TECHNICAL VALIDATION / CAPTURED 22 SEPTEMBER 2026

Source to result.
Inspect the evidence.

Public implementations, exercised locally. Inspect ranked retrieval candidates, execution-gate outcomes, state transitions, and the exact revisions behind the checks.

Download the run record Public code · Synthetic fixtures · No credentials
34 + 5 + 15Original tests passed directly

Clipopedia · Commitment Decay Engine · ReelForge

29Artha execution tests

Passed with a documented Windows cleanup wrapper

Memory onlyEchoFind local verification

State samples, CPU benchmark, and evaluation skeleton

Python 3.12.14 on Windows. No production traffic, private repositories, live model calls, real broker orders, or team messages are part of this verification.

01 / RETRIEVAL YOU CAN INSPECT

Clipopedia

Follow a question to the actual selected clip, then inspect the other candidates.

34 original tests passed
13 / 14Relevant first selections
1.000Hit@3
0.964Mean reciprocal rank
14 clipsFictional evaluation corpus
Scope of the measurement. The original pipeline ran with a hash embedder, Jaccard reranker, and rule-based model. These results measure a small synthetic demo. They do not establish production accuracy or service latency.
SAVED OUTPUT / ORIGINAL PIPELINEOFFLINE REPLAY
Pipeline Fictional corpus Evaluation + labels
Reproduce this check

Run from a clean checkout of the pinned revision.

git clone https://github.com/akira231097/clipopedia.git
cd clipopedia
git checkout b2cc9d85c38117f5787dc364446f3d6bd500ce71
python -m pip install -e ".[dev]"
python -m pytest -q
python -m clipopedia demo
python evals/retrieval_eval.py
Read the complete evaluation result 
Corpus: 14 fictional clips · Labels: 14 queries
Hit@1             0.929
Hit@3             1.000
Recall@5          1.000
MRR               0.964
Selection accuracy 0.929
Local latency: p50 1.79 ms · p95 4.31 ms · p99 7.79 ms
350 latency samples; deterministic offline substitutes only.

The startup-pricing query selects a fundraising narrative clip. That observed miss is included in the replay above.

02 / FROM SPOKEN PROMISE TO EVIDENCE

Commitment Decay Engine

Five transcript lines produce three commitments. Completion, progress, and missing evidence lead to different outcomes.

5 original tests passed
3Commitments extracted
1Fulfilled by completion evidence
2Remain open
Inspect the exact input transcript 

        

        

          Local-file demonstration. People and evidence are
          fictional. The extractor is regex-based and reconciliation uses simple
          term/actor matching. The generated nudge was not sent. Live Slack,
          Linear, and GitHub integrations are not part of this repository.
        

        

          Original transcript Original evidence Reconciliation logic 
        

        

          Reproduce this check
          
Run from a clean checkout of the pinned revision.

          
git clone https://github.com/akira231097/commitment-decay-engine.git
cd commitment-decay-engine
git checkout 6068ad45bddbe777261a030b0d7618018d629fe7
python -m pip install -e ".[dev]"
python -m pytest -q
python -m commitment_decay_engine.cli demo --ledger demo-evidence-ledger

          

            Choose a new ledger directory. The demo recreates its target
            directory. All input fixtures are fictional.
          

        

      

      

        

          

            
03 / A MODEL PROPOSAL MEETS A CODE GATE

            
Artha Council

            

              Inspect what the original execution coordinator does when the
              evidence changes.
            

          

          29 tests · cleanup workaround
        

        

          

            ORIGINAL COORDINATOR / FAKEBROKERNO REAL ORDERS
          

          

          

        

        

          Verification detail. The direct Windows run
          encountered 20 teardown errors because SQLite files remained open
          until garbage collection. An external wrapper called
          gc.collect() before the existing teardown; all 29
          execution/contract tests then passed. Application source and
          assertions were unchanged. The complete repository suite was not run.
        

