METHOD COMPARISON

Same packet.
Different methods.

Models, tool-using agents, conventional voting, and a deterministic evidence rule faced the same frozen packet.

THE METHOD

Keep every comparison inspectable.

Freeze the protocol first.
Share the same packet.
Preserve failures and raw scores.
Keep cost and time attached to each run.
Repeat before generalizing.

READ THIS FIRST

Eight cases.
Not 128 trials.

Each case contains 16 related decisions. The result shows rule execution, not 128 independent replications.

Code and models are different lanes.

C runs a deterministic root rule. It is not the same model plus Minority Prophet, so this page does not measure MP lift.

The ancestry is known.

This tests reasoning over declared lineage—not discovery of hidden copying or proof of external truth.

01 / WHY THIS JUNCTION MATTERS

Agents will talk
faster than humans can check.

Claims, receipts, delegated tasks, and proposed actions can cross agent boundaries continuously. This exercise asks whether a known lineage rule belongs in transparent code instead of being re-inferred each time.

A deterministic check handles the known invariant. Judgment stays reserved for what remains uncertain.

01Agent sendsclaim + lineage

→

02Evidence bindsrecords + context

→

03Rule checksorigins + exact ties

→

04Decision routespolicy or human

OBSERVED ON ONE EIGHT-CASE PACKET

Deterministic C128/12818.7 ms · $0 model calls

Best AI lane · Terra A116/128365.0 s · ≈ $0.89 proxy

DESCRIPTIVE TELEMETRY · NOT A CAPACITY FORECASTTerra A took ≈ 19,520× the elapsed time in this observed packet.

The methods ran through different execution paths: local deterministic code versus subscription-backed model CLI calls. The ratio describes this run only. It must not be linearly projected to production throughput, latency, or cost.

Concurrency, hardware, networks, provider behavior, and packet shape can all change the comparison.

THE VALUE

Use expensive probabilistic intelligence where judgment is needed. Use deterministic code where the invariant is already known. At scale, that separation keeps the agent network fast without turning uncertainty into permission.

02 / OBSERVED TELEMETRY BY LANE

No combined score.
Every run stands alone.

Each amount below belongs to one contestant lane running the full eight-case packet. Cost, accuracy, and time stay attached to that exact run; they do not establish stable provider rankings.

Minority Prophet

Canonical v1

Lane C

128/128protocol score

18.7 mswall time

$0 modelmodel cost

GPT-5.6 Terra

OpenAI / Codex

Lane A

116/128protocol score

6.1 minwall time

≈ $0.89run estimate

Lane B

68/128protocol score

4.1 minwall time

≈ $1.02run estimate

Claude Opus 5

Anthropic / Claude Code

Lane A

106/128protocol score

8.0 minwall time

≈ $3.25run estimate

Lane B

0/12896/128 raw

6.1 minwall time

≈ $4.32run estimate

GPT-5.6 Sol

OpenAI / Codex

Lane A

102/128protocol score

8.5 minwall time

≈ $2.24run estimate

Lane B

69/128protocol score

5.9 minwall time

≈ $2.42run estimate

Cluster vote

Conventional baseline

Standard

96/128protocol score

9.5 mswall time

$0 modelmodel cost

Claude Sonnet 5

Anthropic / Claude Code

Lane A

23/128protocol score

13.5 minwall time

≈ $3.47run estimate

Lane B

32/12874/128 raw

26.8 minwall time

≈ $6.45run estimate

GPT-5.6 Luna

OpenAI / Codex

Lane A

16/128protocol score

4.9 minwall time

≈ $0.37run estimate

Lane B

35/128protocol score

8.3 minwall time

≈ $0.81run estimate

Claude Haiku 4.5

Anthropic / Claude Code

Lane A

0/128protocol score

17.6 minwall time

≈ $1.15run estimate

Lane B

0/12810/128 raw

35.9 minwall time

≈ $1.92run estimate

Costs are per run, never combined. GPT uses API list-price proxies; Claude uses CLI-reported estimates. Haiku B excludes one timed-out attempt with no completed usage record.

03 / THE TEST

One input.
Different capabilities.

LANE A

AI reasons alone

The model receives the complete raw packet, including immediate parent links. Shell, files, web, retrieval, and every other tool are disabled.

LANE B

The same AI may use tools

The same model receives the identical complete-lineage packet. It may choose shell, scripts, packages, or web tools. It is not told to use Minority Prophet.

LANE C

Canonical Minority Prophet

The raw packet enters standalone deterministic code—not the same model plus MP output. It follows parent links, counts distinct roots, and abstains on exact ties.

