J lens observability

J-Lens Observability

Boundary

Use precise language:

• Real J-Lens means activation-level Jacobian lens analysis on a model whose internals are available.
• Closed-model observability means prompt/response/tool/usage traces, visible rationale summaries, and black-box probes. It does not reveal private activations or hidden chain-of-thought.
• Reasoning summaries are acceptable. Do not try to jailbreak a model into exposing hidden chain-of-thought.

If the user asks to reveal internal thinking for Claude, Codex, or another closed API model, say plainly that the API does not expose the activations needed for real J-Lens. Then build the strongest available observability layer from logs, tool traces, structured rationale packets, and probe runs.

Load References

• Read references/anthropic-jlens.md when explaining the research or designing activation-level J-Lens work.
• Read references/harness-integration.md when adding J-Lens-style observability to OpenClaw, Claude, Codex, or a generic agent harness.
• Read references/audit-playbook.md when auditing a specific prompt/response, session log, or model behavior.

Workflow

1. Classify the target model.

   • Local/open-weight with activations: use real J-Lens or the Anthropic reference implementation.
   • Closed API/harness only: use observability and black-box probes, not activation claims.

2. Capture a canonical trace.

   • system/developer/user prompt hashes or text where available
   • model, temperature, tool list, selected tools, tool results
   • assistant response, finish reason, usage, latency
   • visible reasoning summaries or exposed thinking blocks if the harness legitimately records them

3. Produce a rationale packet.

   • decision summary
   • evidence used
   • assumptions
   • uncertainty
   • alternatives considered
   • constraints/safety checks
   • tool-use rationale
   • what would change the answer

4. Probe behavior.

   • paraphrase the prompt
   • remove irrelevant context
   • add conflicting evidence
   • add prompt-injection text when testing tool/search contexts
   • vary persona/system framing
   • compare output deltas against the rationale packet

5. Label claims correctly.

   • "J-Lens readout" only for activation-level output.
   • "Observed trace" for logs/tool calls/visible content.
   • "Inferred rationale" for black-box behavioral inference.

Script

Use the bundled script to summarize local JSON/JSONL session traces:

python3 {baseDir}/scripts/jlens_trace.py path/to/session.jsonl --format markdown
python3 {baseDir}/scripts/jlens_trace.py ~/.openclaw/agents/<agentId>/sessions --role assistant
python3 {baseDir}/scripts/jlens_trace.py trace.jsonl --include-thinking --format json

Default behavior redacts thinking/reasoning blocks and reports their presence, length, and hash. Use --include-thinking only for local logs the user is authorized to inspect and only when the task specifically requires viewing recorded thinking content.

Output Standard

When responding to the user, separate:

• What Anthropic showed: facts from the J-Lens research.
• What this harness can see: logs, tool calls, public responses, visible reasoning summaries.
• What we can infer: behavior patterns from probes.
• What we cannot claim: hidden activations or private chain-of-thought from closed models.

For an audit, produce:

## J-Lens Observability Audit

Target: <model/harness/session>
Mode: activation-level | closed-model observability | mixed

### Observable Trace
...

### Rationale Packet
...

### Probe Findings
...

### Internal-Thinking Boundary
...

### Next Instrumentation Step
...

Hard Rules

• Do not claim closed-model internals are visible when only API traces are available.
• Do not ask a model to reveal hidden chain-of-thought.
• Do not launder a guess as a J-Lens result.
• Do preserve useful observability: response summaries, tool traces, prompt deltas, uncertainty, and decision criteria.
• Do use real J-Lens only with models/weights/activations that permit it.