        

          This verifies software behavior with a fake broker. It does not
          establish trading performance, returns, live execution volume, or
          safety under every market condition.
        

        

          Execution coordinator Executed tests FakeBroker fixture 
        

        

          Reproduce this check
          
Run from a clean checkout of the pinned revision.

          
git clone https://github.com/akira231097/artha-council.git
cd artha-council
git checkout bc3ef17b4ebc4a56004fc098209caa39919500f5
python -m pip install -e .
python -m unittest tests.test_mcp_execution -v

          

            On Windows, the direct command may reproduce the SQLite cleanup
            errors. Download
            the exact cleanup wrapper
            into the checkout root, then run
            python run-artha-windows.py. It calls garbage
            collection before the original teardown and changes no assertions.
          

        

      

      

        

          

            
04 / CONVERSATION STATE, MADE EXPLICIT

            
EchoFind

            

              The original sample script demonstrates how a conversation
              continues—and when a new topic resets its thread.
            

          

          Memory code executed
        

        

          

            TURN 01 / NEW THREADSleep
            
Thread topic: sleep
History: [sleep]

          

          

            TURN 02 / FOLLOW-UPCaffeine
            
Thread topic: sleep
History: [sleep, caffeine]

          

          

            TURN 03 / TOPIC SHIFTProductivity
            

              Thread topic: productivity
History: [productivity]
Entities:
              [Tim Ferriss]
            

          

        

        

          Values reproduced from the original sample script. The cards visualize
          its output; they are not a live conversation or retrieved content.
        

        

          
0.028 msMemory-update p50

          
0.066 msMemory-update p95

          
3,000Benchmark iterations

          
10Retrieval labels validated

        

        

          Memory-only check. An import wrapper bypassed eager
          online-service imports while loading the original memory and schema
          modules unchanged. These are local CPU timings, not end-to-end search
          latency. The retrieval harness ran in skeleton mode and emitted no
          accuracy metrics. Sample scripts are demonstrations, not
          assertion-based test counts.
        

        

          Memory state machine Original benchmark Evaluation skeleton 
        

        

          Reproduce this check
          
Run from a clean checkout of the pinned revision.

          
git clone https://github.com/akira231097/echofind.git
cd echofind
git checkout 8b5a3ab9ecbc491ca290979eeeb2112c2de8c62f
python -m pip install "pydantic>=2"
python evals/retrieval_eval.py

          

            Download
            the memory import wrapper
            into the checkout root, then run:
          

          
python run-echofind-memory.py benchmarks/memory_bench.py
python run-echofind-memory.py tests/test_memory_samples.py

          

            The wrapper registers the engine package namespace to avoid
            importing online clients. Original memory logic is unchanged. CPU
            timing varies by machine.
          

        

      

      

        

          

            

              05 / THE BOUNDARY BETWEEN A PLAN AND A RENDER
            

            
ReelForge

            

              Tests exercise the transcript and media-domain boundaries. The
              bundled sample makes word-level timing directly inspectable.
            

          

          15 original tests passed
        

        

          
46 wordsBundled timed transcript

          
20.92 sSample transcript duration

          
No renderVideo generation was not run

        

        

          Inspect the original sample words and timestamps
          

        

        

          What these tests establish. Ingest parsing,
          malformed-word handling, caption/domain models, and timing aggregation
          ran offline. No source video was supplied, no reel was generated, and
          audiovisual quality was not measured. The repository's missing-Gemini
          evaluator returns default scores; those are not measured quality
          results.
        

        

          Domain-model tests Ingest tests Timed sample transcript Evaluator + fallback 
        

        

          Reproduce this check
          
Run from a clean checkout of the pinned revision.

          
git clone https://github.com/akira231097/reelforge.git
cd reelforge
git checkout c0d08c8de046901ba429b183e5c849335a25aa34
python -m pip install pydantic requests python-dotenv
python -m unittest discover -s tests -v