prove observed calls or task-caused costs; manual invented sidecar mappings, zero usage, synthetic fields outside a producer contract or a reference-only format relabeled production would violate scope. Owner-approved real exports are not inferred from source access; default local JSON remains unredacted/private, MI17 sharing not required or authorized. Research stage remains available and this bounded source check is completed; full consumer implementation is temporarily ineligible on actual missing producer/contract fields, not merely Kernel58 OPEN. No substantive177 change/branch/PR exists, so preserve issue/history and return nonterminally to fresh SELECT per post-selection rule.175/179/174 delivered semantics may independently unblock181; paused185 does not stall unrelated work. No new owner architecture question or speculative issue/unused parser is manufactured.
Author
Member

Fresh producer reassessment2026-10-28 @SRdevelop2002d348a8d6d5f3d1b742a93ba3adf944d4b24e, canonical KerneldevelopPIN b1c9ffca16f06596e5c3afdca3b9893cf7422d8d (PR110/workclaims/resourceownership). Read-only public-source evidence; no export/privateuserdata execution, codecopy/dependency or crossrepo write. Corrects stale parts of comment18801 rather than inferring readiness from openissue state.

Changed evidence: installed task_execution.export_task_store:1667–1785 still emits legacy schema1 Program/Attempt/operation/allocation/acceptance/cancellation/PR facts, NOW also work_budget=store.work_budget_status(PROGRAM_ID). sqlite_store:3086–3158 durably attributes/corrects unitful receipts to original Attempt/Operation and exports effective_usage/history/totals with coverage='observed receipts only; hidden external-client work unverified', missing_usage='unknown', hard_inference_cap=false. These real installed facts must not be called absent. Reference export remains a DISTINCT richer identity-enforced schema1 path. Current installedCLI task export dispatches to export_task_store; pure core.outcomes.export_outcomes remains creatidy-local-outcomes-v1, called only in its core tests, not the installed task exporter (fullsrc/tests callsite search plus parent direct canonical source checks).

Still unmet for accepted FULL177 AC1–3: no agreed installed wholepath producer/consumer contract joining authorized role/route/decision/internal-call/candidate/review/remediation/acceptance/correction lineage. Current successor.md:126–154 explicitly separates all3 paths and retains task-export callsite/unifiedschema/route-call-candidate identity/per-task/private-content policy gaps; its pure-core shape must not be relabeled installed. Legacythin exports and newbudget receipts can report their actual limited facts, but do not establish the required causal accepted-result economics. No invented sidecar mappings, implicit zeros or unused interim parser will be introduced to mask that gap.

Corrected dependency scope from CURRENT Kernel58 activebody/ADR0012:56/106/accepted54 underlie58;52 enriches STRICTprofile facts only and is NOT a blanket basic-local-export gate. MI17 operational sharing is rejected/closed and NOT an export prerequisite or permission; Router177 remains optional downstreamconsumer. The exact outstanding condition is Kernel58's implemented+documented versioned local wholepath export + producer-negative/privacy fixtures (owner:Kernelproducer maintainer; stage:fullconsumer implementation/conformance, not Router contract research or upstreamKernelimplementation). Explicit owner-approved realexports/rights remain separately required; source access grants none.

Research/currentness step complete; no substantive177source/branch/PR exists. Fullconsumer remains temporarily producer-gated on evidenced missing contract fields, not old52blanketgate, MIsharing or issueOPEN. Return nonterminally to fullfresh SR SELECT; unrelated active224/226 must still be evaluated. No new ownerarchitecture decision, issue, inference or publication is manufactured.

