Strategy reference

The Replay — Canonical Reference

Status: LIVE (audit instrument — applies no selection) · Reconciled: 2026-08-31.

Origin: 🏠 HOUSE. Point-in-time discipline is standard practice in research; making it checkable by the reader, live, on any date is the house part. See the origins notes.

Charter: the canonical definition + current-state reference for the point-in-time replay surfaces.

One-line definition: The Replay applies no selection rule and surfaces no names: you choose the symbol and the date, and it replays what was publicly knowable on that date through the same metered API a client calls. It exists to be checked, not to recommend; it invents nothing and quotes no return.

---

1. What it is

An instrument for testing one claim: that our data carries no look-ahead. It comes in two forms.

• The worked one-pager — two historical cases followed end to end, showing what was knowable at each step and what was not.
• The live replay — pick any symbol and any date. The page calls the real entitled API contract, the same authenticated, metered, provenance-stamped endpoint a client consumes, and renders what was publicly knowable on that date with the clock stamped on every value.

This lens has no rule. It selects nothing, ranks nothing, and surfaces no name. The reader supplies both the symbol and the date; the page answers about that one input and stops. It cannot produce a list of candidates, because producing candidates is not what it does.

Where this stands: this is the surface whose entire purpose is to be falsified by the reader. If a value appeared that could not have been known on the chosen date, the claim this instrument exists to support would be broken — and the reader would be the one who found it.

2. Our variation vs. the standard technique

Point-in-time correctness is ordinary research hygiene. Ordinarily it is *asserted* in a methodology note. Here it is made executable, and the differences are deliberate:

• It runs the real contract, not a demo path. The page calls the same API a paying client calls, in-process, with a metered key. A demo that ran a special code path would prove nothing about the product.
• It prints the command to reproduce it. Each panel shows the exact request a client would issue — same date, same answer — so the reader audits the instrument rather than trusting it.
• Errors are part of the demonstration. A malformed date returns a structured error and it is rendered as-is. A surface that hid its own failures could not be used to check anything.
• The clock rides the value. Every point carries *when* it became knowable and *on what basis* — not a footnote elsewhere on the page.
• It quotes no return, anywhere. An instrument for proving honesty cannot also be a pitch.

3. How it works (methodology)

The reader supplies the inputs. A symbol and an as-of date. There is no default list, no suggested name, and no ranking of anything.

Three panels, each with its clock:

• Credibility — the newest series point that was knowable on the chosen date, carrying both the date it became knowable and the basis on which that was decided.
• Attention — the typed state-change queue as it stood, floored at the event snapshot.
• Universe — survivorship-correct membership on that date, with delisted names retained. A universe that quietly drops the companies that failed is the most common way a backtest flatters itself; retaining them is the fence.

The two-tier knowable clock. Where a real public clock was captured for a period, that clock is used. Where none was captured, the point counts as knowable only once its label period has fully ended — the conservative side, because filings usually precede the call, and that is the direction leakage would run. Which of the two rules was applied is named on the payload for every point, so the reader never has to guess whether a date was observed or modelled.

The honesty contract in code: the page lifts every value verbatim from the API envelope. It computes nothing of its own, and therefore has nothing of its own to get wrong.

4. Status, validation & honesty fence

LIVE — and there is no selection claim to validate, which is the correct status for an audit instrument:

• No return, no ranking, no recommendation appears on this surface. The standing no-edge caveat rides every credibility panel.
• What the receipt shows for this lens is a storage-policy note, not a rule: the underlying archive is kept raw and is never normalised away, so any figure can be re-derived from source. That is a statement about how data is kept, and this page does not dress it up as a selection method.
• A clean replay proves an absence, not a merit. It shows that no future information leaked into a past value. It does not show that the value is useful, predictive, or profitable, and no such inference is offered.
• The instrument degrades honestly. With no key provisioned on the box the page states plainly that it is not provisioned rather than erroring — a trust surface that breaks is worse than one that says it cannot answer.
• The key is never rendered, logged, or sent to the browser.

5. Where it lives (code · routes · DB · timers)

• Routes: /dash/replay-any-date (the replay any date code, the live instrument) · /dash/replay (the worked one-pager). Lens key replay-any-date.
• Mechanics: the call is in-process against the versioned API application, using the box's demo key from settings; the tenant is seeded idempotently and self-heals if a selftest teardown removed it.
• Receipt: registered in the evidence receipt code as replay, anchored on the raw daily equity archive — the archive whose knowable rule the replay demonstrates.
• House pattern: an isolated router with one route, importing nothing mutable at import time.

6. Data & provenance

• Archive: the NSE daily equity bhav copy — primary source (Guardrail #8), kept raw. Its manifest clock rule is EOD publish, and the pre-compute doctrine keeps the archive from being normalised away so anything shown can be re-derived from it.
• Knowable basis: every replayed point declares whether its clock was observed (a captured public event) or modelled (label-period end, the conservative fallback). This distinction is the substance of the surface, not a caveat on it.
• Survivorship: delisted names are retained in the universe count. Membership is answered as of the chosen date.

🔴 A caution that belongs on this page: the replay proves the clock, for the fields it covers. It is not a blanket certificate over every number in the estate. Read it as evidence about the specific values it displays, and no wider.

7. Terminology canon

• The Replay — this instrument. It is an audit surface, never a screen, a scan or a strategy.
• Point-in-time (PIT) — reconstructed as of a past date using only what was public by then.
• Knowable date / knowable basis — when a value became public, and whether that date was observed or modelled. 🔴 Never conflate the two; the modelled tier is deliberately conservative.
• Look-ahead — using information that was not yet public. The defect this instrument exists to expose.
• Survivorship-correct — delisted names retained rather than silently dropped.
• Replay ≠ backtest. A replay shows what was knowable. A backtest claims what would have been earned. This surface does only the first.

Reader-facing definitions: the metrics glossary.

8. Decision & session history

• The two-tier knowable clock was adopted so that a period with no captured public clock falls to the conservative side rather than being assumed knowable early.
• The live replay was built over the entitled API rather than a demo path, so the thing demonstrated is the thing sold.
• Delisted retention was chosen deliberately in universe counts; dropping them would have made every historical membership figure flattering and wrong.
• 2026-08-31 — this page created so the Evidence Receipt has a truthful rule to quote for L5. It records that no selection rule exists, because what this lens has instead is a clock, and a clock is not a rule for choosing names.

9. Open items / frozen work

• Coverage is per-field, not estate-wide. The replay covers the panels it renders. Extending it to further fields is real work, and until done, no wider claim should be made from it.
• The instrument depends on a provisioned key on the box. Where absent it states so; that honest degradation is intended, not a defect to hide.
• No study is attached to this surface and none is applicable — it makes no selection claim that could be tested for return.

10. Sources of truth

• Feeds and clock rules: the feed manifest code.
• Definitions: the metrics glossary.
• Falsification ledger: the strategy ledger — where any future claim built on replayed data must be recorded.
• Code: the replay any date code · the versioned API application it calls · the evidence receipt code.

---