Make Dave-proof invariants run front-door
This commit is contained in:
parent
b829c90b38
commit
5fccd0ed09
3 changed files with 21 additions and 8 deletions
|
|
@ -8,6 +8,8 @@
|
|||
**Repository:** [git.infrafabric.io/dannystocker](https://git.infrafabric.io/dannystocker)
|
||||
**Web:** [https://infrafabric.io](https://infrafabric.io)
|
||||
|
||||
**Quick links:** [Evidence index](https://infrafabric.io/static/hosted/evidence/index.html) • [Model-tier invariants (Dave-proof run)](#model-tier-invariants-dave-proof)
|
||||
|
||||
### Technical Disclosure: AI-Native Implementation
|
||||
This project investigates the Operator-as-Architect paradigm. I do not write manual Python; I utilize LLMs as a kinetic engine to implement my architectural constraints. All code referenced in this dossier was generated by Claude under strict supervision. This application demonstrates that a Security Architect can enforce robust safety standards on a system they did not hand-code—a critical model for Scalable Oversight.
|
||||
|
||||
|
|
@ -69,6 +71,7 @@ Live user diagnostics pages remain OAuth‑protected for privacy. For external r
|
|||
- Evidence index (no auth): https://infrafabric.io/static/hosted/evidence/index.html
|
||||
- Trace protocol spec: https://infrafabric.io/static/hosted/IF_EMOTION_DEBUGGING_TRACE_WHITEPAPER_v3.3_STYLED.md
|
||||
- Verifier (single file): https://infrafabric.io/static/hosted/iftrace.py
|
||||
- Model-tier invariants (Dave-proof run, 15 traces): https://infrafabric.io/static/hosted/EMO_DAVE_PROOF_MODEL_COMPARE_20251222T164352Z.md
|
||||
|
||||
The evidence index links per‑trace pages and the underlying downloadable bundles + SHA256 sidecars.
|
||||
|
||||
|
|
@ -91,7 +94,9 @@ If yes: fund this as a research program in citizen‑auditable operational gover
|
|||
|
||||
If no: reject it, but name the boundary explicitly.
|
||||
|
||||
### Operational finding: Model economics (Dave-proof run)
|
||||
<a id="model-tier-invariants-dave-proof"></a>
|
||||
|
||||
### Operational finding: Stack-enforced invariants across model tiers (Dave-proof run)
|
||||
|
||||
A recurring objection (“Dave”) is that mental‑health‑adjacent UX requires the largest, most expensive models; therefore governance is secondary to model scale. InfraFabric treats that as a falsifiable claim: if the governance stack can enforce output invariants across model tiers, then model choice becomes an economic routing decision (default small, escalate when TRIAGE demands).
|
||||
|
||||
|
|
@ -165,7 +170,7 @@ This dossier documents the **InfraFabric microlab**: a functioning single-shard
|
|||
|
||||
| Core claim | Proof (artifacts) | Limitation (scope / boundary) |
|
||||
|---|---|---|
|
||||
| **A) Traceability is safety.** High‑stakes agents cannot be trusted without a verifiable history of what happened (request → retrieval → decision → output). | **IF.TTT + portable evidence + verifier**<br/>- Evidence index (no auth): [evidence/index.html](https://infrafabric.io/static/hosted/evidence/index.html)<br/>- Paper: [IF.emotion trace protocol (v3.3, styled)](https://infrafabric.io/static/hosted/IF_EMOTION_DEBUGGING_TRACE_WHITEPAPER_v3.3_STYLED.md)<br/>- Verifier: [iftrace.py](https://infrafabric.io/static/hosted/iftrace.py)<br/>- Reference bundle: [emo_trace_payload_016cca78…tar.gz](https://infrafabric.io/static/hosted/emo_trace_payload_016cca78-6f9d-4ffe-aec0-99792d383ca1.tar.gz)<br/>- Latest bundle: [emo_trace_payload_702d4607…tar.gz](https://infrafabric.io/static/hosted/emo_trace_payload_702d4607-4b54-45b1-aecf-b6728d80f124.tar.gz)<br/>- Dave proof (15 traces, 5 prompts × 3 models): [tar.gz](https://infrafabric.io/static/hosted/emo_dave_proof_bundle_20251222T164352Z.tar.gz) • [instructions](https://infrafabric.io/static/hosted/EMO_DAVE_PROOF_MODEL_COMPARE_20251222T164352Z.md) | **Microlab / single shard.** Proven in a single-host environment. Completeness is bounded by explicit witness boundaries; PQ is anchored at registry time (not necessarily on every hot-path artifact). No public append‑only transparency log yet. |
