hosted/IF_EMOTION_DEBUGGING_TRACE_WHITEPAPER_v3.0_STYLED.md

239 lines
8.4 KiB
Markdown
Raw Permalink Blame History

This file contains ambiguous Unicode characters

This file contains Unicode characters that might be confused with other characters. If you think that this is intentional, you can safely ignore this warning. Use the Escape button to reveal them.

# IF.EMOTION TRACE PROTOCOL v3.0: AUDITABLE DEBUGGING (WITHOUT WISHFUL THINKING)
*Alternate title: Debugging Emotion — The Immutable Flight Recorder*
**Subject:** End-to-End Traceability, Completeness Witnessing, and PQ-Anchored Evidence Binding
**Protocol:** IF.TTT (Traceable, Transparent, Trustworthy)
**Version:** 3.0 (Narrative Annex + Edge Commitments)
**Date:** 2025-12-21
**Status:** AUDIT REQUIRED
**Citation:** `if://whitepaper/emotion/trace-protocol/v3.0`
---
> Logging is cheap. Evidence is expensive. If your system cant be verified by someone who doesnt trust you, it doesnt have evidence.
## Why This Matters (Decision Context)
AI incidents are not “debugging problems.” They are dispute problems.
If you cannot prove what the system received, what it did, what it returned, and what evidence it claims to have used, you dont have an incident process. You have a debate club with screenshots.
*The goal is not truth. The goal is provenance: what was recorded, when, by which code, under which keys, and whether something was omitted within an explicit boundary.*
---
## Layered Evidence Stack (Where Guarantees Begin)
```mermaid
flowchart TB
U[User Browser] -->|HTTPS| E[Edge: TLS + Routing]
E -->|/oauth2/*| O[oauth2-proxy]
E -->|/api/*| N[nginx]
N --> B[Backend Witness Boundary]
B --> R[Retrieval / RAG]
B --> P[Prompt Build]
B --> M[Model Inference]
B --> X[Postprocess: citations + trace footer]
N --> EDGELOG["Edge attempt log<br/>(JSONL)"]
B --> T1["REQ_SEEN ledger<br/>(hourly JSONL)"]
B --> T2["Trace events<br/>(hash chain JSONL)"]
B --> T3["Signed summary<br/>(output hash + head attestation)"]
T1 --> H["Signed Merkle head<br/>(per hour)"]
T2 --> S["Trace head<br/>(event_hash)"]
H --> BUNDLE["Evidence bundle<br/>(tar.gz + manifest)"]
S --> BUNDLE
T3 --> BUNDLE
EDGELOG --> BUNDLE
BUNDLE --> REG["IF.TTT registry<br/>(PQ-hybrid anchor)"]
BUNDLE --> MIRROR["Public static mirror<br/>/static/hosted"]
```
**Interpretation:** integrity begins at the backend witness boundary. Completeness is only meaningful at and after that boundary until edge witnessing is fully cryptographic.
---
## What This Protocol Guarantees (And What It Refuses To Claim)
This system does not try to “prove the model is true.” That is not a meaningful claim for probabilistic generation.
This system proves something narrower and more valuable:
1) what the system received (as a commitment),
2) what the system did (trace event chain),
3) what the system returned (output hash),
4) what evidence it claims to have used (retrieval IDs + citation handles),
5) that the resulting artifacts are tamper-evident and portable for external review.
If a claim cannot be bound to an artifact, it does not exist.
---
## Evidence Inventory (What Exists, How You Verify It)
| Artifact | What it proves | Verification path |
|---|---|---|
| Edge attempt log | request attempts above the backend boundary | inspect JSONL + correlate request id + trace id + body commitment headers |
| REQ_SEEN ledger | witnessed attempts within backend boundary | Merkle inclusion proof against signed head |
| REQ_SEEN signed head | integrity of a witnessed batch | verify signature + Merkle root |
| trace event chain | ordered sequence of system actions | recompute chain (`prev_hash``event_hash`) |
| signed summary | binding between output + trace head + policy decisions | verify signature + output hash |
| evidence bundle | portable capsule of the above | verify SHA256 + run verifier tool |
| IF.TTT record | anchored registration (PQ-hybrid) | verify registry chain record + signature |
---
## The Inversion (Where Teams Usually Get This Wrong)
The problem isnt hallucination.
The problem is un-auditable systems pretending to be auditable.
Most teams treat tracing as “extra logs.” In probabilistic systems, tracing is the product: it is the only thing that turns a dispute into a verification.
---
