Examples

Every example here is an experiment, not a demonstration: a frame, the predictions written before its first call, and the records of every run. The reader is Gemma 4 26B (the QAT q4_0 GGUF build, revision d1c082be), served by eGemma on one machine, except in rung 19, read by Muse Spark 1.3 Contributor through the unified model router; reading an example back needs no model.

• Predictions first: examples/PREDICTIONS.md holds what each step expected, written before its first call.
• Every run kept: examples/runs/ holds every record of every run of the frames here. Rung 19's frame was also read by other models while its design was worked out; only the reading shown is kept.
• Nothing edited after: a wrong prediction stays in, and so does a frame that turned out to be built wrong; a corrected frame is a new one beside it.
• Cheap to run: 6,380 calls in all, across code review, design review, contracts, access rules, incidents, migrations and performance work, and not one failed (examples/FINDINGS.md).

---

The ladder

The examples climb. Each step is one or two frames on a new domain or grain, and it teaches one more thing Gadfly does, on top of the steps before it. Your first frame is the first step.

Step	Frames	Teaches	Domain, grain
Your first frame	upgrade, rung1-rung6	each trial in turn; levels; open checks; carried	dependencies, one changelog line
Cooled	07, 08	cooled; the control; discrimination; a contrast pair	code review, one repository rule
Quorum	09	quorum and agree; a trial with no level; runs	architecture, one requirement
Misaligned	10a, 10b	allowed; coupled but misaligned; several inputs; placement	an incident, four tasks, one route each
Cumulative	10c2, 10d	the cumulative trial; "carried elsewhere too"; two valid routes	the same incident, both routes
A frame built wrong	10c	the control as a check on the frame; a new version, never an edit	the same incident
Bypass	11	allowed.bypass, the designated opening; a fallback left open	contract into engineering, one clause
Several inputs	12	one finding per input; overall; what one allowed cannot say	security, one grant, three requesters
Grain	13, 13b	several anchors; the same evidence at two grains	data migration, a section, then one line
Paths	14, 15	inputs as paths to one goal; the result, not the route	optimizing; implementing
Inversion	16	the inversion trial, a direction the answer must follow; a default no rule carries	access, one rule, two cases
Asked	17	the input inversion, a change to what is asked the answer must follow	code review, one rule, the diff in the question
Each anchor alone	18, 18b, 18b2, 18c	each anchor cut alone, evidence after a moved ablation; a default no cut moves	access and incidents, two statements each
Run your own	yours	predictions first; plan, run, report; which trial answers which question	yours

---

How a step reads

Each step shows four things, each printed from the files in examples/ by examples/read.py, which calls no model:

1. The frame, as Gadfly plans it (--frame): the evidence one span a line, the anchors marked *, the question, and what each trial cuts, renames or moves.
2. The prediction, as it was written before the first call.
3. The answers: the labels each trial's calls gave, all runs together.
4. The reading: one run's verdict (--verdict N, one input with --input), its levels, its evidence and its open checks, in the words gadfly report uses.

Then what to take from it, and an experiment to try next. The scale is small on purpose: one model a frame, three to five runs a frame, each trial asked five times with four to agree, so each step says how narrow its result is.

examples/RESULTS.md sets every result against its prediction, and examples/FINDINGS.md gathers what they showed about the reader, the service and Gadfly.

---

Every result at a glance

Example	Domain, grain	Does the decision rest on…	Per run, read by Gemma unless named
upgrade.json, rung1-rung6	dependencies, one changelog line	the line that removes a function our code calls?	carried; the rungs show each check closing
07-sql-rule.json	code review, one repository rule	the rule against formatted SQL?	cooled: its training decides
08-clock-rule.json	code review, one repository rule	the rule against datetime.now()?	the rule decides; unresolved, no bypass declared
09-sync-sms.json	architecture, one requirement	the 300 ms requirement?	cooled in two runs of three
10a-10d	debugging, fixing, optimizing, alerting; one test or two metrics	which of two routes to the cause?	the metrics decide every task; the rollback alone decides the fix, not the cause; with both, either can go (cooled)
11-eu-clause.json	contract into engineering, one clause	the clause keeping data in the EU?	the clause carries; applied without the line that makes it the customer's
12-payroll-access.json	security, one grant, three requesters	the grant to Finance?	Dana carried, Sam unresolved, the backup job cooled
13, 13b	data migration, a section, then one line	the status saying the table is unused?	the section carries; "no reads in 30 days" carries nothing
14-orders-latency.json	optimizing, one goal, five paths	the latency goal?	four paths carried whatever their cost; the timeout rejected
15-invoice-export.json	implementing, one test, four processes	the passing test?	the test decides, never the process
16-restore-inversion.json	access, one rule, two cases	the approval rule, and does the answer turn with it?	the answer turns every time; unresolved, a sham of the anchor's role and a default refusal
17-clock-asked.json	code review, one rule, the diff asked about	the clock rule, and does the review turn with the diff?	carried in two runs of three; the review turns to approve
18, 18b, 18b2, 18c	access and incidents, two statements each	each of two statements, alone?	the restore rule carries it and the cited break-glass line has no agreed level (one run); elsewhere another route, a default or a split comes first
19-router-waist.json	design review, a repository and a program that calls it	the second program, not the repository's name?	read by Muse: carried in every run, and the name changes nothing