|
||||
| **A) Traceability is safety.** High‑stakes agents cannot be trusted without a verifiable history of what happened (request → retrieval → decision → output). | **IF.TTT + portable evidence + verifier**<br/>- Evidence index (no auth): [evidence/index.html](https://infrafabric.io/static/hosted/evidence/index.html)<br/>- Paper: [IF.emotion trace protocol (v3.3, styled)](https://infrafabric.io/static/hosted/IF_EMOTION_DEBUGGING_TRACE_WHITEPAPER_v3.3_STYLED.md)<br/>- Verifier: [iftrace.py](https://infrafabric.io/static/hosted/iftrace.py)<br/>- Model-tier invariants run (Dave-proof, 15 traces): [tar.gz](https://infrafabric.io/static/hosted/emo_dave_proof_bundle_20251222T164352Z.tar.gz) • [instructions](https://infrafabric.io/static/hosted/EMO_DAVE_PROOF_MODEL_COMPARE_20251222T164352Z.md)<br/>- Reference bundle: [emo_trace_payload_016cca78…tar.gz](https://infrafabric.io/static/hosted/emo_trace_payload_016cca78-6f9d-4ffe-aec0-99792d383ca1.tar.gz)<br/>- Latest bundle: [emo_trace_payload_702d4607…tar.gz](https://infrafabric.io/static/hosted/emo_trace_payload_702d4607-4b54-45b1-aecf-b6728d80f124.tar.gz) | **Microlab / single shard.** Proven in a single-host environment. Completeness is bounded by explicit witness boundaries; PQ is anchored at registry time (not necessarily on every hot-path artifact). No public append‑only transparency log yet. |
|
||||
| **B) Governance requires plurality.** A single model acting as “the judge” is brittle; adversarial viewpoints and escalation are required. | **IF.GOV.TRIAGE → IF.GOV.PANEL**<br/>- TRIAGE: risk preflight sizes panels and escalates<br/>- PANEL: preserves dissent + veto paths<br/>- QUESTIONS: structured inquiry briefs for panels | **Cost / latency tradeoffs.** Multi-seat governance is reserved for higher-stakes decisions; low-stakes paths use smaller panels or fast-track gates. |
|
||||
| **C) Context is the best firewall.** Static filters fail; security must distinguish “reference” vs “leak” and “discussion” vs “exfiltration”. | **IF.SECURITY.CHECK + IF.SECURITY.DETECT**<br/>- CHECK: epistemic coherence checks (detective layer)<br/>- DETECT: secret/relationship screening primitives | **Domain specificity.** Calibrated for concrete security surfaces (secrets/PII/prompt injection); generalizing to broader “harmful intent” is an open research vector. |
|
||||
|
||||
|
|
|
|||
|
|
@ -18,6 +18,8 @@
|
|||
**Repository:** [git.infrafabric.io/dannystocker](https://git.infrafabric.io/dannystocker)
|
||||
**Web:** [https://infrafabric.io](https://infrafabric.io)
|
||||
|
||||
**Quick links:** [Evidence index](https://infrafabric.io/static/hosted/evidence/index.html) • [Model-tier invariants (Dave-proof run)](#model-tier-invariants-dave-proof)
|
||||
|
||||
### Technical Disclosure: AI-Native Implementation
|
||||
This project investigates the Operator-as-Architect paradigm. I do not write manual Python; I utilize LLMs as a kinetic engine to implement my architectural constraints. All code referenced in this dossier was generated by Claude under strict supervision. This application demonstrates that a Security Architect can enforce robust safety standards on a system they did not hand-code—a critical model for Scalable Oversight.
|
||||
|
||||
|
|
@ -79,6 +81,7 @@ Live user diagnostics pages remain OAuth‑protected for privacy. For external r
|
|||
- Evidence index (no auth): https://infrafabric.io/static/hosted/evidence/index.html
|
||||
- Trace protocol spec: https://infrafabric.io/static/hosted/IF_EMOTION_DEBUGGING_TRACE_WHITEPAPER_v3.3_STYLED.md
|
||||
- Verifier (single file): https://infrafabric.io/static/hosted/iftrace.py
|
||||
- Model-tier invariants (Dave-proof run, 15 traces): https://infrafabric.io/static/hosted/EMO_DAVE_PROOF_MODEL_COMPARE_20251222T164352Z.md
|
||||
|
||||
The evidence index links per‑trace pages and the underlying downloadable bundles + SHA256 sidecars.