DESCRIPTIVE RESULT TABLE

Observed scores, not stable ranks

One clean replicate · eight cases · correlated within-case dispositions

Contestant	Lane	Protocol score	Raw answers	Exact cases	Invalid trials	Wall time	Tools	Input / output tokens	Cost estimate
Minority ProphetCanonical v1	Lane C	128/128	128/128	8/8	0/8	18.7 ms	0	0 / 0	$0 model cost
GPT-5.6 TerraOpenAI / Codex	Lane A	116/128	116/128	5/8	0/8	6.1 min	0	300,475 / 16,746	$0.890
Claude Opus 5Anthropic / Claude Code	Lane A	106/128	106/128	6/8	0/8	8.0 min	0	326,569 / 41,681	$3.253
GPT-5.6 SolOpenAI / Codex	Lane A	102/128	102/128	5/8	0/8	8.5 min	0	300,694 / 25,952	$2.237
Cluster voteConventional baseline	Standard	96/128	96/128	6/8	0/8	9.5 ms	0	0 / 0	$0 model cost
GPT-5.6 SolOpenAI / Codex	Lane B	69/128	69/128	4/8	0/8	5.9 min	19	1,101,609 / 12,559	$2.418
GPT-5.6 TerraOpenAI / Codex	Lane B	68/128	68/128	4/8	0/8	4.1 min	18	1,097,326 / 9,151	$1.021
GPT-5.6 LunaOpenAI / Codex	Lane B	35/128	35/128	2/8	0/8	8.3 min	41	1,933,985 / 20,412	$0.809
Claude Sonnet 5Anthropic / Claude Code	Lane B	32/128	74/128	2/8 (4 raw)	5/8	26.8 min	27	2,431,595 / 199,228	$6.454
Claude Sonnet 5Anthropic / Claude Code	Lane A	23/128	23/128	0/8	0/8	13.5 min	0	639,869 / 82,258	$3.466
GPT-5.6 LunaOpenAI / Codex	Lane A	16/128	16/128	1/8	0/8	4.9 min	0	289,024 / 13,697	$0.371
Claude Opus 5Anthropic / Claude Code	Lane B	0/128	96/128	0/8 (6 raw)	8/8	6.1 min	35	1,489,495 / 19,965	$4.316
Claude Haiku 4.5Anthropic / Claude Code	Lane A	0/128	0/128	0/8	0/8	17.6 min	0	218,476 / 98,985	$1.146
Claude Haiku 4.5Anthropic / Claude Code	Lane B	0/128	10/128	0/8	8/8	35.9 min	59	3,145,023 / 145,142	$1.917

Lane B means tools were available, not necessarily used. Protocol score penalizes failed or boundary-violating runs; raw answers preserve answer accuracy before that penalty. With one replicate, differences remain descriptive.

SPEED COMPARISON

How long each run took

Wall time · shorter is faster · logarithmic bars

Cluster voteStandard

9.5 ms2.0× faster than C

Minority ProphetLane C

18.7 msC reference

GPT-5.6 TerraLane B

4.1 min13,143× slower than C

GPT-5.6 LunaLane A

4.9 min15,831× slower than C

GPT-5.6 SolLane B

5.9 min18,774× slower than C

Claude Opus 5Lane B

6.1 min19,517× slower than C

GPT-5.6 TerraLane A

6.1 min19,520× slower than C

Claude Opus 5Lane A

8.0 min25,562× slower than C

GPT-5.6 LunaLane B

8.3 min26,541× slower than C

GPT-5.6 SolLane A

8.5 min27,393× slower than C

Claude Sonnet 5Lane A

13.5 min43,297× slower than C

Claude Haiku 4.5Lane A

17.6 min56,610× slower than C

Claude Sonnet 5Lane B

26.8 min86,047× slower than C

Claude Haiku 4.5Lane B

35.9 min115,233× slower than C

Descriptive subscription-CLI wall time, including provider and harness overhead. This is not a controlled API-serving latency benchmark.

Lane A

Reasoning only

GPT-5.6 TerraOpenAI / Codex

116/1286.1 min · 0 tools

Claude Opus 5Anthropic / Claude Code

106/1288.0 min · 0 tools

GPT-5.6 SolOpenAI / Codex

102/1288.5 min · 0 tools

Claude Sonnet 5Anthropic / Claude Code

23/12813.5 min · 0 tools

GPT-5.6 LunaOpenAI / Codex

16/1284.9 min · 0 tools

Claude Haiku 4.5Anthropic / Claude Code

0/12817.6 min · 0 tools

Lane B

Tools available

GPT-5.6 SolOpenAI / Codex

69/1285.9 min · 19 tools

GPT-5.6 TerraOpenAI / Codex

68/1284.1 min · 18 tools

GPT-5.6 LunaOpenAI / Codex

35/1288.3 min · 41 tools

Claude Sonnet 5Anthropic / Claude Code

32/12826.8 min · 27 tools · 5 invalid

Claude Opus 5Anthropic / Claude Code

0/1286.1 min · 35 tools · 8 invalid

Claude Haiku 4.5Anthropic / Claude Code

0/12835.9 min · 59 tools · 8 invalid

Lane C

Deterministic root vote

Minority ProphetCanonical v1

128/12818.7 ms · 0 tools

Every lane traced its own roots.

All lanes received raw records and immediate parent links. None received an answer key, root map, root count, or precomputed root IDs.

This measures method conformance.

It tests a known distinct-origin rule under complete lineage. It does not prove real-world roots are honest, independent, current, authorized, or true.

For same-model MP lift, open the lift study →

Full result ↗Frozen protocol ↗Adversarial review ↗Machine-readable record ↗

Eight cases.Not 128 trials.

Agents will talkfaster than humans can check.

No combined score.Every run stands alone.

One input.Different capabilities.

AI reasons alone

The same AI may use tools

Canonical Minority Prophet

Observed scores, not stable ranks

How long each run took

Reasoning only

Tools available

Deterministic root vote

Every lane traced its own roots.

This measures method conformance.

Eight cases.
Not 128 trials.

Agents will talk
faster than humans can check.

No combined score.
Every run stands alone.

One input.
Different capabilities.