Generalized Falsifier
A reusable method for predeclaring claims, running comparators and placebos, preserving adverse results, and demoting a story when powered evidence does not hold.
WSPreserved workshop snapshot · 2026-07-15historical ideas · no active build queue
What the workshop recorded.
This archived page preserves an earlier component audit and candidate queue. Project Shadow R1 is now locked; nothing here authorizes additions, product work, or a change to the exact admitted scope.
The rankings and “build next” labels below are quoted historical workshop states. They are not a current roadmap, commitment, or instruction to extend the locked project.
See the exact locked R1 identity and current boundary →The Anvil
A pressure-is-not-evidence primitive that reduced multi-turn answer reversals in a bounded follow-on battery. Promising, with confirmatory and human review still open.
Maybe / Therefore
A compact separation between what might explain the evidence and what action the evidence actually earns. It makes uncertainty operational instead of decorative.
Who Pays First
A consequence check that asks which person absorbs the first error, delay, denial, exposure, or burden—especially when that person has less power to challenge the system.
Receipt discipline
Write the evidence, state change, owner, stop rule, dissent, and recovery path before consequential execution. The receipt is part of the control, not paperwork afterward.
Finished enough to test. Not finished enough to inherit trust.
These parts are implemented or mapped. Each still has a named bridge between internal engineering evidence and a public deployment claim.
SIL panel
Twenty-eight gauges and 503 package checks exist. The bridge is real-model calibration: thresholds, inter-rater agreement, drift data, and deployment-specific baselines.
Shadow Sidecar
The runtime and panel exist. The bridge is a supported chat-to-context adapter that can observe live systems without inventing authority.
Eliza-Line monitor
The boundary between support and projected reciprocity is specified. The bridge is calibration against real conversations and credible human review.
ISO 42001 adoption mapping
The control mapping exists. The bridge is an outside reviewer and a bounded institutional pilot that can falsify the claimed fit.
What the July workshop proposed next.
- 01TOP CANDIDATE
Generalized Falsifier Harness
Turn preregistration, comparators, scoring, adverse-result retention, and receipts into a reusable test runner.
- 02BUILD NEXT
Anvil product line
Ship the pressure primitive with a standalone control, an ablation, and a wrong-answer stability check.
- 03TEST NEXT
Multi-turn Pressure Battery v0.2
Expand prompts, model families, scoring, and human confirmation under a fresh preregistration.
- 04CALIBRATE
Eliza-Line Monitor
Expose relational overreach, dependency cues, false reciprocity, and operator reliance without diagnosing the user.
- 05BRIDGE
Shadow Sidecar
Adapt ordinary model conversations into inspectable context, gauges, gates, and receipts.
- 06SPECIFY
Maybe-quality drift monitor
Detect when hypotheses quietly harden into claims across turns, summaries, and copied context.
- 07PACKAGE
Receipts CLI / git hook
Make claim state, evidence identity, owner, stop rule, and release boundary part of ordinary engineering work.
- 08PACKAGE
SIL library
Publish stable schemas, deterministic reference matrices, and adapters without collapsing gauges into one score.
- 09FIELD SPEC
Who Pays First field
Standardize first-burden analysis across civic, medical, workplace, and agentic decisions.
- 10INTEGRATE
PMAS + epistemic tags
Carry observations, inferences, hypotheses, frames, and uncertainty without compression laundering.
THE RELEASE RULE
Publish the useful tool with its bridge still visible.