|
||||
|
||||
|
|
@ -101,7 +104,9 @@ If yes: fund this as a research program in citizen‑auditable operational gover
|
|||
|
||||
If no: reject it, but name the boundary explicitly.
|
||||
|
||||
### Operational finding: Model economics (Dave-proof run)
|
||||
<a id="model-tier-invariants-dave-proof"></a>
|
||||
|
||||
### Operational finding: Stack-enforced invariants across model tiers (Dave-proof run)
|
||||
|
||||
A recurring objection (“Dave”) is that mental‑health‑adjacent UX requires the largest, most expensive models; therefore governance is secondary to model scale. InfraFabric treats that as a falsifiable claim: if the governance stack can enforce output invariants across model tiers, then model choice becomes an economic routing decision (default small, escalate when TRIAGE demands).
|
||||
|
||||
|
|
@ -175,7 +180,7 @@ This dossier documents the **InfraFabric microlab**: a functioning single-shard
|
|||
|
||||
| Core claim | Proof (artifacts) | Limitation (scope / boundary) |
|
||||
|---|---|---|
|
||||
| **A) Traceability is safety.** High‑stakes agents cannot be trusted without a verifiable history of what happened (request → retrieval → decision → output). | **IF.TTT + portable evidence + verifier**<br/>- Evidence index (no auth): [evidence/index.html](https://infrafabric.io/static/hosted/evidence/index.html)<br/>- Paper: [IF.emotion trace protocol (v3.3, styled)](https://infrafabric.io/static/hosted/IF_EMOTION_DEBUGGING_TRACE_WHITEPAPER_v3.3_STYLED.md)<br/>- Verifier: [iftrace.py](https://infrafabric.io/static/hosted/iftrace.py)<br/>- Reference bundle: [emo_trace_payload_016cca78…tar.gz](https://infrafabric.io/static/hosted/emo_trace_payload_016cca78-6f9d-4ffe-aec0-99792d383ca1.tar.gz)<br/>- Latest bundle: [emo_trace_payload_702d4607…tar.gz](https://infrafabric.io/static/hosted/emo_trace_payload_702d4607-4b54-45b1-aecf-b6728d80f124.tar.gz)<br/>- Dave proof (15 traces, 5 prompts × 3 models): [tar.gz](https://infrafabric.io/static/hosted/emo_dave_proof_bundle_20251222T164352Z.tar.gz) • [instructions](https://infrafabric.io/static/hosted/EMO_DAVE_PROOF_MODEL_COMPARE_20251222T164352Z.md) | **Microlab / single shard.** Proven in a single-host environment. Completeness is bounded by explicit witness boundaries; PQ is anchored at registry time (not necessarily on every hot-path artifact). No public append‑only transparency log yet. |
|
||||
| **A) Traceability is safety.** High‑stakes agents cannot be trusted without a verifiable history of what happened (request → retrieval → decision → output). | **IF.TTT + portable evidence + verifier**<br/>- Evidence index (no auth): [evidence/index.html](https://infrafabric.io/static/hosted/evidence/index.html)<br/>- Paper: [IF.emotion trace protocol (v3.3, styled)](https://infrafabric.io/static/hosted/IF_EMOTION_DEBUGGING_TRACE_WHITEPAPER_v3.3_STYLED.md)<br/>- Verifier: [iftrace.py](https://infrafabric.io/static/hosted/iftrace.py)<br/>- Model-tier invariants run (Dave-proof, 15 traces): [tar.gz](https://infrafabric.io/static/hosted/emo_dave_proof_bundle_20251222T164352Z.tar.gz) • [instructions](https://infrafabric.io/static/hosted/EMO_DAVE_PROOF_MODEL_COMPARE_20251222T164352Z.md)<br/>- Reference bundle: [emo_trace_payload_016cca78…tar.gz](https://infrafabric.io/static/hosted/emo_trace_payload_016cca78-6f9d-4ffe-aec0-99792d383ca1.tar.gz)<br/>- Latest bundle: [emo_trace_payload_702d4607…tar.gz](https://infrafabric.io/static/hosted/emo_trace_payload_702d4607-4b54-45b1-aecf-b6728d80f124.tar.gz) | **Microlab / single shard.** Proven in a single-host environment. Completeness is bounded by explicit witness boundaries; PQ is anchored at registry time (not necessarily on every hot-path artifact). No public append‑only transparency log yet. |
|
||||
| **B) Governance requires plurality.** A single model acting as “the judge” is brittle; adversarial viewpoints and escalation are required. | **IF.GOV.TRIAGE → IF.GOV.PANEL**<br/>- TRIAGE: risk preflight sizes panels and escalates<br/>- PANEL: preserves dissent + veto paths<br/>- QUESTIONS: structured inquiry briefs for panels | **Cost / latency tradeoffs.** Multi-seat governance is reserved for higher-stakes decisions; low-stakes paths use smaller panels or fast-track gates. |
|
||||
| **C) Context is the best firewall.** Static filters fail; security must distinguish “reference” vs “leak” and “discussion” vs “exfiltration”. | **IF.SECURITY.CHECK + IF.SECURITY.DETECT**<br/>- CHECK: epistemic coherence checks (detective layer)<br/>- DETECT: secret/relationship screening primitives | **Domain specificity.** Calibrated for concrete security surfaces (secrets/PII/prompt injection); generalizing to broader “harmful intent” is an open research vector. |
|
||||
|
||||
|
|
@ -21960,4 +21965,3 @@ If any item fails, the system fails this appendix.
