METHOD
named + versioned
analysis evolved
A deterministic computation layer your assistant calls instead of doing the analysis itself. Every call names its method, pins its data, and returns a run identifier that re-executes to an identical result.
DATA
pinned snapshot
VALIDATION
deflated Sharpe, held out
RUN
re-executes exactly
WHAT IT IS
A computation layer between the model and the number
Trophik sits between an analyst's assistant and the figures it reports. The assistant asks for a valuation, a screen, or a stress test. Trophik computes it with a named method at a pinned version against a stored data snapshot, and returns a run identifier anyone can re-execute. The model writes the memo. It never produces the number.
Explainable by construction
Methods are named, versioned formulas you can read, not weights. The figure in the deliverable has a citation, and the citation runs.
Full lineage
Every run carries its method version, its data snapshot and its inputs hash. Reproducing a decision from six months ago is a command, not an excavation.
How it works
01
Workflows, not primitives
value_company is one call, not fifteen. Coarse, named tools mean the same question takes the same path twice, so an answer cannot drift with the phrasing of a prompt.
02
Snapshots are stored, not named
Vendors restate history. A run pins the actual bytes it read, content addressed, and re-executes against storage rather than the vendor. Otherwise a rerun quietly disagrees with itself three weeks later, in exactly the moment someone is checking your work.
03
Validation is attached, not optional
A screen or a backtested rule comes back with a deflated Sharpe, an overfit probability and a held-out result. The evidence against an idea arrives with the idea.
DETERMINISM & AUDIT TRAIL
The model narrates. The tools compute.
No number in a deliverable is produced by a language model. Each call names its method, pins its version, and pins its data snapshot, then returns a run identifier that re-executes to an identical result.
- 01Named, versioned methods, so there is no silent drift between runs
- 02Point-in-time data, snapshot-pinned, with no lookahead
- 03Deflated Sharpe, overfit and permutation tests, held-out validation
- 04A shareable run page for the analysis behind a decision
RUN RECORD
run_id8f21c4d90
methodvaluation.blended_multiples
method_version2.1.0
data_snapshotedgar:0000320193:2026-08-09T04:00Z
snapshot_hashsha256:9c11…04ab
inputs_hashsha256:4a7d…e11c
reproducehttps://trophik.ai/r/8f21c4d90
Illustrative record. The shape of a published run, not a computed result.
WHY IT MATTERS
Most trading is algorithmic. Access is not.
Existing tools are black-box or coder-only. A portfolio manager who cannot see the method cannot defend the decision, and a firm that cannot reproduce a number cannot substantiate it.
Black-box models leave nothing to review.
Coder-only platforms exclude the people who own the mandate.
Intellectual property
PATENT PENDING, SEPTEMBER 2025Core claims
The provisional covers: A gaming interface to algorithmic trading; an ecology- and evolution-inspired framework; strategy evolution via selection, mutation, mating, predation, and mimicry; multi-regime island validation; full strategy lineage and auditability.
Why it is defensible
The process is protected, not just the models. Explainability and compliance are architectural, not bolted on. The system is difficult to replicate without the full design.
Planned expansion
Continuations on evolutionary rule discovery and ecological capital allocation and risk guardrails.
Access
METERED ON MONITORED POSITIONS AND STANDING MODELS
Local
free
1 saved method
Runs on your machine with your own data credentials. Full determinism, local run records.
Install, SeptemberTeam
$500/mo
10 monitored positions or standing models
Hosted run history, shareable run pages, scheduled re-runs, diff two runs, audit export.
Request accessFirm
$1,500/mo
50 positions or models
Team plus the full validation suite, custom methods, and priority support.
Request accessEnterprise
from $5,000/mo
unlimited
SSO, self-host or VPC, retention policy, compliance export, DPA.
Request accessWe sell compute, orchestration, storage, and audit. Your data credentials stay yours.
STATUS
Where we are
SEP 2025
Provisional patent application filed.
JUL 2026
Research platform live with the first design partner: market-wide screening, thesis monitoring, committee memos.
AUG 2026
Deterministic valuation engine reproduces an analyst's own model exactly, computed from stored inputs with no spreadsheet in the loop.
NEXT
MCP surface and hosted run history. Private beta by introduction.
MCP QUICKSTART
Call it from the assistant you already use
A small set of named, coarse-grained tools. One call, one method version, one pinned snapshot, one run identifier. Primitives stay behind an advanced flag so composition cannot drift.
Get a tokenMCP SERVER LIVE IN SEPTEMBER
{ "mcpServers": { "trophik": { "url": "https://mcp.trophik.ai" } } }value_company value one name with a named method against a pinned snapshot
explain_run given a run id, return its method, version, snapshot and inputs
compare_runs given two run ids, return the field-level diff{
"run_id": "8f21c4d90",
"method_version": "2.1.0",
"data_snapshot": "edgar:0000320193:2026-08-09T04:00Z",
"reproduce": "trophik.ai/r/8f21c4d90"
}Private beta, by introduction
We are taking a small number of institutional design partners. Tell us what you re-run most often and where the answer changes.