WEBVTT

1
00:00:00.000 --> 00:00:04.020
A research project can leave you with code in one place, results
in another,

2
00:00:04.240 --> 00:00:06.760
and an important decision buried in a conversation.

3
00:00:06.960 --> 00:00:11.860
When you return, which result belongs to which experiment, and
what should happen next?

4
00:00:12.280 --> 00:00:17.240
That is the idea behind Dr. Claw: a persistent workspace for
human-directed research.

5
00:00:17.520 --> 00:00:23.720
It wraps existing command-line coding agents, adding
orchestration, state, and audit views around their execution.

6
00:00:23.960 --> 00:00:29.380
The researcher governs goals, constraints, and acceptance; the
system handles the executable workflow.

7
00:00:29.680 --> 00:00:35.540
We will follow an original project illustration, then examine a
recorded recovery and the paper's pilot comparison.

8
00:00:36.247 --> 00:00:43.807
The workspace maintains four state objects: a Task Graph, an
Artifact Store, a Decision Log, and an Execution Trace.

9
00:00:44.747 --> 00:00:48.687
The Task Graph represents what must be done and the dependencies
between steps.

10
00:00:49.467 --> 00:00:53.927
The Artifact Store holds the produced files and outputs,
connecting completed work to the plan.

11
00:00:54.787 --> 00:01:00.347
The Decision Log retains interventions, while the Execution
Trace preserves what actions and observations occurred.

12
00:01:01.207 --> 00:01:05.647
Together, these objects make an evolving project inspectable
beyond its latest message.

13
00:01:06.347 --> 00:01:12.407
The mechanism is continuity of recorded state: later steps build
on work that can be reviewed and recovered.

14
00:01:13.273 --> 00:01:16.053
Suppose you want to compare two ways of preparing a dataset.

15
00:01:16.453 --> 00:01:20.473
You specify the question, allowed operations, and what evidence
will be acceptable.

16
00:01:21.193 --> 00:01:24.633
The workspace converts that brief into a dependency-aware task
plan:

17
00:01:25.013 --> 00:01:28.873
prepare inputs, run experiments, analyze results, and draft the
report.

18
00:01:29.433 --> 00:01:36.873
Its loop is Plan, Execute, Verify, and Write-back. Outputs enter
the artifact store and task statuses update.

19
00:01:37.473 --> 00:01:40.393
At a checkpoint, you can inspect the evidence and revise the
next step

20
00:01:40.733 --> 00:01:43.513
rather than accepting a finished-looking report automatically.

21
00:01:44.193 --> 00:01:48.153
The intervention is retained, so a changed direction remains
part of the process history.

22
00:01:48.713 --> 00:01:53.373
That lets the next executor work from the shared project context
without losing the researcher's decision.

23
00:01:53.893 --> 00:01:58.753
The interface provides linked views for orchestration, skills,
and the task list.

24
00:01:59.413 --> 00:02:05.533
Skills are reusable instructions for stages such as literature
processing, experiments, analysis, and writing.

25
00:02:06.173 --> 00:02:12.253
The task's stage and type can suggest skills; users can filter
the catalogue or invoke a skill manually.

26
00:02:13.073 --> 00:02:17.853
Versioning and schema checks make an activated skill
identifiable for later replay and audit.

27
00:02:18.653 --> 00:02:24.093
The task list exposes progress, objectives, and linked skills,
and provides an entry point for the next action.

28
00:02:24.793 --> 00:02:26.793
A suggested skill is not guaranteed to run.

29
00:02:27.273 --> 00:02:31.273
The paper distinguishes recommending a capability from verifying
its invocation.

30
00:02:32.273 --> 00:02:37.833
The paper demonstrates recovery by inducing a wrong-path failure
within one research mini-project.

31
00:02:38.573 --> 00:02:43.053
The error is captured in the execution trace, while the prior
project state remains available.

32
00:02:43.793 --> 00:02:49.073
The system halts on the error, then recovers in place with all
five pre-existing files retained.

33
00:02:49.793 --> 00:02:52.553
The recorded walkthrough contains twenty-three tool events.

34
00:02:53.253 --> 00:02:56.313
Corrected artifacts are added without wiping the earlier
attempt.

35
00:02:57.113 --> 00:02:59.293
This is a single-condition design demonstration,

36
00:02:59.553 --> 00:03:03.593
not a matched experiment showing faster or better recovery than
another system.

37
00:03:04.093 --> 00:03:10.373
For our illustration, the useful idea is to fix the failed input
step while keeping the completed work and its history.

38
00:03:10.973 --> 00:03:15.433
The automated pilot compares Dr. Claw with the bare command-line
agent it wraps,

39
00:03:15.773 --> 00:03:18.093
using the same GPT-five point four backend.

40
00:03:18.793 --> 00:03:22.253
Three medical-research tasks receive the same open-ended
instruction,

41
00:03:22.593 --> 00:03:24.893
without a list of all expected research components.

42
00:03:25.533 --> 00:03:30.893
Completeness is the fraction of twenty-one research-practice
elements present in the produced files.

43
00:03:31.533 --> 00:03:37.353
The checks cover items such as reproducibility, multiple models,
calibration, subgroup analysis,

44
00:03:37.753 --> 00:03:39.293
and real literature citations.

45
00:03:39.993 --> 00:03:45.193
This is a coverage measure. It does not establish that every
analysis or scientific claim is correct.

46
00:03:46.153 --> 00:03:51.713
The wrapper changes as one bundle: task graph, state, skills,
and workflow instructions.

47
00:03:52.113 --> 00:03:54.253
The comparison does not isolate one component.

48
00:03:54.787 --> 00:03:58.167
Dr. Claw's pooled completeness is zero point nine five two,

49
00:03:58.586 --> 00:04:01.487
compared with zero point eight seven three for the bare agent.

50
00:04:01.987 --> 00:04:05.287
It exceeds the bare agent on two tasks and ties on the third.

51
00:04:06.067 --> 00:04:11.567
The observed difference comes from research-practice coverage,
including subgroup analysis and real citations,

52
00:04:11.967 --> 00:04:13.507
rather than stronger modeling checks.

53
00:04:14.227 --> 00:04:19.727
There is only one run per task. The reported bootstrap interval
for the pooled difference includes zero.

54
00:04:20.467 --> 00:04:23.887
The pilot therefore gives directional evidence about this
orchestration bundle,

55
00:04:24.327 --> 00:04:26.826
without a statistically confirmed superiority claim.

56
00:04:27.367 --> 00:04:30.127
Persisted audit objects are architectural affordances,

57
00:04:30.567 --> 00:04:34.367
not an independent performance score that the bare agent was
designed to satisfy.

58
00:04:35.047 --> 00:04:40.987
Dr. Claw's contribution is a workspace that links human
decisions with existing agent execution

59
00:04:41.007 --> 00:04:42.547
through persistent research state.

60
00:04:43.187 --> 00:04:49.247
The original project illustration shows how a plan, its
artifacts, and later interventions can stay connected.

61
00:04:49.847 --> 00:04:53.367
The recorded recovery shows prior files surviving a failed step,

62
00:04:53.747 --> 00:04:56.907
while the pilot separately measures expected-component coverage.

63
00:04:57.687 --> 00:05:02.847
These are different pieces of evidence: recoverability, process
visibility, and output coverage.

64
00:05:03.647 --> 00:05:08.707
The researcher's acceptance decision still matters, because a
traceable process needs its conclusions checked.

65
00:05:09.507 --> 00:05:16.187
The useful unit is the continuing project: what was attempted,
what changed, and what evidence supports the next step.