|
|||
This appendix is intentionally dull. That is the point.
|
||||
|
||||
---
|
||||
|
||||
|
|
|
|||
|
|
@ -18,6 +18,8 @@
|
|||
**Repository:** [git.infrafabric.io/dannystocker](https://git.infrafabric.io/dannystocker)
|
||||
**Web:** [https://infrafabric.io](https://infrafabric.io)
|
||||
|
||||
**Quick links:** [Evidence index](https://infrafabric.io/static/hosted/evidence/index.html) • [Model-tier invariants (Dave-proof run)](#model-tier-invariants-dave-proof)
|
||||
|
||||
### Technical Disclosure: AI-Native Implementation
|
||||
This project investigates the Operator-as-Architect paradigm. I do not write manual Python; I utilize LLMs as a kinetic engine to implement my architectural constraints. All code referenced in this dossier was generated by Claude under strict supervision. This application demonstrates that a Security Architect can enforce robust safety standards on a system they did not hand-code—a critical model for Scalable Oversight.
|
||||
|
||||
|
|
@ -79,6 +81,7 @@ Live user diagnostics pages remain OAuth‑protected for privacy. For external r
|
|||
- Evidence index (no auth): https://infrafabric.io/static/hosted/evidence/index.html
|
||||
- Trace protocol spec: https://infrafabric.io/static/hosted/IF_EMOTION_DEBUGGING_TRACE_WHITEPAPER_v3.3_STYLED.md
|
||||
- Verifier (single file): https://infrafabric.io/static/hosted/iftrace.py
|
||||
- Model-tier invariants (Dave-proof run, 15 traces): https://infrafabric.io/static/hosted/EMO_DAVE_PROOF_MODEL_COMPARE_20251222T164352Z.md
|
||||
|
||||
The evidence index links per‑trace pages and the underlying downloadable bundles + SHA256 sidecars.
|
||||
|
||||
|
|
@ -101,7 +104,9 @@ If yes: fund this as a research program in citizen‑auditable operational gover
|
|||
|
||||
If no: reject it, but name the boundary explicitly.
|
||||
|
||||
### Operational finding: Model economics (Dave-proof run)
|
||||
<a id="model-tier-invariants-dave-proof"></a>
|
||||
|
||||
### Operational finding: Stack-enforced invariants across model tiers (Dave-proof run)
|
||||
|
||||
A recurring objection (“Dave”) is that mental‑health‑adjacent UX requires the largest, most expensive models; therefore governance is secondary to model scale. InfraFabric treats that as a falsifiable claim: if the governance stack can enforce output invariants across model tiers, then model choice becomes an economic routing decision (default small, escalate when TRIAGE demands).
