CKC: the `science` profile

The science profile is for empirical discovery: experiments, measurements, datasets, and replication. The trusted base here is softer than in the proof profile, because there is no kernel. So honesty rests on two things instead: provenance (what data, protocol, code, and seed produced the number) and uncertainty (effect size, interval, replication). It also keeps a code correction (repro-fix) separate from a genuine new finding.

Types

Type	Use it when you	Typical impact
`conjecture`	register a hypothesis or prediction, ideally pre-registered	additive
`experiment`	record the execution of a protocol and its raw outcome	additive
`result`	record a measured finding drawn from experiments	additive
`replicate`	report an independent reproduction, whether it succeeds or fails	additive
`null`	report a null or negative result	additive
`data`	add, curate, clean, or version a dataset	additive, see note
`protocol`, `method`	add or change an experimental or analysis protocol	additive, see note
`analysis`	add or revise a statistical or computational analysis	additive, see note
`repro-fix`	fix a bug in analysis or code that changes the numbers	patch, or `!` if conclusions flip
`retract`	withdraw a previously reported finding	breaking, `!`

Note: a data, protocol, or analysis change that alters earlier conclusions is breaking (! plus Invalidates:), the same as a data correction.

Status

Status: values use the sci.* namespace:

`Status:`	Meaning
`sci.hypothesis`	predicted, not yet tested
`sci.piloted`	preliminary or underpowered evidence
`sci.measured`	a result from a single, adequately run study
`sci.supported`	corroborated across analyses or conditions
`sci.replicated`	independently reproduced
`sci.not-replicated`	an independent reproduction failed
`sci.falsified`	contradicted by evidence

Trusted-base footers

ASSUMES: <untested assumption>: a modelling or identification assumption the result rests on.
UNREPLICATED: <scope>: a single, un-replicated run, for example one machine or one cohort.

Protocol fidelity

Provenance says what produced the number; protocol fidelity says the procedure that ran is the procedure that was registered. The two fail independently: an experiment can carry a complete Dataset:, Seed:, and Code: chain and still deviate silently from its Pre-Registration:, and no provenance check can see it, the same way a dropped hypothesis is invisible to the kernel. An undisclosed as-run deviation is this profile's transcription error.

Protocol-Deviations: none | <log>: none asserts that the as-run procedure matches the cited registration or protocol, item by item; otherwise enumerate each deviation with a one-line impact note, for example n=812 not n=1000 (power recomputed: 0.83).

Protocol-Deviations: qualifies a Pre-Registration: or Protocol: footer; without one it has nothing to be compared against and is omitted. A sci.measured or sci.supported commit that cites a Pre-Registration: or Protocol: SHOULD carry it, because those statuses then mean "true as registered", not merely "measured". A deviation that changes the conclusion is not a fidelity note; it is a breaking change (! plus Invalidates:), as elsewhere in this profile.

Provenance and uncertainty footers

Dataset: <id|doi> [sha256:...], Raw-Data: <uri> sha256:...
Protocol: <id>, Method: <one line>, Pre-Registration: <registry url>
Protocol-Deviations: none | <log>: see Protocol fidelity.
Sample-Size: n=..., Replicates: biological=... technical=...
Effect-Size: ..., CI: 95% [..., ...], P-Value: ...
Code: git:<sha>, Environment: <container sha256>, Seed: ..., Hardware: ..., Metric: ...
Ethics-Approval: <id> where it applies.

Identity

Empirical claims do not get an identifier for free, because there is no canonical formal statement. Each referenced claim or hypothesis MUST be a curated slug, for example conjecture:tensor-rank-helps, registered in the claims registry and optionally bound to a DOI or pre-registration once published. See Identifiers.

Examples

conjecture(tensor-rank): tensor rank K>1 beats K=1 on non-additive targets

Claim-ID: conjecture:tensor-rank-helps
Status: sci.hypothesis
Pre-Registration: osf.io/abcd1

experiment(non-additive): K=1 vs K>1 on XOR, radial, and multiplicative targets

Status: sci.measured
Metric: MSE
Sample-Size: n=1000
Seed: 0..4
Hardware: M4 (MPS)
Effect-Size: ΔMSE −0.41 (K=8 vs K=1)
Pre-Registration: osf.io/abcd1
Protocol-Deviations: none
UNREPLICATED: single machine, 5 seeds; needs an independent run
Closes: conjecture:tensor-rank-helps

replicate(non-additive): an independent run fails to reproduce the K>1 advantage

Status: sci.not-replicated
Dataset: D-2026-014 sha256:aa80...
Sample-Size: n=1200
Effect-Size: ΔMSE −0.03
CI: 95% [−0.09, 0.04]
Refutes: conjecture:tensor-rank-helps

repro-fix(scaling): correct an off-by-one in the epoch windowing

The direction is unchanged and the effect size is revised down. This is not a new finding.

Affects: result:igl-linear-scaling
Effect-Size: +6.8pp → +5.1pp
Seed: 0
Impact: patch

CKC: the science profile