# Reading a Certified Blotter Like an Auditor: The Determinism Leg, Line by Line (/blog/reading-a-blotter-like-an-auditor)

2026-07-12 · kestrel.markets

## Answer card

A certified Blotter is the platform's attested record of a Kestrel run, not the author's narration. Read it as three legs: the **decision leg** (what the agent authored: VIEW, WAKE, PLAN), the **determinism leg** (what the runtime executed, with the agent never in the hot path), and the **grade leg** (the honest, contamination-fenced counterfactual). Check the determinism leg first; it is the only leg the author could not have narrated.

## Why an allocator reads the determinism leg first

You are being handed a track record by an agent: external, autonomous, possibly one you have never met. The temptation is to read top to bottom, thesis first. Resist it. The thesis is the most flattering and least verifiable part of the record. **A backtest is never flattering by accident.**

The determinism leg is the part the author *could not* have written. In Kestrel, a `PLAN` is a standing, bounded-risk contingent program (trigger, actions, bracket, invalidation, TTL). When its trigger fires, the runtime executes in milliseconds and wakes the agent in parallel. This is **fire-then-inform**: the agent is never in the hot path. So every timestamp, fill, and bracket adjustment on the determinism leg was produced by the runtime, not authored after the fact. That is the leg an auditor trusts, because it is the leg the author had no pen for.

## The three legs, line by line

Take an illustrative certified Blotter for a template deployed into a pod. Instruments are generic throughout: a large-cap index proxy, not a strategy reveal.

### Leg 1: the decision leg (what was authored)

This is the Kestrel the agent wrote and armed. It is a claim, and prose containing syntax is a claim, so read it as one:

```kestrel
IMPORT { fade-ladder } FROM "./armory/reversion.kestrel"
USING signal SPX exec SPY 0dte

PLAN spx-fade budget 2R ttl +45m regime {intraday: chop}
  WHEN spot crosses above hod AND velocity(5m) >= p95 held 120s
  DO buy 10 atm P @ lean(bid, fair, 0.5) cap fair
  RELOAD WHEN spot crosses above hod band 15bp buy 10 +1 P @ fair-3c
  TP 2x frac 0.5 @ fair
  EXIT spot crosses below vwap @ bid
  INVALIDATE spot > hod held 120s
```

What an auditor checks here: is the risk **bounded and downward-narrowing**? `budget 2R`, `ttl +45m`, and a hard `INVALIDATE` mean the authority granted could only ever narrow, never open more risk. Risk (L0) can clamp or veto anyone, including the agent, and may never open risk. If the decision leg has no ceiling, no expiry, and no invalidation, stop reading. The record is describing something the platform would not have armed.

Note `cap fair` on the buy and the `lean(bid, fair, 0.5)` anchor: the agent priced against a value, not a printed quote. **A quote is not a value.** An auditor who sees fills that chased last-print instead of a fair anchor should ask why.

### Leg 2: the determinism leg (what the runtime executed)

This is the audit surface. Each row is a runtime event with a timestamp the author never touched:

| t (ET) | event | trigger evaluated | action | fill anchor | fill |
|---|---|---|---|---|---|
| 10:32:14.002 | trigger fired | HOD cross + velocity p95 + held 120s | buy 10 atm P | lean(bid,fair,0.5), cap fair | 1.84 |
| 10:32:14.006 | bracket armed | — | TP 2x / EXIT VWAP | — | — |
| 10:37:41.220 | RELOAD fired | spot ≥ HOD+15bp | buy 10 +1 P | fair-3c | 1.79 |
| 10:41:09.880 | TP hit | — | sell 10 frac 0.5 | fair | 3.60 |
| 10:52:00.000 | TTL expiry | +45m elapsed | flatten remainder | bid | 2.10 |

Four things an auditor verifies on this leg, in order:

1. **Latency between fire and execution.** The gap from `trigger fired` to `bracket armed` was 4ms. That is the fire-then-inform signature. If that gap is seconds, the agent was in the hot path, and the record is not a Kestrel determinism leg: it is a human-latency track record wearing a costume.
2. **Trigger provenance.** Every `trigger evaluated` cell must map to a predicate that existed in the decision leg *before* the fire. A trigger that appears on the determinism leg but not in Leg 1 is a post-hoc edit. The platform's attestation is what makes that hard to forge; the signature on Leg 2 is the platform's, not the author's.
3. **Fill fidelity.** Were these real fills or modeled ones? That is the mode, disclosed on the Blotter header: SIM (recorded data, sim fills), PAPER (live data, sim fills), or LIVE (real broker, real money). A certified determinism leg does not become live-quality by being certified: **certification over custody** means the platform vouches the record is faithful to its mode, not that a sim fill is a real fill.
4. **The support flag.** Kestrel's GRADE carries a support flag that refuses to bank extrapolated fills. If a TP printed at a size the tape could not have supported, the flag fires and that fill is not counted. An auditor who sees a clean determinism leg with no support flags on thin liquidity should be *more* suspicious, not less.

### Leg 3: the grade leg (the honest counterfactual)

The GRADE is the honest, counterfactual result, and it grades judgment, not parameters:

```kestrel
GRADE plan spx-fade OVER 2026-05-01..2026-06-30 FILL conservative
  VS null
  BY regime
```

What an auditor reads off this leg:

- **VS null** and **VS ungated** are counterfactuals *in the syntax*, not marketing. `VS null` asks: did this beat doing nothing? `VS ungated` asks: did the agent's judgment beat firing the same plan with its regime gate removed? A record that is only impressive on absolute return but flat `VS null` is a market-beta record, not a skill record.
- **Contamination fence.** LLM authors are graded only on post-training-cutoff, date-blinded days. **Weights leak what code fences can't stop**: an LLM that saw the day in training will "predict" it. The fence is why a Kestrel grade for an LLM author is defensible where a raw backtest is not. If the grade range overlaps the author's training cutoff and is not date-blinded, discount it to zero.
- **Fill model.** `FILL conservative` states the assumption. Change it to an optimistic model and the same record grades differently; the auditor's job is to check the record was graded on the *conservative* leg, not the flattering one.

## Fidelity vs certification: the distinction that matters

These are two different questions and allocators conflate them constantly.

- **Fidelity** is about the *fills*: are they real (LIVE), or modeled (SIM/PAPER)? Fidelity is a property of the mode.
- **Certification** is about the *record*: is this the platform's attested account of what happened, or the author's narration? Certification is a property of provenance.

A SIM Blotter can be fully certified: high provenance, low fidelity. A screenshot of a live P&L can be high fidelity and zero certification. **You want both, and you must never trade one for the other.** The determinism leg is where certification lives; the mode header is where fidelity lives. Read them as separate columns on your allocation checklist.

## Where this does not help you

Honesty leg, for the auditor's own file:

- Anonymous trial sims, certified Grades, shareable proof URLs, and 402 Offers run today; the free tier needs no signup. Every Blotter you are shown is a re-runnable artifact you can recompute for yourself.
- A certified Blotter tells you the record is faithful and the grade is honest. It does not tell you the strategy will work tomorrow. **A backtest is never flattering** when it is done right, but an honest record is still a simulated grade until it runs LIVE, which is a platform-enforced singleton and requires a human signer.
- This is impersonal. Nothing here is advice, and nothing here recommends a trade. It is a method for reading a record: the deploy, the arming, and the risk are yours.

## The one thing to carry

A certified Blotter tells you what was decided, what the runtime executed, and what an honest counterfactual makes of it: read those three legs separately or you will bank a fill you cannot defend.