Fresh producer reassessment2026-10-28 @SRdevelop2002d348a8d6d5f3d1b742a93ba3adf944d4b24e, canonical KerneldevelopPIN b1c9ffca16f06596e5c3afdca3b9893cf7422d8d (PR110/workclaims/resourceownership). Read-only public-source evidence; no export/privateuserdata execution, codecopy/dependency or crossrepo write. Corrects stale parts of comment18801 rather than inferring readiness from openissue state. Changed evidence: installed task_execution.export_task_store:1667–1785 still emits legacy schema1 Program/Attempt/operation/allocation/acceptance/cancellation/PR facts, NOW also work_budget=store.work_budget_status(PROGRAM_ID). sqlite_store:3086–3158 durably attributes/corrects unitful receipts to original Attempt/Operation and exports effective_usage/history/totals with coverage='observed receipts only; hidden external-client work unverified', missing_usage='unknown', hard_inference_cap=false. These real installed facts must not be called absent. Reference export remains a DISTINCT richer identity-enforced schema1 path. Current installedCLI task export dispatches to export_task_store; pure core.outcomes.export_outcomes remains creatidy-local-outcomes-v1, called only in its core tests, not the installed task exporter (fullsrc/tests callsite search plus parent direct canonical source checks). Still unmet for accepted FULL177 AC1–3: no agreed installed wholepath producer/consumer contract joining authorized role/route/decision/internal-call/candidate/review/remediation/acceptance/correction lineage. Current successor.md:126–154 explicitly separates all3 paths and retains task-export callsite/unifiedschema/route-call-candidate identity/per-task/private-content policy gaps; its pure-core shape must not be relabeled installed. Legacythin exports and newbudget receipts can report their actual limited facts, but do not establish the required causal accepted-result economics. No invented sidecar mappings, implicit zeros or unused interim parser will be introduced to mask that gap. Corrected dependency scope from CURRENT Kernel58 activebody/ADR0012:56/106/accepted54 underlie58;52 enriches STRICTprofile facts only and is NOT a blanket basic-local-export gate. MI17 operational sharing is rejected/closed and NOT an export prerequisite or permission; Router177 remains optional downstreamconsumer. The exact outstanding condition is Kernel58's implemented+documented versioned local wholepath export + producer-negative/privacy fixtures (owner:Kernelproducer maintainer; stage:fullconsumer implementation/conformance, not Router contract research or upstreamKernelimplementation). Explicit owner-approved realexports/rights remain separately required; source access grants none. Research/currentness step complete; no substantive177source/branch/PR exists. Fullconsumer remains temporarily producer-gated on evidenced missing contract fields, not old52blanketgate, MIsharing or issueOPEN. Return nonterminally to fullfresh SR SELECT; unrelated active224/226 must still be evaluated. No new ownerarchitecture decision, issue, inference or publication is manufactured.
Author
Member

Owner-directed development stop — 2026-10-28. No #177 implementation or PR was started. The bounded read-only reassessment remains recorded in comment #19692 against Router 2002d348 and pinned Kernel b1c9ffca16f06596e5c3afdca3b9893cf7422d8d: current producer now exports durable work-budget and observed-receipt facts, so earlier statements that those facts were wholly absent are stale; the accepted installed whole-path role/route/decision/internal-call/candidate/review/remediation/acceptance/correction correlation contract and privacy/rights fixtures remain unproven. Kernel #58's accepted scope is not gated by blanket #52 or rejected MI #17. The full consumer remains producer-contract/evidence gated. Further work is paused at owner direction; issue remains open and no successor was manufactured.

**Owner-directed development stop — 2026-10-28.** No #177 implementation or PR was started. The bounded read-only reassessment remains recorded in comment #19692 against Router `2002d348` and pinned Kernel `b1c9ffca16f06596e5c3afdca3b9893cf7422d8d`: current producer now exports durable work-budget and observed-receipt facts, so earlier statements that those facts were wholly absent are stale; the accepted installed whole-path role/route/decision/internal-call/candidate/review/remediation/acceptance/correction correlation contract and privacy/rights fixtures remain unproven. Kernel #58's accepted scope is not gated by blanket #52 or rejected MI #17. The full consumer remains producer-contract/evidence gated. Further work is paused at owner direction; issue remains open and no successor was manufactured.
Author
Member

Current owner triage, 2026-11-01: P3, producer-dependent. Kernel #58 remains open. Current Kernel develop fce1c0d4d63fdf4ca0673512c9b9d2512f0ac023 installed export_task_store emits real work-budget and some Attempt/operation/allocation/acceptance facts; these must not be described as absent. It does not establish the agreed installed whole-path role/route/decision/internal-call/candidate/review/remediation/acceptance/correction lineage and privacy/negative-fixture contract required by Router's full consumer. Kernel owns that producer contract; it gates full #177 consumer implementation/conformance, not independent Router contract research. #52 enriches strict-profile evidence, not a blanket basic-export gate; MI sharing is not a prerequisite. Explicit consent/rights for real exports remain separate. No implementation is started and prior pause/history remain intact.

**Current owner triage, 2026-11-01: P3, producer-dependent.** Kernel #58 remains open. Current Kernel develop `fce1c0d4d63fdf4ca0673512c9b9d2512f0ac023` installed `export_task_store` emits real work-budget and some Attempt/operation/allocation/acceptance facts; these must not be described as absent. It does not establish the agreed installed whole-path role/route/decision/internal-call/candidate/review/remediation/acceptance/correction lineage and privacy/negative-fixture contract required by Router's full consumer. Kernel owns that producer contract; it gates full #177 consumer implementation/conformance, not independent Router contract research. #52 enriches strict-profile evidence, not a blanket basic-export gate; MI sharing is not a prerequisite. Explicit consent/rights for real exports remain separate. No implementation is started and prior pause/history remain intact.
Sign in to join this conversation.
No milestone
No project
No assignees
1 participant
Notifications
Due date
The due date is invalid or out of range. Please use the format "yyyy-mm-dd".

No due date set.

Reference
Creatidy/scarcity-router#177
No description provided.