Turn consequential assumptions
into models that update
when the evidence changes.
FP1 structures claims, grades evidence, maps dependencies, pre-registers what would prove a call wrong, and records how new evidence changes the options available to a decision-maker.
Published state as of 5 September 2026. Grade status, reconstruction notes, and publication history are preserved in the calibration record.
Every step from evidence to decision remains inspectable.
Eight stages connect an observation to a decision brief. Software supports retrieval and formatting, published rules govern repeatable steps, and consequential judgments remain attributable to a named reviewer.
Current state, September 2026: people execute the published eight-stage workflow with software supporting retrieval and formatting. The planned engine for stages 1 through 4 remains in development and does not produce current published material. Method of record: NCB-004. Machine-readable: /data/pipeline.json.
Designed for agentic execution under declared analytical rules.
The method separates software-supported steps, repeatable rules, and accountable judgment. The same inspectable record lets automation expand without erasing provenance or responsibility.
See the evidence behind the conclusion.
FP1 Evidence turns source material into discrete claims, grades the supporting evidence, preserves uncertainty and disagreement, and records what would overturn each conclusion. Receipts over verdicts: read the sources before repeating a claim.
Verification before interpretation
Context. Framing, loaded language, one-sided sourcing, counterarguments the source never answered, and the moments where it was fair.
Integrity. Examines one transcript at a time, surfacing inspectable patterns in presentation, omission, and source use without assigning motive or personal trust.
Analysis. The strategic, historical and synthesis reads, applied to the same evidence record rather than to the headline.
Why it is built this way
A bias score compresses presentation and factual accuracy into one number that hides the evidence a reader needs to inspect.
FP1 Evidence refuses the score and shows the record instead: the claim, what supports it, how independent that support is, and what would change the answer.
The Radar shows what changed and which choices it affects.
The evidence layer establishes what the record supports. The Radar tracks whether consequential assumptions are moving toward or away from the thresholds that were fixed before tracking opened. Every reading opens with the same seven fields, in the same order, so two readings can be compared line for line.
- NarrowedFinancing structures that assume utilization holds above the writedown line.
- OpenedDiligence on second-owner economics: what the installed base is worth to an operator who did not pay to build it.
- UnchangedAnything resting on end demand for compute. This reading measures the financing and utilization of supply.
- ClosedNone this reading.
Reading R-012, sweep of 29 June 2026, 17 primary sources. Each field is backed by entries in the same evidence record described above. Machine-readable: /data/readings.json.
Commit to the test before the outcome is known.
Plenty of people have models. Far fewer fix, in public, the condition that would prove the model wrong, and then hold themselves to a date. The Register is where FP1 does that, including the entries where publication ran late.
| Tracked dimension | Threshold | Provable direction | Opened | Grade date | Published |
|---|---|---|---|---|---|
| FNC-1 / SOX spread | Closes to −40 pts, or widens past −120 pts | Both | 27 July 2026 | 31 August 2026 | Held |
| Belief Index | Two weeks above 55, or any single week below 40 | Both | 27 July 2026 | 31 August 2026 | Held · near-miss |
| Biological substrate | Not registered | — | — | — | Declined, reason published |
Thresholds are fixed before the tracking window opens. If one turns out to be wrong, it is voided in public and re-registered with a new date rather than quietly moved. The Register also records what FP1 declined to register, the verdict rule fixed before each window opened, and the difference between a grade date and the date a result was actually published. Machine-readable: /data/register.json.
Three analytical lenses interrogate the same evidence record.
Domain evidence is gathered and logged before Hé, Darśan, and Manticus apply different methods to the same underlying claims. Where their conclusions conflict, the decision brief records convergence, documented dissent, or insufficient evidence.
Built to be challenged.
Four commitments that are cheap to state and expensive to keep. They are the reason a reader can argue with an FP1 output rather than only agree or disagree with it.
Pre-registration
What would make a call fail is stated first, with a date, before the tracking window opens. A threshold re-set inside its own window voids the claim for that signal and is retired in public.
Source independence
Five citations to one wire story is one source. Quantity of citations is graded separately from whether those citations are actually different sources.
Disagreement preserved
Lens disagreement resolves in one of three states: convergence, documented dissent, or insufficient evidence. It is recorded, never smoothed into a house view.
Self-falsification
The method carries four stated conditions under which it must be revised or retired, published alongside it. Each forces a restructuring rather than a quiet patch.
The instruments are backed by methods, applications, and theory.
The instruments sit on published methods, selected applications that show the framework in use, and a broader theoretical corpus that supplies candidate models, analogues and hypotheses.
NCB-003 on the composite index and NCB-004 on signal scoring and historical thresholds. These are the documentation of record for how an FP1 signal is built, scored, graded and retired.
Sixteen dated Readings and Briefings, published from 30 March to 27 July 2026 as a finite public testbed. The complete development record is preserved.
Theoretical and conceptual research that supplies candidate models, analogues and hypotheses to the analytical system. It is upstream of the instruments, not a product of them.
Long-form arguments and interviews with researchers working on agency, active inference, living systems and adaptive institutions.
Designed to sit inside an existing intelligence workflow.
FP1 is deliberately modular. Its evidence structures, analytical methods and instruments can operate as a standalone research workflow or inside an existing intelligence, research, policy, risk or decision-support environment.
Claim-level structure under existing coverage, so every conclusion carries its sources, its independence grade and its falsifier.
Monitoring tied to pivotal uncertainties rather than to headlines, with explicit scoring of which strategic options an event opens or forecloses.
Positions that can be defended in front of a committee, because the threshold and the disproof were fixed and dated in advance.
A pipeline where each stage declares what software supports, which published rules govern it, and where consequential judgment remains attributable. Any output can be re-graded on the record.
An independent research and software company for auditable decision intelligence.
FP1 develops methods and instruments for consequential decisions made under changing evidence. The work combines structured evidence, explicit falsification, preserved analytical disagreement and human accountability at the judgment layer.
Two principals. David Lovejoy leads the methodology and instrument track. John Clippinger leads the long-form theoretical track the analytical work draws on.
Get in touch
Tell us the decision in front of you and the domain that needs tracking.
First Principles First is a research and software company. Published analysis is research, not investment advice.