|
||||
|
||||
|
|
@ -175,7 +180,7 @@ This dossier documents the **InfraFabric microlab**: a functioning single-shard
|
|||
|
||||
| Core claim | Proof (artifacts) | Limitation (scope / boundary) |
|
||||
|---|---|---|
|
||||
| **A) Traceability is safety.** High‑stakes agents cannot be trusted without a verifiable history of what happened (request → retrieval → decision → output). | **IF.TTT + portable evidence + verifier**<br/>- Evidence index (no auth): [evidence/index.html](https://infrafabric.io/static/hosted/evidence/index.html)<br/>- Paper: [IF.emotion trace protocol (v3.3, styled)](https://infrafabric.io/static/hosted/IF_EMOTION_DEBUGGING_TRACE_WHITEPAPER_v3.3_STYLED.md)<br/>- Verifier: [iftrace.py](https://infrafabric.io/static/hosted/iftrace.py)<br/>- Reference bundle: [emo_trace_payload_016cca78…tar.gz](https://infrafabric.io/static/hosted/emo_trace_payload_016cca78-6f9d-4ffe-aec0-99792d383ca1.tar.gz)<br/>- Latest bundle: [emo_trace_payload_702d4607…tar.gz](https://infrafabric.io/static/hosted/emo_trace_payload_702d4607-4b54-45b1-aecf-b6728d80f124.tar.gz)<br/>- Dave proof (15 traces, 5 prompts × 3 models): [tar.gz](https://infrafabric.io/static/hosted/emo_dave_proof_bundle_20251222T164352Z.tar.gz) • [instructions](https://infrafabric.io/static/hosted/EMO_DAVE_PROOF_MODEL_COMPARE_20251222T164352Z.md) | **Microlab / single shard.** Proven in a single-host environment. Completeness is bounded by explicit witness boundaries; PQ is anchored at registry time (not necessarily on every hot-path artifact). No public append‑only transparency log yet. |
|
||||
| **A) Traceability is safety.** High‑stakes agents cannot be trusted without a verifiable history of what happened (request → retrieval → decision → output). | **IF.TTT + portable evidence + verifier**<br/>- Evidence index (no auth): [evidence/index.html](https://infrafabric.io/static/hosted/evidence/index.html)<br/>- Paper: [IF.emotion trace protocol (v3.3, styled)](https://infrafabric.io/static/hosted/IF_EMOTION_DEBUGGING_TRACE_WHITEPAPER_v3.3_STYLED.md)<br/>- Verifier: [iftrace.py](https://infrafabric.io/static/hosted/iftrace.py)<br/>- Model-tier invariants run (Dave-proof, 15 traces): [tar.gz](https://infrafabric.io/static/hosted/emo_dave_proof_bundle_20251222T164352Z.tar.gz) • [instructions](https://infrafabric.io/static/hosted/EMO_DAVE_PROOF_MODEL_COMPARE_20251222T164352Z.md)<br/>- Reference bundle: [emo_trace_payload_016cca78…tar.gz](https://infrafabric.io/static/hosted/emo_trace_payload_016cca78-6f9d-4ffe-aec0-99792d383ca1.tar.gz)<br/>- Latest bundle: [emo_trace_payload_702d4607…tar.gz](https://infrafabric.io/static/hosted/emo_trace_payload_702d4607-4b54-45b1-aecf-b6728d80f124.tar.gz) | **Microlab / single shard.** Proven in a single-host environment. Completeness is bounded by explicit witness boundaries; PQ is anchored at registry time (not necessarily on every hot-path artifact). No public append‑only transparency log yet. |
|
||||
| **B) Governance requires plurality.** A single model acting as “the judge” is brittle; adversarial viewpoints and escalation are required. | **IF.GOV.TRIAGE → IF.GOV.PANEL**<br/>- TRIAGE: risk preflight sizes panels and escalates<br/>- PANEL: preserves dissent + veto paths<br/>- QUESTIONS: structured inquiry briefs for panels | **Cost / latency tradeoffs.** Multi-seat governance is reserved for higher-stakes decisions; low-stakes paths use smaller panels or fast-track gates. |
|
||||
| **C) Context is the best firewall.** Static filters fail; security must distinguish “reference” vs “leak” and “discussion” vs “exfiltration”. | **IF.SECURITY.CHECK + IF.SECURITY.DETECT**<br/>- CHECK: epistemic coherence checks (detective layer)<br/>- DETECT: secret/relationship screening primitives | **Domain specificity.** Calibrated for concrete security surfaces (secrets/PII/prompt injection); generalizing to broader “harmful intent” is an open research vector. |
|
||||
|
||||
|
|
@ -21960,4 +21965,3 @@ If any item fails, the system fails this appendix.
|
|||
This appendix is intentionally dull. That is the point.
|
||||
|
||||
---
|
||||
|
||||
|
|
|
|||
Loading…
Add table
Reference in a new issue