LHC Collider v2.1
Visual failure → runtime failure → one-byte repair → exact reconstruction and runtime replay.
- ARTIFACT
- ELF64
- BYTES
- 2,048
PARS / VERIFICATION ARCHITECTURE FOR AI REASONING EXPERIMENTAL
Reasoning should survive contact with evidence.
PARS forces difficult AI work through explicit contracts, competing mechanisms, hostile tests, reconstruction, and a final invariant gate—so the claim cannot outrun the evidence.
CANDIDATE STATE
READY FOR ADVERSARIAL RUNRun the complete PARS cycle. A hostile perturbation will break one invariant; Build must stay blocked until Reconstruction and retest repair the candidate.
INV-001 PASS
INV-002 PASS
INV-003 UNTESTED
CLAIM BOUNDARY
PENDING
run the adversarial pass
PARSE→BRANCH→TRANSFORM→PERTURB→TEST→RECONSTRUCT→BUILD
PUBLIC RECORD7historical / development cases
CLASSIFICATION3PASS · FAIL · INCONCLUSIVE
DISCIPLINE↺failures and repairs preserved
PROSPECTIVE CLAIM—benchmark still pending
01 / WHY
Generation is easy.
Knowing when to trust it is not.
PARS treats difficult work as a contract-preservation and evidence-classification problem rather than merely an answer-generation problem.
It keeps competing mechanisms alive, protects a NULL/OTHER possibility, prioritizes discriminating evidence, and preserves failure history instead of smoothing it away.
02 / THE ENGINE
Seven stages. No one-way street.
Changed evidence can send the process backward. A broken representation returns to Transform. A missing mechanism returns to Branch. A changed contract returns to Parse.
Parse the real task.
Separate facts, assumptions, preferences, unknowns, prohibited paths, acceptance criteria, and the narrowest defensible claim.
INPUTambiguous request
OUTPUTfrozen contract + invariants
03 / BREAK IT
The final candidate does not get the benefit of the doubt.
Every acceptance-critical requirement is represented as an invariant. Before Build, PARS replays every active invariant against the exact selected candidate.
If one fails, Build is blocked. If one cannot be verified, PARS must obtain evidence or return an explicitly inconclusive result.
FINAL INVARIANT GATE3 ACTIVE
INV-001Hard requirement preservedPASS
INV-002Evidence chain completePASS
BUILD AUTHORIZEDClick INV-003 to inject a failure.
04 / EVIDENCE > CONFIDENCE
Evidence has an order.
PARS ranks stronger evidence above plausibility. Duplicated evidence counts once. Authority can affect reliability, but does not determine truth.
01Proof / counterexamplestrongest
02Independent measurement
03Controlled experiment
04Held-out prediction
05Reproducible simulation
06Historical replay
07Expert / source comparison
08Visual inspection
09Intuitionweakest
05 / RECONSTRUCTION
New evidence should change the structure, not just the confidence score.
When material evidence changes, PARS removes invalid influence, inserts corrections, rebuilds affected dependencies, reruns downstream checks, and preserves the failure history.
ASSUMPTION A
EVIDENCE B
COLLAPSED
COLLAPSED
CLAIM C
BUILD D
RECONSTRUCT
affected state only
affected state only
06 / FIELD RECORDS
Failures stay in the record.
Historical cases are development evidence, not a prospective proof of PARS superiority. They show the method, boundaries, repairs, and proof burden.
WASM Hyperlattice
Cross-format direct serialization with zero imports, perturbation, regeneration, and exact identity checks.
- ARTIFACT
- WASM
- BYTES
- 9,433
BVIS-001 audit
Technical checks passed while upstream byte-stream provenance remained absent. Strict result: INCONCLUSIVE.
- RESULT
- INCONCLUSIVE
- CLAIM
- BOUNDED
07 / VERIFIED ADAPTATION
Learning claims are classified, not implied.
Prompting, context, memory, recursive search, repeated inference, and controller edits do not magically become model training. PARS separates four adaptation classes.
RL0Current-task adaptationEvidence changes reasoning during this task.
RL1Verified strategy memoryScoped strategy retained with evidence and failures.
RL2Verified controller adaptationChild controller promoted only after separated evaluation.
RL3Parameter learningActual model parameters change through training.
08 / RUN PARS
Install the skill.
Freeze the contract.
Attack the answer.
The public repository packages PARS as the installable Codex skill $apply-pars-deep, together with the operative candidate specification, whitepaper, evidence templates, and execution references.
WINDOWS / POWERSHELL
git clone https://github.com/rookepoole/PARS.git "$env:USERPROFILE\.codex\skills\apply-pars-deep"invoke$apply-pars-deep
!
CLAIM BOUNDARY
Experimental means experimental.
PARS v1.25.0-candidate.6 is experimental and unpromoted. The public whitepaper defines a prospective comparison against simpler prompting, but that comparison has not yet been reported as completed. Historical cases are development evidence unless prospectively regenerated under a frozen protocol.
Read the source authority boundary →CREATED BY
Rooke Poole
Independent AI systems builder and researcher. Creator of PARS and PARS-Agent.