The Evidence Hound

With 139 of 140 edits informed by prior reads, this builder practically refuses to touch a file they haven't looked at first.

@NitishKumar-ai

Built With

claude-sonnet-5claude-opus-5antigravity-agent

Real metrics extracted from Claude Code sessions.

Sessions

140

Total Edits

99.3%

Evidence Before Edit

3.7%

Invalid Action Rate

Evidence-Before-Edit:-2.6%

Invalid Action Rate:+0.3%

Live Vibe Replay

claude-opus-5

toolu_0|Executing Bash

1 / 50

Running tool

Bash

Timestamp: 10:39:29 AM

Derived from 8 sessions with real work in them · 3 too thin to count. Every badge names the numbers that earned it.

First Lightmilestone
Captured a first session with real work in it.
215 tool calls across 221 assistant turns · 2026-08-08
Toolsmithmilestone
Reached for 12 different tools. Breadth, not depth — a cheap badge, and labelled as one.
15 distinct tools used · 2026-08-08
Reads Firstratio
Informed 90% of edits with a read, grep, or test — held across 50 edits.
informed 102 of 102 edits (100.0%) · 2026-08-08
Flawless Runratio
Finished a single session of 50+ tool calls without one failing.
58 tool calls, zero failures · 2026-08-08

6 still locked

Polyglotmilestone
Worked with three or more models across at least ten real sessions.
8/10 sessions, with 3 models already used
Clean Handsratio
Kept failed tool calls under 2% across 200 calls.
3.7/2 % of calls failed, needs to come down
Fortnightconsistency
Did real work on 14 separate days.
1/14 distinct days
Long Haulconsistency
Showed up in eight different calendar weeks. The one badge that takes two months.
1/8 distinct weeks
Hands Onconsistency
Stopped and redirected the agent in 60% of 20+ sessions.
8/20 sessions, 3 with a course correction
Getting Bettertrend
Evidence-before-edit is up 10 points or more against your own earlier sessions.
-2.6/10 points of evidence-before-edit improvement

Four scores, each backed by the session data that produced it — click any card to verify the receipts yourself.

3 of 4 scored · 1 locked

○ provisional
Evidence Disciplineratio
10.0/ 10
139/140 edits informed (99.3%)
○ provisional
Prompt Crafttrend
5.7/ 10
steering 6.4% → 4.5%
Execution Hygienetrend
8/10 qualifying sessions
○ provisional
Learning Velocitytrend
Holding steady
flat on Evidence Discipline, Execution Hygiene, Prompt Craft.

View evidence trail

evidence_before_edit: 139/140 edits informed (99.3%)

steering_rate_event: steering 6.4% → 4.5%

invalid_action_rate: locked (8/10 qualifying sessions)

Bands as of Aug 8, 2026