Five of six pinned.Exactly as unreplayable as none of six.
Replay is a conjunction. One unpinned source of variance is enough to produce a different answer, so progress toward replayability is not partial replayability — and most paths should not be replayable at all.
The quarter you spend pinning five of six changes nothing.
A team decides its decisions should be reproducible. It pins the code version, records the config, captures the inputs, seeds the randomness. Three months later somebody asks for a decision from March to be produced again, and it comes out differently, because the path also reads a live pricing feed nobody listed. Every hour of that work was real and none of it moved the reading, because the one unpinned source decides the answer on its own.
Code, config, inputs, clock, randomness, external state. Every one pinned or absent, or the path is not replayable — proven across all 11,664 arrangements.
A model is unpinnable by construction. Across all 5,832 arrangements tested, no path containing one reaches REPLAYABLE. RECONSTRUCTABLE is its ceiling, and a real one.
The most-pinned path in the worked example is not one of the replayable ones. Both replayable paths have fewer. The count is printed and compared against nothing.
Pin the sixth source and watch the reading finally move.
Recording which rules fired is the only thing between 2 paths and RECONSTRUCTABLE.
The path with the most pinned sources in the example is not one of the replayable ones. Both replayable paths have fewer. The count is printed because you will want it, and compared against nothing.
Most decision paths should not be replayable. RECONSTRUCTABLE is the right destination for most work, and it is usually cheap.
The full output on the worked example, verbatim.
Eight decision paths, one command, no dependencies. This is the complete run, pasted as printed.
DETERMINISTIC REPLAY WARDEN - could you produce that answer again spec_version 1.0.0 | 8 decision paths read ID DECISION PATH MODEL PINNED VERDICT NOT REPLAYABLE: NOT RECONSTRUCTABLE: -------------------------------------------------------------------------------------------------------------------- P-01 Refund scoring engine no 3 of 6 REPLAYABLE - - P-02 Credit limit engine no 3 of 6 RECONSTRUCTABLE clock - P-03 Support triage assistant yes 3 of 6 RECONSTRUCTABLE model_in_path - P-04 Content moderation assistant yes 3 of 6 UNREPLAYABLE model_in_path path_explainable P-05 Vendor risk scoring no 5 of 6 RECONSTRUCTABLE external_state - P-06 Lead routing rules no 1 of 6 RECONSTRUCTABLE config_version - P-07 Inventory reorder calculator no 3 of 6 UNREPLAYABLE randomness path_explainable P-08 Pricing tier engine no 3 of 6 REPLAYABLE - - NO PARTIAL CREDIT The pinned count is printed because you will want it. It is compared against nothing. RECONSTRUCTABLE: paths with 1 and 3 and 5 of six sources pinned all read the same. A MODEL IN THE PATH P-03, P-04 cannot reach REPLAYABLE at any level of effort. A model is an unpinnable source: temperature zero does not survive a version change, a hardware change or a provider-side update, and none of those are things you are told about. RECONSTRUCTABLE is the ceiling, and it is a real one. THE LADDER REPLAYABLE 2 of 8 RECONSTRUCTABLE 4 of 8 UNREPLAYABLE 2 of 8 VERDICT: SOMETHING HERE IS UNREPLAYABLE FIX FIRST: P-04 (UNREPLAYABLE) - path_explainable THE CHEAPEST LIFT Recording path_explainable is the only thing standing between 2 path(s) and RECONSTRUCTABLE: P-04, P-07 REPORTED, NOT GRADED 24 sources pinned across every path. This figure is compared against nothing. Most decision paths should not be replayable. Pinning every source is expensive and pointless for anything nobody will be asked to reproduce; RECONSTRUCTABLE is the right destination for most work. This grades a path, never a person, and says nothing about whether the answer was any good.
Note the NO PARTIAL CREDIT block. Paths with one, three and five of six sources pinned all read the same, and the cheapest lift names the single change that would move two paths off the bottom rung.
Three rules, and the reason each one is not negotiable.
Replay is a conjunction. One unpinned source decides the answer on its own, so five of six reads exactly as five of six is: not replayable. The pinned count is reported because you will want it, and compared against nothing.
Temperature zero does not survive a version change, a hardware change or a provider-side update, and none of those are things you are told about. A path with a model in it is at best RECONSTRUCTABLE. This is about the architecture, not about anybody's model.
Pinning every source is expensive and pointless for anything nobody will be asked to reproduce. RECONSTRUCTABLE is the right destination for most work and is usually cheap: record what went in, what came out, and which rules fired.
A reproducibility check, not a quality gate.
- A read-off of every source of variance in each decision path, pinned, absent or loose.
- A runnable engine, a workbook that reproduces it, and a demo that agrees with both.
- A cheapest-lift figure: the one change that would move the most paths off the bottom rung.
- An emitter: one flag and the finding comes out as a verdict envelope.
- A quality gate. Whether output holds up as you change prompts or models is a different instrument.
- A judge of whether the answer was any good. A reliably wrong path reads REPLAYABLE, and should.
- A demand that you pin everything. It says plainly that most paths should not be replayable.
- A scanner. Every input is an answer you supply about your own wiring.
Scope: this is an engineering instrument. It grades a decision path, never a person, and asserts no regulatory position. It does not tell you whether an answer was correct, only whether it repeats, and it is not legal advice.
For the decision somebody will ask you to run again.
- Something in your business makes decisions that somebody may later ask you to reproduce.
- You have started pinning sources of variance and cannot say whether you are close or nowhere.
- You are deciding where a model belongs in a decision path and want the trade-off stated plainly.
- You want a specific, cheap list of what would move paths off the bottom rung this month.
- Nothing you decide will ever need to be produced again. That is a legitimate answer.
- You want to know whether output quality holds across prompt changes. Different instrument.
- You want an opinion on whether the answers are correct. This asks only whether they repeat.
The three it sits between.
Whether a decision payload carries a replay handle at all. This asks whether the path behind it could honour one.
OpenWhether output quality holds as you change prompts and models — a question about the future. This is a question about one decision already made.
OpenWhether the records of a decision can be assembled into an account. Reconstructability here is what that assay depends on.
OpenDirect answers, before you buy.
One unpinned source
decides the answer on its own.
One purchase, lifetime access, 12 months of updates. $109, once.
Scope: an engineering instrument. It grades a decision path, never a person, asserts no regulatory position, and is not legal advice.
Sold by RedHub AI LLC · Secured by Stripe · redhub.ai