THE OLD UNIT
ONE ANSWER
PROMPT→COMPLETION
- No evidence loop
- No observed consequences
- Confidence without accountability
FLUENT
NOT GROUNDED
NOT GROUNDED
AN EXPERIMENT ABOUT THE DIFFERENCE BETWEEN ANSWERS AND WORK
THE MODEL
WAS NOT
THE PRODUCT.
A model can sound right. A harness has to show its work.
Free* Intelligence is a field note and 70-second film. Three fictional prompts, archived on 2 August 2026, show a one-shot model inventing treaties, people, institutions, and quotations. The page argues the failure was architectural, not a matter of price. It also ships askcline, a Bash wrapper that runs tasks through the official Cline CLI. The receipts are historical demonstrations, not a benchmark.
* “Free” described the original experiment. The wrapper is free; model and provider charges may vary.
THE CURRENT UNIT
ONE LOOP
CONTEXT→TOOLS→OBSERVATIONS→CHECKS
- Repository and project rules
- Tool results become evidence
- Human review remains mandatory
AGENT
HARNESS
HARNESS
MODELS PROPOSE◆HARNESSES ACT◆EVIDENCE DECIDES◆HUMANS REMAIN ACCOUNTABLE
THE RECEIPT
FIRST, THE MODEL
INVENTED HISTORY.
We supplied fictional premises. The one-shot system supplied treaties, directors, institutions, quotations, and consequences with newsroom confidence.
Archive note: prompts and outputs were preserved from 2 August 2026. The exact model configuration was not preserved, so these are historical receipts—not a reproducible benchmark.
CAPTURE 01
DIPLOMATIC CHEESE
FICTIONAL PREMISE
$
ARCHIVED MODEL ANSWER
WHAT THE SYSTEM INVENTED
THE DIAGNOSIS
THE FAILURE WAS
ARCHITECTURAL.
The answer was not wrong because it was free. It was ungrounded because the path ended at text.
No retrieval contract
The system was not required to find evidence, resolve identities, or challenge the premise.
No observation loop
Nothing happened after the answer, so there was no real-world feedback to correct it.
No permission boundary
A completion has no explicit policy for what may be read, changed, executed, or blocked.
No proof of work
There was no diff, command output, test result, citation ledger, or accountable reviewer.
PROMPT → ANSWERIS NOT AN AGENT.
THE UPGRADE
WE CHANGED THE UNIT
OF INTELLIGENCE.
Not from a weak model to a stronger model. From a single answer to a governed loop that can inspect, act, observe, retry, and report evidence.
THE OLD PATH
A convincing answer with no accountable loop.
my-repositoryzsh
$ askcline "inspect the repo, implement the issue, run every relevant check, fix failures, and leave it ready to merge"
01 · READ CONTEXT
02 · USE TOOLS
03 · OBSERVE RESULTS
04 · RETRY FAILURES
05 · REPORT EVIDENCE
FIELD NOTE 003 · 70 SECONDS
THE RECEIPT.
THE FAILURE.
THE LOOP.
The rebuilt film tells the complete story: why the one-shot demo failed, what the harness changes, and why no agent system escapes human responsibility.
Download the filmFREE* INTELLIGENCEMODELS PROPOSE · EVIDENCE DECIDES
WHAT THE HARNESS ADDS
MORE CAPABILITY.
MORE RESPONSIBILITY.
A harness can make work observable and governable. It cannot make an autonomous action automatically correct.
Rules before edits
Repository instructions and project rules shape the task before code changes begin.
Actions with consequences
Files, terminal commands, browser operations, and integrations turn text into real work.
Permissions and hooks
Command policies and lifecycle hooks can block, validate, log, or enrich operations.
Checks that can be inspected
Diffs, test output, build results, citations, and traces create an audit trail.
A human still decides
The final diff and reported checks must be reviewed before merge, release, or reliance.
AUTONOMY WARNINGThe default task path uses auto-approval and can edit files or run commands. This wrapper is not a sandbox. Use a clean branch, keep permissions narrow, and review the evidence.
WHAT IS HISTORICAL
THE ONE-SHOT RECEIPTS
- Captured on 2 August 2026
- Prompts and outputs preserved
- Exact model configuration not preserved
- Useful as demonstration, not benchmark
WHAT IS TRUE NOW
ONE CLINE HARNESS PATH
- No direct model HTTP client
- Agent tasks use the official Cline CLI
- Read-only modes use Plan mode
- CI verifies behavior and static claims
INSTALL ASKCLINE
ONE COMMAND.
ONE HARNESS.
NO BYPASS.
The wrapper delegates authentication, providers, sessions, tools, and context management to Cline. It never reads private token files.
- 1Authenticate Cline
cline auth - 2Install the wrapper
Preparing GitHub Pages URL… - 3Add it to PATH
export PATH="$HOME/.local/bin:$PATH" - 4Run from a clean branch
askcline "inspect, implement, validate, retry, and finish"