v1.2.0 — R3.2 Topological Invariance + κC Compression Threshold + Hash Chain

SAS Multimetric Tribunal v1.2.0

Date: 2026-10-28
Commits: b060a0c (κC) · e401095 (R3.2) · aa4b83f (hash chain) · b46c4f5 (README)

New in v1.2.0

κC Compression Threshold (commit b060a0c)

• Replaced binary compression cutoff with a continuous κC curve
• R3.1 abstentions: 58 → 1 (−98%) across N1 (1,800 A→B pairs)
• 57 pairs with CR > 0.5 recovered and correctly evaluated

R3.2 Topological Invariance (commit e401095)

• New 8th tribunal module: detects semantic equivalence under paraphrasing
• Zero external dependencies (no NLTK, no WordNet, no LLMs)
• O(1) inline canonical clusters (~800 entries)
• Deterministic lemmatizer, passive-to-active voice conversion, strict anchoring
• Conditional gate: PARAPHRASE_RESCUE / CONTRADICTION_VETO / INACTIVE
• 7-test suite: ALL PASS
• Internal module hash: 85236b0c851270bf49d4a4114a53bad8835d916df25aa00b32f128b7fafa0adc

Forensic Hash Chain (commit aa4b83f)

• 6 nested SHA-256 levels covering 103 .py files
• Anchored to Bitcoin via OpenTimestamps
• Reproducible script: python hash_chain/build_hash_chain.py

Enhanced Public Demo

• Displays zone (COLLAPSE/RUPTURE/COHERENT) and risk level (CRITICAL/HIGH/LOW)
• Includes R3.2 topological score, lexical overlap, and audit hash
• Requires a minimum of 30 characters, with a clear message

CI/CD Smoke Test

• Runs manually or every 6 hours (designed to avoid race conditions with Render deployments)
• Validates 19 endpoints, including v2/audit, whoami, diff, batch, and interaction stability
• Resilient to rate limiting (429 errors) and Render-specific 502/503 errors

Metrics

Metric	Value
Evaluation modules	8 (7 core + R3.1 + optional R3.2)
F1 (core, 1,800 pairs)	98.99%
F1 (+R3.1 Guard, N1)	0.833 ± 0.0055
FinQA detection	87.0%
R3.1 abstentions	58 → 1 (−98%)
External dependencies	0
Latency	~100 µs/pair
Hash chain	103 files, 6 levels, Bitcoin-anchored

• Paper v9 (Qeios): DOI 10.32388/T420IV
• Zenodo: DOI 10.5281/zenodo.19702379
• TAD Registry: EX-2026-18792778
• Landing Page: https://leesintheblindmonk1999.github.io/sas-landing/
• API Docs: https://sas-api.onrender.com/docs
• PyPI: https://pypi.org/project/sas-client/
• npm: https://www.npmjs.com/package/sas-audit-client

SAS v1.2.0 — CLI Onboarding, SourceTargetGuard, Production Hardening

SAS v1.2.0

Developer onboarding and production hardening release.

Added

• Official sas-client v0.2.0 support from PyPI.
• Terminal onboarding with sas request-key.
• Automatic Free API key delivery by email.
• sas demo-audit, sas whoami, sas plans, sas public-stats, sas public-activity.
• SourceTargetGuard for source-response factual slot mutations.
• manipulation_alert.triggered synchronization with SourceTargetGuard.
• /robots.txt with crawler guidance.
• HEAD / for uptime monitors.
• Safer 422 validation responses with examples.
• New documentation:
  • docs/manifold.md
  • docs/api.md
  • docs/billing.md
  • docs/benchmark.md
  • docs/security.md

Changed

• Legacy public key sas_test_key_2026 is now limited to 5 requests/day.
• Root benchmark claim is presented more conservatively.
• Public metrics and activity remain aggregate/anonymized.

Fixed

• /v1/diff no longer treats same-structure factual mutations as IDENTICAL.
• Eiffel Tower Paris/1889/France → Berlin/1950/Germany now returns:
  • isi = 0.25
  • verdict = MANIFOLD_RUPTURE
  • manipulation_alert.triggered = true
  • sources = ["SourceTargetGuard"]
• Reduced crawler/monitor false errors:
  • /robots.txt no longer 404
  • HEAD / no longer 405
• Detector timeout handling is safer outside the main thread.

Verified

• Hosted API: https://sas-api.onrender.com
• Docs: https://sas-api.onrender.com/docs
• PyPI client: https://pypi.org/project/sas-client/
• DOI: https://doi.org/10.5281/zenodo.19702379

SAS v1.1.0 — Benchmark 98.8% · 0 FP · API Online

SAS v1.1.0 — First Public Release

API: https://sas-api.onrender.com
DOI: https://doi.org/10.5281/zenodo.19689077
OTS Proof: 0713acbbf50e1a0054f545e5eb68078744f9c5a09d4bc370b5224bb81183a6fe

Benchmark (April 29, 2026 — 2,000 pairs)

Metric	Result
Accuracy	98.80%
Precision	100.00%
Recall	97.60%
F1	98.79%
False Positives	0

Confusion matrix: TP=976 · FP=0 · FN=24 · TN=1000

What's included

• Core detection engine: TDA + NIG + κD = 0.56
• Extended modules: E9 (Logical Contradiction) · E10 (Fact Grounding) · E11 (Temporal Inconsistency) · E12 (Topic Shift)
• FastAPI endpoints: /v1/audit · /v1/diff · /v1/chat · /health
• Authentication: API key system with admin generation
• Rate limiting: 50 req/day free tier
• Hosted API on Render (always-on)

Registry

• TAD Argentina: EX-2026-18792778
• Zenodo: 10.5281/zenodo.19689077
• License: GPL-3.0 + Durante Invariance License v1.0
• Author: Gonzalo Emir Durante