## The Human Constraint (Why This Doesnt Ship)
**The resistance:** “This is too much work; well add observability later.”
**The psychology:** Shipping dopamine beats audit discipline, and nobody gets promoted for preventing future disputes.
**The unstick:** Make “no trace, no trust” a deployment gate. If it cant be verified by an adversarial reviewer, it doesnt ship.
---
## Core Protocol (Technical Specification)
### Trace ID contract
Every request to `/api/chat/completions` receives a Trace ID. This includes denials.
- Client request header (optional): `X-IF-Client-Trace: <uuid>` (if present, backend uses it as Trace ID)
- Response header: `X-IF-Emotion-Trace: <uuid>`
- Response header: `X-IF-Emotion-Trace-Sig: <sig>` (app-level integrity)
- User output: final line `Trace: <uuid>`
### Body commitment (tamper-evident “what I sent”)
The client may send:
- `X-IF-Body-Sha256: sha256=<hex>` (or `<hex>`)
The backend computes `sha256(request_body_bytes)` and records:
- `request_body_sha256` (computed),
- `claimed_body_sha256` (from header),
- `body_commit_ok` (match flag).
This provides a commitment without storing full request bodies in witness layers.
### Edge attempt logging (interim completeness above backend)
Because auth/rate limits can deny requests before the backend can write REQ_SEEN, nginx writes an edge attempt log:
- File: `/var/log/nginx/edge_req_seen.jsonl`
- Captures: `req_id`, `status`, `client_trace`, `body_sha256` (header), timing, and (if available) `auth_email`.
This is a pragmatic bridge. It is not yet a cryptographic edge ledger.
### Backend completeness (REQ_SEEN) and boundary
REQ_SEEN records every request attempt that reaches the backend witness boundary as a privacy-preserving commitment:
- Ledger: `/opt/if-emotion/data/req_seen/<YYYYMMDDTHH>.jsonl`
- Signed head: `/opt/if-emotion/data/req_seen/heads/<YYYYMMDDTHH>.json`
REQ_SEEN completeness is only valid for requests that reach the backend process. Requests blocked before the backend are out of scope until a cryptographic witness is moved to the edge.
### Privacy for witnessed prompts (avoid rainbow tables)
REQ_SEEN stores a keyed commitment by default:
- `user_text_hash_alg: hmac-sha256`
- `user_text_hash: hmac_sha256(secret, user_text_bytes)`
- `user_text_sha256_prefix: <short prefix>` (debug-only correlation)
Key source:
- `IF_REQ_SEEN_HMAC_KEY` if configured, otherwise falls back to `IF_TTT_SIGNING_SECRET`.
### Trace events and signed summary
- Trace events: `/opt/if-emotion/data/trace_events.jsonl` (hash chain)
- Signed summary: `/opt/if-emotion/data/ttt_signed_log.jsonl`
- Optional payload capture (OAuth-gated): `/opt/if-emotion/data/trace_payloads/<trace_id>.json`
Signed completion records include `code_identity` (code SHA + config hash) so “which code ran?” is auditable even without remote attestation.
### Post-quantum boundary
Evidence bundles are PQ-hybrid signed when registered into IF.TTT:
- `pq_status: hybrid-fips204`
- `pq_algo: ML-DSA-87`
Hot-path signatures (Ed25519) are not post-quantum. The correct claim is “PQ-anchored at registry time,” not “PQ everywhere.”
---
## Verifier Tooling
Primary verifier:
- `https://infrafabric.io/static/hosted/iftrace.py`
Runbook:
- `https://infrafabric.io/static/hosted/VERIFY_EMO_TRACE_016cca78-6f9d-4ffe-aec0-99792d383ca1.md`
---
## Reference Proof Run (Transport + Integrity)
Bundle:
- `https://infrafabric.io/static/hosted/emo_trace_payload_016cca78-6f9d-4ffe-aec0-99792d383ca1.tar.gz`
Directory (discovery):
- `https://infrafabric.io/static/hosted/`
SHA256:
- `7101ff9c38fc759a66157f6a6ab9c0936af547d0ec77a51b5d05db07069966c8`
One-line verification:
- `curl -fsSL https://infrafabric.io/static/hosted/emo_trace_payload_016cca78-6f9d-4ffe-aec0-99792d383ca1.tar.gz | sha256sum`
Expected output:
- `7101ff9c38fc759a66157f6a6ab9c0936af547d0ec77a51b5d05db07069966c8 -`
IF.TTT citation handle (PQ-hybrid signed):
- `if://citation/c24fe95e-226c-4efc-ba22-5ddcc37ff7d2/v1`
---
## Narrative Annex (Human-Readable Index)
This paper deliberately separates:
- Evidence: JSONL + signed summary inside the bundle.
- Narrative: deterministic projection that points back to evidence by `event_hash`.
Annex (public):
- `https://infrafabric.io/static/hosted/IF_EMOTION_TRACE_REFERENCE_016cca78_IF_STORY.md`
The narrative is useful during review, but it is not the source of truth.
---
> The models answer is ephemeral. The trace is the product. If you cant prove what happened, you are not running an AI system — you are running a narrative.